Merge branch 'main' into nvidia-e2e-notebook

llama_stack/apis/agents/agents.py

@@ -13,7 +13,7 @@ from typing import Annotated, Any, Literal, Protocol, runtime_checkable
 from pydantic import BaseModel, ConfigDict, Field
 
 from llama_stack.apis.common.content_types import URL, ContentDelta, InterleavedContent
-from llama_stack.apis.common.responses import PaginatedResponse
+from llama_stack.apis.common.responses import Order, PaginatedResponse
 from llama_stack.apis.inference import (
     CompletionMessage,
     ResponseFormat,
@@ -31,6 +31,8 @@ from llama_stack.apis.tools import ToolDef
 from llama_stack.schema_utils import json_schema_type, register_schema, webmethod
 
 from .openai_responses import (
+    ListOpenAIResponseInputItem,
+    ListOpenAIResponseObject,
     OpenAIResponseInput,
     OpenAIResponseInputTool,
     OpenAIResponseObject,
@@ -579,14 +581,14 @@ class Agents(Protocol):
     #
     # Both of these APIs are inherently stateful.
 
-    @webmethod(route="/openai/v1/responses/{id}", method="GET")
+    @webmethod(route="/openai/v1/responses/{response_id}", method="GET")
     async def get_openai_response(
         self,
-        id: str,
+        response_id: str,
     ) -> OpenAIResponseObject:
         """Retrieve an OpenAI response by its ID.
 
-        :param id: The ID of the OpenAI response to retrieve.
+        :param response_id: The ID of the OpenAI response to retrieve.
         :returns: An OpenAIResponseObject.
         """
         ...
@@ -596,6 +598,7 @@ class Agents(Protocol):
         self,
         input: str | list[OpenAIResponseInput],
         model: str,
+        instructions: str | None = None,
         previous_response_id: str | None = None,
         store: bool | None = True,
         stream: bool | None = False,
@@ -610,3 +613,43 @@ class Agents(Protocol):
         :returns: An OpenAIResponseObject.
         """
         ...
+
+    @webmethod(route="/openai/v1/responses", method="GET")
+    async def list_openai_responses(
+        self,
+        after: str | None = None,
+        limit: int | None = 50,
+        model: str | None = None,
+        order: Order | None = Order.desc,
+    ) -> ListOpenAIResponseObject:
+        """List all OpenAI responses.
+
+        :param after: The ID of the last response to return.
+        :param limit: The number of responses to return.
+        :param model: The model to filter responses by.
+        :param order: The order to sort responses by when sorted by created_at ('asc' or 'desc').
+        :returns: A ListOpenAIResponseObject.
+        """
+        ...
+
+    @webmethod(route="/openai/v1/responses/{response_id}/input_items", method="GET")
+    async def list_openai_response_input_items(
+        self,
+        response_id: str,
+        after: str | None = None,
+        before: str | None = None,
+        include: list[str] | None = None,
+        limit: int | None = 20,
+        order: Order | None = Order.desc,
+    ) -> ListOpenAIResponseInputItem:
+        """List input items for a given OpenAI response.
+
+        :param response_id: The ID of the response to retrieve input items for.
+        :param after: An item ID to list items after, used for pagination.
+        :param before: An item ID to list items before, used for pagination.
+        :param include: Additional fields to include in the response.
+        :param limit: A limit on the number of objects to be returned. Limit can range between 1 and 100, and the default is 20.
+        :param order: The order to return the input items in. Default is desc.
+        :returns: An ListOpenAIResponseInputItem.
+        """
+        ...

llama_stack/apis/agents/openai_responses.py

@@ -10,6 +10,9 @@ from pydantic import BaseModel, Field
 
 from llama_stack.schema_utils import json_schema_type, register_schema
 
+# NOTE(ashwin): this file is literally a copy of the OpenAI responses API schema. We should probably
+# take their YAML and generate this file automatically. Their YAML is available.
+
 
 @json_schema_type
 class OpenAIResponseError(BaseModel):
@@ -79,16 +82,45 @@ class OpenAIResponseOutputMessageWebSearchToolCall(BaseModel):
 
 @json_schema_type
 class OpenAIResponseOutputMessageFunctionToolCall(BaseModel):
-    arguments: str
     call_id: str
     name: str
+    arguments: str
     type: Literal["function_call"] = "function_call"
+    id: str | None = None
+    status: str | None = None
+
+
+@json_schema_type
+class OpenAIResponseOutputMessageMCPCall(BaseModel):
     id: str
-    status: str
+    type: Literal["mcp_call"] = "mcp_call"
+    arguments: str
+    name: str
+    server_label: str
+    error: str | None = None
+    output: str | None = None
+
+
+class MCPListToolsTool(BaseModel):
+    input_schema: dict[str, Any]
+    name: str
+    description: str | None = None
+
+
+@json_schema_type
+class OpenAIResponseOutputMessageMCPListTools(BaseModel):
+    id: str
+    type: Literal["mcp_list_tools"] = "mcp_list_tools"
+    server_label: str
+    tools: list[MCPListToolsTool]
 
 
 OpenAIResponseOutput = Annotated[
-    OpenAIResponseMessage | OpenAIResponseOutputMessageWebSearchToolCall | OpenAIResponseOutputMessageFunctionToolCall,
+    OpenAIResponseMessage
+    | OpenAIResponseOutputMessageWebSearchToolCall
+    | OpenAIResponseOutputMessageFunctionToolCall
+    | OpenAIResponseOutputMessageMCPCall
+    | OpenAIResponseOutputMessageMCPListTools,
     Field(discriminator="type"),
 ]
 register_schema(OpenAIResponseOutput, name="OpenAIResponseOutput")
@@ -117,6 +149,16 @@ class OpenAIResponseObjectStreamResponseCreated(BaseModel):
     type: Literal["response.created"] = "response.created"
 
 
+@json_schema_type
+class OpenAIResponseObjectStreamResponseOutputTextDelta(BaseModel):
+    content_index: int
+    delta: str
+    item_id: str
+    output_index: int
+    sequence_number: int
+    type: Literal["response.output_text.delta"] = "response.output_text.delta"
+
+
 @json_schema_type
 class OpenAIResponseObjectStreamResponseCompleted(BaseModel):
     response: OpenAIResponseObject
@@ -124,7 +166,9 @@ class OpenAIResponseObjectStreamResponseCompleted(BaseModel):
 
 
 OpenAIResponseObjectStream = Annotated[
-    OpenAIResponseObjectStreamResponseCreated | OpenAIResponseObjectStreamResponseCompleted,
+    OpenAIResponseObjectStreamResponseCreated
+    | OpenAIResponseObjectStreamResponseOutputTextDelta
+    | OpenAIResponseObjectStreamResponseCompleted,
     Field(discriminator="type"),
 ]
 register_schema(OpenAIResponseObjectStream, name="OpenAIResponseObjectStream")
@@ -186,13 +230,50 @@ class OpenAIResponseInputToolFileSearch(BaseModel):
     # TODO: add filters
 
 
+class ApprovalFilter(BaseModel):
+    always: list[str] | None = None
+    never: list[str] | None = None
+
+
+class AllowedToolsFilter(BaseModel):
+    tool_names: list[str] | None = None
+
+
+@json_schema_type
+class OpenAIResponseInputToolMCP(BaseModel):
+    type: Literal["mcp"] = "mcp"
+    server_label: str
+    server_url: str
+    headers: dict[str, Any] | None = None
+
+    require_approval: Literal["always"] | Literal["never"] | ApprovalFilter = "never"
+    allowed_tools: list[str] | AllowedToolsFilter | None = None
+
+
 OpenAIResponseInputTool = Annotated[
-    OpenAIResponseInputToolWebSearch | OpenAIResponseInputToolFileSearch | OpenAIResponseInputToolFunction,
+    OpenAIResponseInputToolWebSearch
+    | OpenAIResponseInputToolFileSearch
+    | OpenAIResponseInputToolFunction
+    | OpenAIResponseInputToolMCP,
     Field(discriminator="type"),
 ]
 register_schema(OpenAIResponseInputTool, name="OpenAIResponseInputTool")
 
 
-class OpenAIResponseInputItemList(BaseModel):
+class ListOpenAIResponseInputItem(BaseModel):
     data: list[OpenAIResponseInput]
     object: Literal["list"] = "list"
+
+
+@json_schema_type
+class OpenAIResponseObjectWithInput(OpenAIResponseObject):
+    input: list[OpenAIResponseInput]
+
+
+@json_schema_type
+class ListOpenAIResponseObject(BaseModel):
+    data: list[OpenAIResponseObjectWithInput]
+    has_more: bool
+    first_id: str
+    last_id: str
+    object: Literal["list"] = "list"

llama_stack/apis/common/deployment_types.py

@@ -1,30 +0,0 @@
-# Copyright (c) Meta Platforms, Inc. and affiliates.
-# All rights reserved.
-#
-# This source code is licensed under the terms described in the LICENSE file in
-# the root directory of this source tree.
-
-from enum import Enum
-from typing import Any
-
-from pydantic import BaseModel
-
-from llama_stack.apis.common.content_types import URL
-from llama_stack.schema_utils import json_schema_type
-
-
-@json_schema_type
-class RestAPIMethod(Enum):
-    GET = "GET"
-    POST = "POST"
-    PUT = "PUT"
-    DELETE = "DELETE"
-
-
-@json_schema_type
-class RestAPIExecutionConfig(BaseModel):
-    url: URL
-    method: RestAPIMethod
-    params: dict[str, Any] | None = None
-    headers: dict[str, Any] | None = None
-    body: dict[str, Any] | None = None

llama_stack/apis/common/responses.py

@@ -4,6 +4,7 @@
 # This source code is licensed under the terms described in the LICENSE file in
 # the root directory of this source tree.
 
+from enum import Enum
 from typing import Any
 
 from pydantic import BaseModel
@@ -11,6 +12,11 @@ from pydantic import BaseModel
 from llama_stack.schema_utils import json_schema_type
 
 
+class Order(Enum):
+    asc = "asc"
+    desc = "desc"
+
+
 @json_schema_type
 class PaginatedResponse(BaseModel):
     """A generic paginated response that follows a simple format.

llama_stack/apis/inference/inference.py

@@ -19,6 +19,7 @@ from pydantic import BaseModel, Field, field_validator
 from typing_extensions import TypedDict
 
 from llama_stack.apis.common.content_types import ContentDelta, InterleavedContent, InterleavedContentItem
+from llama_stack.apis.common.responses import Order
 from llama_stack.apis.models import Model
 from llama_stack.apis.telemetry.telemetry import MetricResponseMixin
 from llama_stack.models.llama.datatypes import (
@@ -833,11 +834,6 @@ class ListOpenAIChatCompletionResponse(BaseModel):
     object: Literal["list"] = "list"
 
 
-class Order(Enum):
-    asc = "asc"
-    desc = "desc"
-
-
 @runtime_checkable
 @trace_protocol
 class InferenceProvider(Protocol):

llama_stack/apis/tools/rag_tool.py

@@ -76,6 +76,7 @@ class RAGQueryConfig(BaseModel):
     :param chunk_template: Template for formatting each retrieved chunk in the context.
         Available placeholders: {index} (1-based chunk ordinal), {chunk.content} (chunk content string), {metadata} (chunk metadata dict).
         Default: "Result {index}\\nContent: {chunk.content}\\nMetadata: {metadata}\\n"
+    :param mode: Search mode for retrieval—either "vector" or "keyword". Default "vector".
     """
 
     # This config defines how a query is generated using the messages
@@ -84,6 +85,7 @@ class RAGQueryConfig(BaseModel):
     max_tokens_in_context: int = 4096
     max_chunks: int = 5
     chunk_template: str = "Result {index}\nContent: {chunk.content}\nMetadata: {metadata}\n"
+    mode: str | None = None
 
     @field_validator("chunk_template")
     def validate_chunk_template(cls, v: str) -> str:

llama_stack/apis/tools/tools.py

@@ -27,18 +27,10 @@ class ToolParameter(BaseModel):
     default: Any | None = None
 
 
-@json_schema_type
-class ToolHost(Enum):
-    distribution = "distribution"
-    client = "client"
-    model_context_protocol = "model_context_protocol"
-
-
 @json_schema_type
 class Tool(Resource):
     type: Literal[ResourceType.tool] = ResourceType.tool
     toolgroup_id: str
-    tool_host: ToolHost
     description: str
     parameters: list[ToolParameter]
     metadata: dict[str, Any] | None = None
@@ -76,8 +68,8 @@ class ToolInvocationResult(BaseModel):
 
 
 class ToolStore(Protocol):
-    def get_tool(self, tool_name: str) -> Tool: ...
-    def get_tool_group(self, toolgroup_id: str) -> ToolGroup: ...
+    async def get_tool(self, tool_name: str) -> Tool: ...
+    async def get_tool_group(self, toolgroup_id: str) -> ToolGroup: ...
 
 
 class ListToolGroupsResponse(BaseModel):

llama_stack/cli/download.py

@@ -9,6 +9,7 @@ import asyncio
 import json
 import os
 import shutil
+import sys
 from dataclasses import dataclass
 from datetime import datetime, timezone
 from functools import partial
@@ -377,14 +378,15 @@ def _meta_download(
     downloader = ParallelDownloader(max_concurrent_downloads=max_concurrent_downloads)
     asyncio.run(downloader.download_all(tasks))
 
-    cprint(f"\nSuccessfully downloaded model to {output_dir}", "green")
+    cprint(f"\nSuccessfully downloaded model to {output_dir}", color="green", file=sys.stderr)
     cprint(
         f"\nView MD5 checksum files at: {output_dir / 'checklist.chk'}",
-        "white",
+        file=sys.stderr,
     )
     cprint(
         f"\n[Optionally] To run MD5 checksums, use the following command: llama model verify-download --model-id {model_id}",
-        "yellow",
+        color="yellow",
+        file=sys.stderr,
     )
 
 

llama_stack/cli/stack/_build.py

@@ -79,6 +79,7 @@ def run_stack_build_command(args: argparse.Namespace) -> None:
             cprint(
                 f"Could not find template {args.template}. Please run `llama stack build --list-templates` to check out the available templates",
                 color="red",
+                file=sys.stderr,
             )
             sys.exit(1)
         build_config = available_templates[args.template]
@@ -88,6 +89,7 @@ def run_stack_build_command(args: argparse.Namespace) -> None:
             cprint(
                 f"Please specify a image-type ({' | '.join(e.value for e in ImageType)}) for {args.template}",
                 color="red",
+                file=sys.stderr,
             )
             sys.exit(1)
     elif args.providers:
@@ -97,6 +99,7 @@ def run_stack_build_command(args: argparse.Namespace) -> None:
                 cprint(
                     "Could not parse `--providers`. Please ensure the list is in the format api1=provider1,api2=provider2",
                     color="red",
+                    file=sys.stderr,
                 )
                 sys.exit(1)
             api, provider = api_provider.split("=")
@@ -105,6 +108,7 @@ def run_stack_build_command(args: argparse.Namespace) -> None:
                 cprint(
                     f"{api} is not a valid API.",
                     color="red",
+                    file=sys.stderr,
                 )
                 sys.exit(1)
             if provider in providers_for_api:
@@ -113,6 +117,7 @@ def run_stack_build_command(args: argparse.Namespace) -> None:
                 cprint(
                     f"{provider} is not a valid provider for the {api} API.",
                     color="red",
+                    file=sys.stderr,
                 )
                 sys.exit(1)
         distribution_spec = DistributionSpec(
@@ -123,6 +128,7 @@ def run_stack_build_command(args: argparse.Namespace) -> None:
             cprint(
                 f"Please specify a image-type (container | conda | venv) for {args.template}",
                 color="red",
+                file=sys.stderr,
             )
             sys.exit(1)
 
@@ -151,12 +157,14 @@ def run_stack_build_command(args: argparse.Namespace) -> None:
                 cprint(
                     f"No current conda environment detected or specified, will create a new conda environment with the name `llamastack-{name}`",
                     color="yellow",
+                    file=sys.stderr,
                 )
                 image_name = f"llamastack-{name}"
             else:
                 cprint(
                     f"Using conda environment {image_name}",
                     color="green",
+                    file=sys.stderr,
                 )
         else:
             image_name = f"llamastack-{name}"
@@ -169,9 +177,10 @@ def run_stack_build_command(args: argparse.Namespace) -> None:
             """,
             ),
             color="green",
+            file=sys.stderr,
         )
 
-        print("Tip: use <TAB> to see options for the providers.\n")
+        cprint("Tip: use <TAB> to see options for the providers.\n", color="green", file=sys.stderr)
 
         providers = dict()
         for api, providers_for_api in get_provider_registry().items():
@@ -213,6 +222,7 @@ def run_stack_build_command(args: argparse.Namespace) -> None:
                 cprint(
                     f"Could not parse config file {args.config}: {e}",
                     color="red",
+                    file=sys.stderr,
                 )
                 sys.exit(1)
 
@@ -239,22 +249,25 @@ def run_stack_build_command(args: argparse.Namespace) -> None:
         cprint(
             f"Error building stack: {exc}",
             color="red",
+            file=sys.stderr,
         )
-        cprint("Stack trace:", color="red")
+        cprint("Stack trace:", color="red", file=sys.stderr)
         traceback.print_exc()
         sys.exit(1)
+
     if run_config is None:
         cprint(
             "Run config path is empty",
             color="red",
+            file=sys.stderr,
         )
         sys.exit(1)
 
     if args.run:
         config_dict = yaml.safe_load(run_config.read_text())
         config = parse_and_maybe_upgrade_config(config_dict)
-        if not os.path.exists(config.external_providers_dir):
-            os.makedirs(config.external_providers_dir, exist_ok=True)
+        if config.external_providers_dir and not config.external_providers_dir.exists():
+            config.external_providers_dir.mkdir(exist_ok=True)
         run_args = formulate_run_args(args.image_type, args.image_name, config, args.template)
         run_args.extend([str(os.getenv("LLAMA_STACK_PORT", 8321)), "--config", run_config])
         run_command(run_args)
@@ -304,6 +317,7 @@ def _generate_run_config(
                 cprint(
                     f"Failed to import provider {provider_type} for API {api} - assuming it's external, skipping",
                     color="yellow",
+                    file=sys.stderr,
                 )
                 # Set config_type to None to avoid UnboundLocalError
                 config_type = None
@@ -331,10 +345,7 @@ def _generate_run_config(
     # For non-container builds, the run.yaml is generated at the very end of the build process so it
     # makes sense to display this message
     if build_config.image_type != LlamaStackImageType.CONTAINER.value:
-        cprint(
-            f"You can now run your stack with `llama stack run {run_config_file}`",
-            color="green",
-        )
+        cprint(f"You can now run your stack with `llama stack run {run_config_file}`", color="green", file=sys.stderr)
     return run_config_file
 
 
@@ -372,7 +383,7 @@ def _run_stack_build_command_from_build_config(
     # Generate the run.yaml so it can be included in the container image with the proper entrypoint
     # Only do this if we're building a container image and we're not using a template
     if build_config.image_type == LlamaStackImageType.CONTAINER.value and not template_name and config_path:
-        cprint("Generating run.yaml file", color="green")
+        cprint("Generating run.yaml file", color="yellow", file=sys.stderr)
         run_config_file = _generate_run_config(build_config, build_dir, image_name)
 
     with open(build_file_path, "w") as f:
@@ -396,11 +407,13 @@ def _run_stack_build_command_from_build_config(
             run_config_file = build_dir / f"{template_name}-run.yaml"
             shutil.copy(path, run_config_file)
 
-        cprint("Build Successful!", color="green")
-        cprint("You can find the newly-built template here: " + colored(template_path, "light_blue"))
+        cprint("Build Successful!", color="green", file=sys.stderr)
+        cprint(f"You can find the newly-built template here: {template_path}", color="light_blue", file=sys.stderr)
         cprint(
             "You can run the new Llama Stack distro via: "
-            + colored(f"llama stack run {template_path} --image-type {build_config.image_type}", "light_blue")
+            + colored(f"llama stack run {template_path} --image-type {build_config.image_type}", "light_blue"),
+            color="green",
+            file=sys.stderr,
         )
         return template_path
     else:

llama_stack/cli/stack/list_stacks.py

@@ -0,0 +1,56 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the terms described in the LICENSE file in
+# the root directory of this source tree.
+
+import argparse
+from pathlib import Path
+
+from llama_stack.cli.subcommand import Subcommand
+from llama_stack.cli.table import print_table
+
+
+class StackListBuilds(Subcommand):
+    """List built stacks in .llama/distributions directory"""
+
+    def __init__(self, subparsers: argparse._SubParsersAction):
+        super().__init__()
+        self.parser = subparsers.add_parser(
+            "list",
+            prog="llama stack list",
+            description="list the build stacks",
+            formatter_class=argparse.ArgumentDefaultsHelpFormatter,
+        )
+        self._add_arguments()
+        self.parser.set_defaults(func=self._list_stack_command)
+
+    def _get_distribution_dirs(self) -> dict[str, Path]:
+        """Return a dictionary of distribution names and their paths"""
+        distributions = {}
+        dist_dir = Path.home() / ".llama" / "distributions"
+
+        if dist_dir.exists():
+            for stack_dir in dist_dir.iterdir():
+                if stack_dir.is_dir():
+                    distributions[stack_dir.name] = stack_dir
+        return distributions
+
+    def _list_stack_command(self, args: argparse.Namespace) -> None:
+        distributions = self._get_distribution_dirs()
+
+        if not distributions:
+            print("No stacks found in ~/.llama/distributions")
+            return
+
+        headers = ["Stack Name", "Path"]
+        headers.extend(["Build Config", "Run Config"])
+        rows = []
+        for name, path in distributions.items():
+            row = [name, str(path)]
+            # Check for build and run config files
+            build_config = "Yes" if (path / f"{name}-build.yaml").exists() else "No"
+            run_config = "Yes" if (path / f"{name}-run.yaml").exists() else "No"
+            row.extend([build_config, run_config])
+            rows.append(row)
+        print_table(rows, headers, separate_rows=True)

llama_stack/cli/stack/remove.py

@@ -0,0 +1,115 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the terms described in the LICENSE file in
+# the root directory of this source tree.
+
+import argparse
+import shutil
+import sys
+from pathlib import Path
+
+from termcolor import cprint
+
+from llama_stack.cli.subcommand import Subcommand
+from llama_stack.cli.table import print_table
+
+
+class StackRemove(Subcommand):
+    """Remove the build stack"""
+
+    def __init__(self, subparsers: argparse._SubParsersAction):
+        super().__init__()
+        self.parser = subparsers.add_parser(
+            "rm",
+            prog="llama stack rm",
+            description="Remove the build stack",
+            formatter_class=argparse.ArgumentDefaultsHelpFormatter,
+        )
+        self._add_arguments()
+        self.parser.set_defaults(func=self._remove_stack_build_command)
+
+    def _add_arguments(self) -> None:
+        self.parser.add_argument(
+            "name",
+            type=str,
+            nargs="?",
+            help="Name of the stack to delete",
+        )
+        self.parser.add_argument(
+            "--all",
+            "-a",
+            action="store_true",
+            help="Delete all stacks (use with caution)",
+        )
+
+    def _get_distribution_dirs(self) -> dict[str, Path]:
+        """Return a dictionary of distribution names and their paths"""
+        distributions = {}
+        dist_dir = Path.home() / ".llama" / "distributions"
+
+        if dist_dir.exists():
+            for stack_dir in dist_dir.iterdir():
+                if stack_dir.is_dir():
+                    distributions[stack_dir.name] = stack_dir
+        return distributions
+
+    def _list_stacks(self) -> None:
+        """Display available stacks in a table"""
+        distributions = self._get_distribution_dirs()
+        if not distributions:
+            cprint("No stacks found in ~/.llama/distributions", color="red", file=sys.stderr)
+            sys.exit(1)
+
+        headers = ["Stack Name", "Path"]
+        rows = [[name, str(path)] for name, path in distributions.items()]
+        print_table(rows, headers, separate_rows=True)
+
+    def _remove_stack_build_command(self, args: argparse.Namespace) -> None:
+        distributions = self._get_distribution_dirs()
+
+        if args.all:
+            confirm = input("Are you sure you want to delete ALL stacks? [yes-i-really-want/N] ").lower()
+            if confirm != "yes-i-really-want":
+                cprint("Deletion cancelled.", color="green", file=sys.stderr)
+                return
+
+            for name, path in distributions.items():
+                try:
+                    shutil.rmtree(path)
+                    cprint(f"Deleted stack: {name}", color="green", file=sys.stderr)
+                except Exception as e:
+                    cprint(
+                        f"Failed to delete stack {name}: {e}",
+                        color="red",
+                        file=sys.stderr,
+                    )
+                    sys.exit(1)
+
+        if not args.name:
+            self._list_stacks()
+            if not args.name:
+                return
+
+        if args.name not in distributions:
+            self._list_stacks()
+            cprint(
+                f"Stack not found: {args.name}",
+                color="red",
+                file=sys.stderr,
+            )
+            sys.exit(1)
+
+        stack_path = distributions[args.name]
+
+        confirm = input(f"Are you sure you want to delete stack '{args.name}'? [y/N] ").lower()
+        if confirm != "y":
+            cprint("Deletion cancelled.", color="green", file=sys.stderr)
+            return
+
+        try:
+            shutil.rmtree(stack_path)
+            cprint(f"Successfully deleted stack: {args.name}", color="green", file=sys.stderr)
+        except Exception as e:
+            cprint(f"Failed to delete stack {args.name}: {e}", color="red", file=sys.stderr)
+            sys.exit(1)

llama_stack/cli/stack/run.py

@@ -6,6 +6,7 @@
 
 import argparse
 import os
+import subprocess
 from pathlib import Path
 
 from llama_stack.cli.stack.utils import ImageType
@@ -60,6 +61,11 @@ class StackRun(Subcommand):
             help="Image Type used during the build. This can be either conda or container or venv.",
             choices=[e.value for e in ImageType],
         )
+        self.parser.add_argument(
+            "--enable-ui",
+            action="store_true",
+            help="Start the UI server",
+        )
 
     # If neither image type nor image name is provided, but at the same time
     # the current environment has conda breadcrumbs, then assume what the user
@@ -83,6 +89,8 @@ class StackRun(Subcommand):
         from llama_stack.distribution.utils.config_dirs import DISTRIBS_BASE_DIR
         from llama_stack.distribution.utils.exec import formulate_run_args, run_command
 
+        if args.enable_ui:
+            self._start_ui_development_server(args.port)
         image_type, image_name = self._get_image_type_and_name(args)
 
         # Check if config is required based on image type
@@ -170,3 +178,44 @@ class StackRun(Subcommand):
                     run_args.extend(["--env", f"{key}={value}"])
 
             run_command(run_args)
+
+    def _start_ui_development_server(self, stack_server_port: int):
+        logger.info("Attempting to start UI development server...")
+        # Check if npm is available
+        npm_check = subprocess.run(["npm", "--version"], capture_output=True, text=True, check=False)
+        if npm_check.returncode != 0:
+            logger.warning(
+                f"'npm' command not found or not executable. UI development server will not be started. Error: {npm_check.stderr}"
+            )
+            return
+
+        ui_dir = REPO_ROOT / "llama_stack" / "ui"
+        logs_dir = Path("~/.llama/ui/logs").expanduser()
+        try:
+            # Create logs directory if it doesn't exist
+            logs_dir.mkdir(parents=True, exist_ok=True)
+
+            ui_stdout_log_path = logs_dir / "stdout.log"
+            ui_stderr_log_path = logs_dir / "stderr.log"
+
+            # Open log files in append mode
+            stdout_log_file = open(ui_stdout_log_path, "a")
+            stderr_log_file = open(ui_stderr_log_path, "a")
+
+            process = subprocess.Popen(
+                ["npm", "run", "dev"],
+                cwd=str(ui_dir),
+                stdout=stdout_log_file,
+                stderr=stderr_log_file,
+                env={**os.environ, "NEXT_PUBLIC_LLAMA_STACK_BASE_URL": f"http://localhost:{stack_server_port}"},
+            )
+            logger.info(f"UI development server process started in {ui_dir} with PID {process.pid}.")
+            logger.info(f"Logs: stdout -> {ui_stdout_log_path}, stderr -> {ui_stderr_log_path}")
+            logger.info(f"UI will be available at http://localhost:{os.getenv('LLAMA_STACK_UI_PORT', 8322)}")
+
+        except FileNotFoundError:
+            logger.error(
+                "Failed to start UI development server: 'npm' command not found. Make sure npm is installed and in your PATH."
+            )
+        except Exception as e:
+            logger.error(f"Failed to start UI development server in {ui_dir}: {e}")

llama_stack/cli/stack/stack.py

@@ -7,12 +7,14 @@
 import argparse
 from importlib.metadata import version
 
+from llama_stack.cli.stack.list_stacks import StackListBuilds
 from llama_stack.cli.stack.utils import print_subcommand_description
 from llama_stack.cli.subcommand import Subcommand
 
 from .build import StackBuild
 from .list_apis import StackListApis
 from .list_providers import StackListProviders
+from .remove import StackRemove
 from .run import StackRun
 
 
@@ -41,5 +43,6 @@ class StackParser(Subcommand):
         StackListApis.create(subparsers)
         StackListProviders.create(subparsers)
         StackRun.create(subparsers)
-
+        StackRemove.create(subparsers)
+        StackListBuilds.create(subparsers)
         print_subcommand_description(self.parser, subparsers)

llama_stack/distribution/build.py

@@ -6,6 +6,7 @@
 
 import importlib.resources
 import logging
+import sys
 from pathlib import Path
 
 from pydantic import BaseModel
@@ -43,8 +44,20 @@ def get_provider_dependencies(
     # Extract providers based on config type
     if isinstance(config, DistributionTemplate):
         providers = config.providers
+
+        # TODO: This is a hack to get the dependencies for internal APIs into build
+        # We should have a better way to do this by formalizing the concept of "internal" APIs
+        # and providers, with a way to specify dependencies for them.
+        run_configs = config.run_configs
+        additional_pip_packages: list[str] = []
+        if run_configs:
+            for run_config in run_configs.values():
+                run_config_ = run_config.run_config(name="", providers={}, container_image=None)
+                if run_config_.inference_store:
+                    additional_pip_packages.extend(run_config_.inference_store.pip_packages)
     elif isinstance(config, BuildConfig):
         providers = config.distribution_spec.providers
+        additional_pip_packages = config.additional_pip_packages
     deps = []
     registry = get_provider_registry(config)
     for api_str, provider_or_providers in providers.items():
@@ -72,6 +85,9 @@ def get_provider_dependencies(
         else:
             normal_deps.append(package)
 
+    if additional_pip_packages:
+        normal_deps.extend(additional_pip_packages)
+
     return list(set(normal_deps)), list(set(special_deps))
 
 
@@ -80,10 +96,11 @@ def print_pip_install_help(config: BuildConfig):
 
     cprint(
         f"Please install needed dependencies using the following commands:\n\nuv pip install {' '.join(normal_deps)}",
-        "yellow",
+        color="yellow",
+        file=sys.stderr,
     )
     for special_dep in special_deps:
-        cprint(f"uv pip install {special_dep}", "yellow")
+        cprint(f"uv pip install {special_dep}", color="yellow", file=sys.stderr)
     print()
 
 

llama_stack/distribution/client.py

@@ -6,6 +6,7 @@
 
 import inspect
 import json
+import sys
 from collections.abc import AsyncIterator
 from enum import Enum
 from typing import Any, Union, get_args, get_origin
@@ -96,13 +97,13 @@ def create_api_client_class(protocol) -> type:
                             try:
                                 data = json.loads(data)
                                 if "error" in data:
-                                    cprint(data, "red")
+                                    cprint(data, color="red", file=sys.stderr)
                                     continue
 
                                 yield parse_obj_as(return_type, data)
                             except Exception as e:
-                                print(f"Error with parsing or validation: {e}")
-                                print(data)
+                                cprint(f"Error with parsing or validation: {e}", color="red", file=sys.stderr)
+                                cprint(data, color="red", file=sys.stderr)
 
         def httpx_request_params(self, method_name: str, *args, **kwargs) -> dict:
             webmethod, sig = self.routes[method_name]

llama_stack/distribution/datatypes.py

@@ -25,7 +25,8 @@ from llama_stack.apis.tools import Tool, ToolGroup, ToolGroupInput, ToolRuntime
 from llama_stack.apis.vector_dbs import VectorDB, VectorDBInput
 from llama_stack.apis.vector_io import VectorIO
 from llama_stack.providers.datatypes import Api, ProviderSpec
-from llama_stack.providers.utils.kvstore.config import KVStoreConfig
+from llama_stack.providers.utils.kvstore.config import KVStoreConfig, SqliteKVStoreConfig
+from llama_stack.providers.utils.sqlstore.sqlstore import SqlStoreConfig
 
 LLAMA_STACK_BUILD_CONFIG_VERSION = "2"
 LLAMA_STACK_RUN_CONFIG_VERSION = "2"
@@ -220,21 +221,38 @@ class LoggingConfig(BaseModel):
 class AuthProviderType(str, Enum):
     """Supported authentication provider types."""
 
-    KUBERNETES = "kubernetes"
+    OAUTH2_TOKEN = "oauth2_token"
     CUSTOM = "custom"
 
 
 class AuthenticationConfig(BaseModel):
     provider_type: AuthProviderType = Field(
         ...,
-        description="Type of authentication provider (e.g., 'kubernetes', 'custom')",
+        description="Type of authentication provider",
     )
-    config: dict[str, str] = Field(
+    config: dict[str, Any] = Field(
         ...,
         description="Provider-specific configuration",
     )
 
 
+class AuthenticationRequiredError(Exception):
+    pass
+
+
+class QuotaPeriod(str, Enum):
+    DAY = "day"
+
+
+class QuotaConfig(BaseModel):
+    kvstore: SqliteKVStoreConfig = Field(description="Config for KV store backend (SQLite only for now)")
+    anonymous_max_requests: int = Field(default=100, description="Max requests for unauthenticated clients per period")
+    authenticated_max_requests: int = Field(
+        default=1000, description="Max requests for authenticated clients per period"
+    )
+    period: QuotaPeriod = Field(default=QuotaPeriod.DAY, description="Quota period to set")
+
+
 class ServerConfig(BaseModel):
     port: int = Field(
         default=8321,
@@ -262,6 +280,10 @@ class ServerConfig(BaseModel):
         default=None,
         description="The host the server should listen on",
     )
+    quota: QuotaConfig | None = Field(
+        default=None,
+        description="Per client quota request configuration",
+    )
 
 
 class StackRunConfig(BaseModel):
@@ -297,6 +319,13 @@ Configuration for the persistence store used by the distribution registry. If no
 a default SQLite store will be used.""",
     )
 
+    inference_store: SqlStoreConfig | None = Field(
+        default=None,
+        description="""
+Configuration for the persistence store used by the inference API. If not specified,
+a default SQLite store will be used.""",
+    )
+
     # registry of "resources" in the distribution
     models: list[ModelInput] = Field(default_factory=list)
     shields: list[ShieldInput] = Field(default_factory=list)
@@ -345,6 +374,10 @@ class BuildConfig(BaseModel):
         description="Path to directory containing external provider implementations. The providers packages will be resolved from this directory. "
         "pip_packages MUST contain the provider package name.",
     )
+    additional_pip_packages: list[str] = Field(
+        default_factory=list,
+        description="Additional pip packages to install in the distribution. These packages will be installed in the distribution environment.",
+    )
 
     @field_validator("external_providers_dir")
     @classmethod

llama_stack/distribution/inspect.py

@@ -31,7 +31,7 @@ async def get_provider_impl(config, deps):
 
 
 class DistributionInspectImpl(Inspect):
-    def __init__(self, config, deps):
+    def __init__(self, config: DistributionInspectConfig, deps):
         self.config = config
         self.deps = deps
 
@@ -39,22 +39,36 @@ class DistributionInspectImpl(Inspect):
         pass
 
     async def list_routes(self) -> ListRoutesResponse:
-        run_config = self.config.run_config
+        run_config: StackRunConfig = self.config.run_config
 
         ret = []
         all_endpoints = get_all_api_endpoints()
         for api, endpoints in all_endpoints.items():
-            providers = run_config.providers.get(api.value, [])
-            ret.extend(
-                [
-                    RouteInfo(
-                        route=e.route,
-                        method=e.method,
-                        provider_types=[p.provider_type for p in providers],
+            # Always include provider and inspect APIs, filter others based on run config
+            if api.value in ["providers", "inspect"]:
+                ret.extend(
+                    [
+                        RouteInfo(
+                            route=e.route,
+                            method=e.method,
+                            provider_types=[],  # These APIs don't have "real" providers - they're internal to the stack
+                        )
+                        for e in endpoints
+                    ]
+                )
+            else:
+                providers = run_config.providers.get(api.value, [])
+                if providers:  # Only process if there are providers for this API
+                    ret.extend(
+                        [
+                            RouteInfo(
+                                route=e.route,
+                                method=e.method,
+                                provider_types=[p.provider_type for p in providers],
+                            )
+                            for e in endpoints
+                        ]
                     )
-                    for e in endpoints
-                ]
-            )
 
         return ListRoutesResponse(data=ret)
 

llama_stack/distribution/library_client.py

@@ -9,6 +9,7 @@ import inspect
 import json
 import logging
 import os
+import sys
 from concurrent.futures import ThreadPoolExecutor
 from enum import Enum
 from pathlib import Path
@@ -210,10 +211,11 @@ class AsyncLlamaStackAsLibraryClient(AsyncLlamaStackClient):
             self.endpoint_impls = None
             self.impls = await construct_stack(self.config, self.custom_provider_registry)
         except ModuleNotFoundError as _e:
-            cprint(_e.msg, "red")
+            cprint(_e.msg, color="red", file=sys.stderr)
             cprint(
                 "Using llama-stack as a library requires installing dependencies depending on the template (providers) you choose.\n",
-                "yellow",
+                color="yellow",
+                file=sys.stderr,
             )
             if self.config_path_or_template_name.endswith(".yaml"):
                 # Convert Provider objects to their types
@@ -234,7 +236,13 @@ class AsyncLlamaStackAsLibraryClient(AsyncLlamaStackClient):
                 cprint(
                     f"Please run:\n\n{prefix}llama stack build --template {self.config_path_or_template_name} --image-type venv\n\n",
                     "yellow",
+                    file=sys.stderr,
                 )
+            cprint(
+                "Please check your internet connection and try again.",
+                "red",
+                file=sys.stderr,
+            )
             raise _e
 
         if Api.telemetry in self.impls:
@@ -261,9 +269,11 @@ class AsyncLlamaStackAsLibraryClient(AsyncLlamaStackClient):
             raise ValueError("Client not initialized")
 
         # Create headers with provider data if available
-        headers = {}
+        headers = options.headers or {}
         if self.provider_data:
-            headers["X-LlamaStack-Provider-Data"] = json.dumps(self.provider_data)
+            keys = ["X-LlamaStack-Provider-Data", "x-llamastack-provider-data"]
+            if all(key not in headers for key in keys):
+                headers["X-LlamaStack-Provider-Data"] = json.dumps(self.provider_data)
 
         # Use context manager for provider data
         with request_provider_data_context(headers):

llama_stack/distribution/resolver.py

@@ -47,7 +47,7 @@ from llama_stack.providers.datatypes import (
     RemoteProviderSpec,
     ScoringFunctionsProtocolPrivate,
     ShieldsProtocolPrivate,
-    ToolsProtocolPrivate,
+    ToolGroupsProtocolPrivate,
     VectorDBsProtocolPrivate,
 )
 
@@ -93,7 +93,7 @@ def api_protocol_map_for_compliance_check() -> dict[Api, Any]:
 def additional_protocols_map() -> dict[Api, Any]:
     return {
         Api.inference: (ModelsProtocolPrivate, Models, Api.models),
-        Api.tool_groups: (ToolsProtocolPrivate, ToolGroups, Api.tool_groups),
+        Api.tool_groups: (ToolGroupsProtocolPrivate, ToolGroups, Api.tool_groups),
         Api.vector_io: (VectorDBsProtocolPrivate, VectorDBs, Api.vector_dbs),
         Api.safety: (ShieldsProtocolPrivate, Shields, Api.shields),
         Api.datasetio: (DatasetsProtocolPrivate, Datasets, Api.datasets),
@@ -140,7 +140,7 @@ async def resolve_impls(
 
     sorted_providers = sort_providers_by_deps(providers_with_specs, run_config)
 
-    return await instantiate_providers(sorted_providers, router_apis, dist_registry)
+    return await instantiate_providers(sorted_providers, router_apis, dist_registry, run_config)
 
 
 def specs_for_autorouted_apis(apis_to_serve: list[str] | set[str]) -> dict[str, dict[str, ProviderWithSpec]]:
@@ -243,7 +243,10 @@ def sort_providers_by_deps(
 
 
 async def instantiate_providers(
-    sorted_providers: list[tuple[str, ProviderWithSpec]], router_apis: set[Api], dist_registry: DistributionRegistry
+    sorted_providers: list[tuple[str, ProviderWithSpec]],
+    router_apis: set[Api],
+    dist_registry: DistributionRegistry,
+    run_config: StackRunConfig,
 ) -> dict:
     """Instantiates providers asynchronously while managing dependencies."""
     impls: dict[Api, Any] = {}
@@ -258,7 +261,7 @@ async def instantiate_providers(
         if isinstance(provider.spec, RoutingTableProviderSpec):
             inner_impls = inner_impls_by_provider_id[f"inner-{provider.spec.router_api.value}"]
 
-        impl = await instantiate_provider(provider, deps, inner_impls, dist_registry)
+        impl = await instantiate_provider(provider, deps, inner_impls, dist_registry, run_config)
 
         if api_str.startswith("inner-"):
             inner_impls_by_provider_id[api_str][provider.provider_id] = impl
@@ -308,6 +311,7 @@ async def instantiate_provider(
     deps: dict[Api, Any],
     inner_impls: dict[str, Any],
     dist_registry: DistributionRegistry,
+    run_config: StackRunConfig,
 ):
     provider_spec = provider.spec
     if not hasattr(provider_spec, "module"):
@@ -327,7 +331,7 @@ async def instantiate_provider(
         method = "get_auto_router_impl"
 
         config = None
-        args = [provider_spec.api, deps[provider_spec.routing_table_api], deps]
+        args = [provider_spec.api, deps[provider_spec.routing_table_api], deps, run_config]
     elif isinstance(provider_spec, RoutingTableProviderSpec):
         method = "get_routing_table_impl"
 

llama_stack/distribution/routers/__init__.py

@@ -7,18 +7,10 @@
 from typing import Any
 
 from llama_stack.distribution.datatypes import RoutedProtocol
+from llama_stack.distribution.stack import StackRunConfig
 from llama_stack.distribution.store import DistributionRegistry
 from llama_stack.providers.datatypes import Api, RoutingTable
-
-from .routing_tables import (
-    BenchmarksRoutingTable,
-    DatasetsRoutingTable,
-    ModelsRoutingTable,
-    ScoringFunctionsRoutingTable,
-    ShieldsRoutingTable,
-    ToolGroupsRoutingTable,
-    VectorDBsRoutingTable,
-)
+from llama_stack.providers.utils.inference.inference_store import InferenceStore
 
 
 async def get_routing_table_impl(
@@ -27,6 +19,14 @@ async def get_routing_table_impl(
     _deps,
     dist_registry: DistributionRegistry,
 ) -> Any:
+    from ..routing_tables.benchmarks import BenchmarksRoutingTable
+    from ..routing_tables.datasets import DatasetsRoutingTable
+    from ..routing_tables.models import ModelsRoutingTable
+    from ..routing_tables.scoring_functions import ScoringFunctionsRoutingTable
+    from ..routing_tables.shields import ShieldsRoutingTable
+    from ..routing_tables.toolgroups import ToolGroupsRoutingTable
+    from ..routing_tables.vector_dbs import VectorDBsRoutingTable
+
     api_to_tables = {
         "vector_dbs": VectorDBsRoutingTable,
         "models": ModelsRoutingTable,
@@ -45,16 +45,15 @@ async def get_routing_table_impl(
     return impl
 
 
-async def get_auto_router_impl(api: Api, routing_table: RoutingTable, deps: dict[str, Any]) -> Any:
-    from .routers import (
-        DatasetIORouter,
-        EvalRouter,
-        InferenceRouter,
-        SafetyRouter,
-        ScoringRouter,
-        ToolRuntimeRouter,
-        VectorIORouter,
-    )
+async def get_auto_router_impl(
+    api: Api, routing_table: RoutingTable, deps: dict[str, Any], run_config: StackRunConfig
+) -> Any:
+    from .datasets import DatasetIORouter
+    from .eval_scoring import EvalRouter, ScoringRouter
+    from .inference import InferenceRouter
+    from .safety import SafetyRouter
+    from .tool_runtime import ToolRuntimeRouter
+    from .vector_io import VectorIORouter
 
     api_to_routers = {
         "vector_io": VectorIORouter,
@@ -76,6 +75,12 @@ async def get_auto_router_impl(api: Api, routing_table: RoutingTable, deps: dict
         if dep_api in deps:
             api_to_dep_impl[dep_name] = deps[dep_api]
 
+    # TODO: move pass configs to routers instead
+    if api == Api.inference and run_config.inference_store:
+        inference_store = InferenceStore(run_config.inference_store)
+        await inference_store.initialize()
+        api_to_dep_impl["store"] = inference_store
+
     impl = api_to_routers[api.value](routing_table, **api_to_dep_impl)
     await impl.initialize()
     return impl

llama_stack/distribution/routers/datasets.py

@@ -0,0 +1,71 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the terms described in the LICENSE file in
+# the root directory of this source tree.
+
+from typing import Any
+
+from llama_stack.apis.common.responses import PaginatedResponse
+from llama_stack.apis.datasetio import DatasetIO
+from llama_stack.apis.datasets import DatasetPurpose, DataSource
+from llama_stack.log import get_logger
+from llama_stack.providers.datatypes import RoutingTable
+
+logger = get_logger(name=__name__, category="core")
+
+
+class DatasetIORouter(DatasetIO):
+    def __init__(
+        self,
+        routing_table: RoutingTable,
+    ) -> None:
+        logger.debug("Initializing DatasetIORouter")
+        self.routing_table = routing_table
+
+    async def initialize(self) -> None:
+        logger.debug("DatasetIORouter.initialize")
+        pass
+
+    async def shutdown(self) -> None:
+        logger.debug("DatasetIORouter.shutdown")
+        pass
+
+    async def register_dataset(
+        self,
+        purpose: DatasetPurpose,
+        source: DataSource,
+        metadata: dict[str, Any] | None = None,
+        dataset_id: str | None = None,
+    ) -> None:
+        logger.debug(
+            f"DatasetIORouter.register_dataset: {purpose=} {source=} {metadata=} {dataset_id=}",
+        )
+        await self.routing_table.register_dataset(
+            purpose=purpose,
+            source=source,
+            metadata=metadata,
+            dataset_id=dataset_id,
+        )
+
+    async def iterrows(
+        self,
+        dataset_id: str,
+        start_index: int | None = None,
+        limit: int | None = None,
+    ) -> PaginatedResponse:
+        logger.debug(
+            f"DatasetIORouter.iterrows: {dataset_id}, {start_index=} {limit=}",
+        )
+        return await self.routing_table.get_provider_impl(dataset_id).iterrows(
+            dataset_id=dataset_id,
+            start_index=start_index,
+            limit=limit,
+        )
+
+    async def append_rows(self, dataset_id: str, rows: list[dict[str, Any]]) -> None:
+        logger.debug(f"DatasetIORouter.append_rows: {dataset_id}, {len(rows)} rows")
+        return await self.routing_table.get_provider_impl(dataset_id).append_rows(
+            dataset_id=dataset_id,
+            rows=rows,
+        )

llama_stack/distribution/routers/eval_scoring.py

@@ -0,0 +1,148 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the terms described in the LICENSE file in
+# the root directory of this source tree.
+
+from typing import Any
+
+from llama_stack.apis.eval import BenchmarkConfig, Eval, EvaluateResponse, Job
+from llama_stack.apis.scoring import (
+    ScoreBatchResponse,
+    ScoreResponse,
+    Scoring,
+    ScoringFnParams,
+)
+from llama_stack.log import get_logger
+from llama_stack.providers.datatypes import RoutingTable
+
+logger = get_logger(name=__name__, category="core")
+
+
+class ScoringRouter(Scoring):
+    def __init__(
+        self,
+        routing_table: RoutingTable,
+    ) -> None:
+        logger.debug("Initializing ScoringRouter")
+        self.routing_table = routing_table
+
+    async def initialize(self) -> None:
+        logger.debug("ScoringRouter.initialize")
+        pass
+
+    async def shutdown(self) -> None:
+        logger.debug("ScoringRouter.shutdown")
+        pass
+
+    async def score_batch(
+        self,
+        dataset_id: str,
+        scoring_functions: dict[str, ScoringFnParams | None] = None,
+        save_results_dataset: bool = False,
+    ) -> ScoreBatchResponse:
+        logger.debug(f"ScoringRouter.score_batch: {dataset_id}")
+        res = {}
+        for fn_identifier in scoring_functions.keys():
+            score_response = await self.routing_table.get_provider_impl(fn_identifier).score_batch(
+                dataset_id=dataset_id,
+                scoring_functions={fn_identifier: scoring_functions[fn_identifier]},
+            )
+            res.update(score_response.results)
+
+        if save_results_dataset:
+            raise NotImplementedError("Save results dataset not implemented yet")
+
+        return ScoreBatchResponse(
+            results=res,
+        )
+
+    async def score(
+        self,
+        input_rows: list[dict[str, Any]],
+        scoring_functions: dict[str, ScoringFnParams | None] = None,
+    ) -> ScoreResponse:
+        logger.debug(f"ScoringRouter.score: {len(input_rows)} rows, {len(scoring_functions)} functions")
+        res = {}
+        # look up and map each scoring function to its provider impl
+        for fn_identifier in scoring_functions.keys():
+            score_response = await self.routing_table.get_provider_impl(fn_identifier).score(
+                input_rows=input_rows,
+                scoring_functions={fn_identifier: scoring_functions[fn_identifier]},
+            )
+            res.update(score_response.results)
+
+        return ScoreResponse(results=res)
+
+
+class EvalRouter(Eval):
+    def __init__(
+        self,
+        routing_table: RoutingTable,
+    ) -> None:
+        logger.debug("Initializing EvalRouter")
+        self.routing_table = routing_table
+
+    async def initialize(self) -> None:
+        logger.debug("EvalRouter.initialize")
+        pass
+
+    async def shutdown(self) -> None:
+        logger.debug("EvalRouter.shutdown")
+        pass
+
+    async def run_eval(
+        self,
+        benchmark_id: str,
+        benchmark_config: BenchmarkConfig,
+    ) -> Job:
+        logger.debug(f"EvalRouter.run_eval: {benchmark_id}")
+        return await self.routing_table.get_provider_impl(benchmark_id).run_eval(
+            benchmark_id=benchmark_id,
+            benchmark_config=benchmark_config,
+        )
+
+    async def evaluate_rows(
+        self,
+        benchmark_id: str,
+        input_rows: list[dict[str, Any]],
+        scoring_functions: list[str],
+        benchmark_config: BenchmarkConfig,
+    ) -> EvaluateResponse:
+        logger.debug(f"EvalRouter.evaluate_rows: {benchmark_id}, {len(input_rows)} rows")
+        return await self.routing_table.get_provider_impl(benchmark_id).evaluate_rows(
+            benchmark_id=benchmark_id,
+            input_rows=input_rows,
+            scoring_functions=scoring_functions,
+            benchmark_config=benchmark_config,
+        )
+
+    async def job_status(
+        self,
+        benchmark_id: str,
+        job_id: str,
+    ) -> Job:
+        logger.debug(f"EvalRouter.job_status: {benchmark_id}, {job_id}")
+        return await self.routing_table.get_provider_impl(benchmark_id).job_status(benchmark_id, job_id)
+
+    async def job_cancel(
+        self,
+        benchmark_id: str,
+        job_id: str,
+    ) -> None:
+        logger.debug(f"EvalRouter.job_cancel: {benchmark_id}, {job_id}")
+        await self.routing_table.get_provider_impl(benchmark_id).job_cancel(
+            benchmark_id,
+            job_id,
+        )
+
+    async def job_result(
+        self,
+        benchmark_id: str,
+        job_id: str,
+    ) -> EvaluateResponse:
+        logger.debug(f"EvalRouter.job_result: {benchmark_id}, {job_id}")
+        return await self.routing_table.get_provider_impl(benchmark_id).job_result(
+            benchmark_id,
+            job_id,
+        )

llama_stack/distribution/routers/inference.py

@@ -14,14 +14,9 @@ from openai.types.chat import ChatCompletionToolParam as OpenAIChatCompletionToo
 from pydantic import Field, TypeAdapter
 
 from llama_stack.apis.common.content_types import (
-    URL,
     InterleavedContent,
     InterleavedContentItem,
 )
-from llama_stack.apis.common.responses import PaginatedResponse
-from llama_stack.apis.datasetio import DatasetIO
-from llama_stack.apis.datasets import DatasetPurpose, DataSource
-from llama_stack.apis.eval import BenchmarkConfig, Eval, EvaluateResponse, Job
 from llama_stack.apis.inference import (
     BatchChatCompletionResponse,
     BatchCompletionResponse,
@@ -32,8 +27,11 @@ from llama_stack.apis.inference import (
     EmbeddingsResponse,
     EmbeddingTaskType,
     Inference,
+    ListOpenAIChatCompletionResponse,
     LogProbConfig,
     Message,
+    OpenAICompletionWithInputMessages,
+    Order,
     ResponseFormat,
     SamplingParams,
     StopReason,
@@ -51,89 +49,18 @@ from llama_stack.apis.inference.inference import (
     OpenAIResponseFormatParam,
 )
 from llama_stack.apis.models import Model, ModelType
-from llama_stack.apis.safety import RunShieldResponse, Safety
-from llama_stack.apis.scoring import (
-    ScoreBatchResponse,
-    ScoreResponse,
-    Scoring,
-    ScoringFnParams,
-)
-from llama_stack.apis.shields import Shield
 from llama_stack.apis.telemetry import MetricEvent, MetricInResponse, Telemetry
-from llama_stack.apis.tools import (
-    ListToolDefsResponse,
-    RAGDocument,
-    RAGQueryConfig,
-    RAGQueryResult,
-    RAGToolRuntime,
-    ToolRuntime,
-)
-from llama_stack.apis.vector_io import Chunk, QueryChunksResponse, VectorIO
 from llama_stack.log import get_logger
 from llama_stack.models.llama.llama3.chat_format import ChatFormat
 from llama_stack.models.llama.llama3.tokenizer import Tokenizer
 from llama_stack.providers.datatypes import HealthResponse, HealthStatus, RoutingTable
+from llama_stack.providers.utils.inference.inference_store import InferenceStore
+from llama_stack.providers.utils.inference.stream_utils import stream_and_store_openai_completion
 from llama_stack.providers.utils.telemetry.tracing import get_current_span
 
 logger = get_logger(name=__name__, category="core")
 
 
-class VectorIORouter(VectorIO):
-    """Routes to an provider based on the vector db identifier"""
-
-    def __init__(
-        self,
-        routing_table: RoutingTable,
-    ) -> None:
-        logger.debug("Initializing VectorIORouter")
-        self.routing_table = routing_table
-
-    async def initialize(self) -> None:
-        logger.debug("VectorIORouter.initialize")
-        pass
-
-    async def shutdown(self) -> None:
-        logger.debug("VectorIORouter.shutdown")
-        pass
-
-    async def register_vector_db(
-        self,
-        vector_db_id: str,
-        embedding_model: str,
-        embedding_dimension: int | None = 384,
-        provider_id: str | None = None,
-        provider_vector_db_id: str | None = None,
-    ) -> None:
-        logger.debug(f"VectorIORouter.register_vector_db: {vector_db_id}, {embedding_model}")
-        await self.routing_table.register_vector_db(
-            vector_db_id,
-            embedding_model,
-            embedding_dimension,
-            provider_id,
-            provider_vector_db_id,
-        )
-
-    async def insert_chunks(
-        self,
-        vector_db_id: str,
-        chunks: list[Chunk],
-        ttl_seconds: int | None = None,
-    ) -> None:
-        logger.debug(
-            f"VectorIORouter.insert_chunks: {vector_db_id}, {len(chunks)} chunks, ttl_seconds={ttl_seconds}, chunk_ids={[chunk.metadata['document_id'] for chunk in chunks[:3]]}{' and more...' if len(chunks) > 3 else ''}",
-        )
-        return await self.routing_table.get_provider_impl(vector_db_id).insert_chunks(vector_db_id, chunks, ttl_seconds)
-
-    async def query_chunks(
-        self,
-        vector_db_id: str,
-        query: InterleavedContent,
-        params: dict[str, Any] | None = None,
-    ) -> QueryChunksResponse:
-        logger.debug(f"VectorIORouter.query_chunks: {vector_db_id}")
-        return await self.routing_table.get_provider_impl(vector_db_id).query_chunks(vector_db_id, query, params)
-
-
 class InferenceRouter(Inference):
     """Routes to an provider based on the model"""
 
@@ -141,10 +68,12 @@ class InferenceRouter(Inference):
         self,
         routing_table: RoutingTable,
         telemetry: Telemetry | None = None,
+        store: InferenceStore | None = None,
     ) -> None:
         logger.debug("Initializing InferenceRouter")
         self.routing_table = routing_table
         self.telemetry = telemetry
+        self.store = store
         if self.telemetry:
             self.tokenizer = Tokenizer.get_instance()
             self.formatter = ChatFormat(self.tokenizer)
@@ -607,9 +536,31 @@ class InferenceRouter(Inference):
 
         provider = self.routing_table.get_provider_impl(model_obj.identifier)
         if stream:
-            return await provider.openai_chat_completion(**params)
+            response_stream = await provider.openai_chat_completion(**params)
+            if self.store:
+                return stream_and_store_openai_completion(response_stream, model, self.store, messages)
+            return response_stream
         else:
-            return await self._nonstream_openai_chat_completion(provider, params)
+            response = await self._nonstream_openai_chat_completion(provider, params)
+            if self.store:
+                await self.store.store_chat_completion(response, messages)
+            return response
+
+    async def list_chat_completions(
+        self,
+        after: str | None = None,
+        limit: int | None = 20,
+        model: str | None = None,
+        order: Order | None = Order.desc,
+    ) -> ListOpenAIChatCompletionResponse:
+        if self.store:
+            return await self.store.list_chat_completions(after, limit, model, order)
+        raise NotImplementedError("List chat completions is not supported: inference store is not configured.")
+
+    async def get_chat_completion(self, completion_id: str) -> OpenAICompletionWithInputMessages:
+        if self.store:
+            return await self.store.get_chat_completion(completion_id)
+        raise NotImplementedError("Get chat completion is not supported: inference store is not configured.")
 
     async def _nonstream_openai_chat_completion(self, provider: Inference, params: dict) -> OpenAIChatCompletion:
         response = await provider.openai_chat_completion(**params)
@@ -642,295 +593,3 @@ class InferenceRouter(Inference):
                     status=HealthStatus.ERROR, message=f"Health check failed: {str(e)}"
                 )
         return health_statuses
-
-
-class SafetyRouter(Safety):
-    def __init__(
-        self,
-        routing_table: RoutingTable,
-    ) -> None:
-        logger.debug("Initializing SafetyRouter")
-        self.routing_table = routing_table
-
-    async def initialize(self) -> None:
-        logger.debug("SafetyRouter.initialize")
-        pass
-
-    async def shutdown(self) -> None:
-        logger.debug("SafetyRouter.shutdown")
-        pass
-
-    async def register_shield(
-        self,
-        shield_id: str,
-        provider_shield_id: str | None = None,
-        provider_id: str | None = None,
-        params: dict[str, Any] | None = None,
-    ) -> Shield:
-        logger.debug(f"SafetyRouter.register_shield: {shield_id}")
-        return await self.routing_table.register_shield(shield_id, provider_shield_id, provider_id, params)
-
-    async def run_shield(
-        self,
-        shield_id: str,
-        messages: list[Message],
-        params: dict[str, Any] = None,
-    ) -> RunShieldResponse:
-        logger.debug(f"SafetyRouter.run_shield: {shield_id}")
-        return await self.routing_table.get_provider_impl(shield_id).run_shield(
-            shield_id=shield_id,
-            messages=messages,
-            params=params,
-        )
-
-
-class DatasetIORouter(DatasetIO):
-    def __init__(
-        self,
-        routing_table: RoutingTable,
-    ) -> None:
-        logger.debug("Initializing DatasetIORouter")
-        self.routing_table = routing_table
-
-    async def initialize(self) -> None:
-        logger.debug("DatasetIORouter.initialize")
-        pass
-
-    async def shutdown(self) -> None:
-        logger.debug("DatasetIORouter.shutdown")
-        pass
-
-    async def register_dataset(
-        self,
-        purpose: DatasetPurpose,
-        source: DataSource,
-        metadata: dict[str, Any] | None = None,
-        dataset_id: str | None = None,
-    ) -> None:
-        logger.debug(
-            f"DatasetIORouter.register_dataset: {purpose=} {source=} {metadata=} {dataset_id=}",
-        )
-        await self.routing_table.register_dataset(
-            purpose=purpose,
-            source=source,
-            metadata=metadata,
-            dataset_id=dataset_id,
-        )
-
-    async def iterrows(
-        self,
-        dataset_id: str,
-        start_index: int | None = None,
-        limit: int | None = None,
-    ) -> PaginatedResponse:
-        logger.debug(
-            f"DatasetIORouter.iterrows: {dataset_id}, {start_index=} {limit=}",
-        )
-        return await self.routing_table.get_provider_impl(dataset_id).iterrows(
-            dataset_id=dataset_id,
-            start_index=start_index,
-            limit=limit,
-        )
-
-    async def append_rows(self, dataset_id: str, rows: list[dict[str, Any]]) -> None:
-        logger.debug(f"DatasetIORouter.append_rows: {dataset_id}, {len(rows)} rows")
-        return await self.routing_table.get_provider_impl(dataset_id).append_rows(
-            dataset_id=dataset_id,
-            rows=rows,
-        )
-
-
-class ScoringRouter(Scoring):
-    def __init__(
-        self,
-        routing_table: RoutingTable,
-    ) -> None:
-        logger.debug("Initializing ScoringRouter")
-        self.routing_table = routing_table
-
-    async def initialize(self) -> None:
-        logger.debug("ScoringRouter.initialize")
-        pass
-
-    async def shutdown(self) -> None:
-        logger.debug("ScoringRouter.shutdown")
-        pass
-
-    async def score_batch(
-        self,
-        dataset_id: str,
-        scoring_functions: dict[str, ScoringFnParams | None] = None,
-        save_results_dataset: bool = False,
-    ) -> ScoreBatchResponse:
-        logger.debug(f"ScoringRouter.score_batch: {dataset_id}")
-        res = {}
-        for fn_identifier in scoring_functions.keys():
-            score_response = await self.routing_table.get_provider_impl(fn_identifier).score_batch(
-                dataset_id=dataset_id,
-                scoring_functions={fn_identifier: scoring_functions[fn_identifier]},
-            )
-            res.update(score_response.results)
-
-        if save_results_dataset:
-            raise NotImplementedError("Save results dataset not implemented yet")
-
-        return ScoreBatchResponse(
-            results=res,
-        )
-
-    async def score(
-        self,
-        input_rows: list[dict[str, Any]],
-        scoring_functions: dict[str, ScoringFnParams | None] = None,
-    ) -> ScoreResponse:
-        logger.debug(f"ScoringRouter.score: {len(input_rows)} rows, {len(scoring_functions)} functions")
-        res = {}
-        # look up and map each scoring function to its provider impl
-        for fn_identifier in scoring_functions.keys():
-            score_response = await self.routing_table.get_provider_impl(fn_identifier).score(
-                input_rows=input_rows,
-                scoring_functions={fn_identifier: scoring_functions[fn_identifier]},
-            )
-            res.update(score_response.results)
-
-        return ScoreResponse(results=res)
-
-
-class EvalRouter(Eval):
-    def __init__(
-        self,
-        routing_table: RoutingTable,
-    ) -> None:
-        logger.debug("Initializing EvalRouter")
-        self.routing_table = routing_table
-
-    async def initialize(self) -> None:
-        logger.debug("EvalRouter.initialize")
-        pass
-
-    async def shutdown(self) -> None:
-        logger.debug("EvalRouter.shutdown")
-        pass
-
-    async def run_eval(
-        self,
-        benchmark_id: str,
-        benchmark_config: BenchmarkConfig,
-    ) -> Job:
-        logger.debug(f"EvalRouter.run_eval: {benchmark_id}")
-        return await self.routing_table.get_provider_impl(benchmark_id).run_eval(
-            benchmark_id=benchmark_id,
-            benchmark_config=benchmark_config,
-        )
-
-    async def evaluate_rows(
-        self,
-        benchmark_id: str,
-        input_rows: list[dict[str, Any]],
-        scoring_functions: list[str],
-        benchmark_config: BenchmarkConfig,
-    ) -> EvaluateResponse:
-        logger.debug(f"EvalRouter.evaluate_rows: {benchmark_id}, {len(input_rows)} rows")
-        return await self.routing_table.get_provider_impl(benchmark_id).evaluate_rows(
-            benchmark_id=benchmark_id,
-            input_rows=input_rows,
-            scoring_functions=scoring_functions,
-            benchmark_config=benchmark_config,
-        )
-
-    async def job_status(
-        self,
-        benchmark_id: str,
-        job_id: str,
-    ) -> Job:
-        logger.debug(f"EvalRouter.job_status: {benchmark_id}, {job_id}")
-        return await self.routing_table.get_provider_impl(benchmark_id).job_status(benchmark_id, job_id)
-
-    async def job_cancel(
-        self,
-        benchmark_id: str,
-        job_id: str,
-    ) -> None:
-        logger.debug(f"EvalRouter.job_cancel: {benchmark_id}, {job_id}")
-        await self.routing_table.get_provider_impl(benchmark_id).job_cancel(
-            benchmark_id,
-            job_id,
-        )
-
-    async def job_result(
-        self,
-        benchmark_id: str,
-        job_id: str,
-    ) -> EvaluateResponse:
-        logger.debug(f"EvalRouter.job_result: {benchmark_id}, {job_id}")
-        return await self.routing_table.get_provider_impl(benchmark_id).job_result(
-            benchmark_id,
-            job_id,
-        )
-
-
-class ToolRuntimeRouter(ToolRuntime):
-    class RagToolImpl(RAGToolRuntime):
-        def __init__(
-            self,
-            routing_table: RoutingTable,
-        ) -> None:
-            logger.debug("Initializing ToolRuntimeRouter.RagToolImpl")
-            self.routing_table = routing_table
-
-        async def query(
-            self,
-            content: InterleavedContent,
-            vector_db_ids: list[str],
-            query_config: RAGQueryConfig | None = None,
-        ) -> RAGQueryResult:
-            logger.debug(f"ToolRuntimeRouter.RagToolImpl.query: {vector_db_ids}")
-            return await self.routing_table.get_provider_impl("knowledge_search").query(
-                content, vector_db_ids, query_config
-            )
-
-        async def insert(
-            self,
-            documents: list[RAGDocument],
-            vector_db_id: str,
-            chunk_size_in_tokens: int = 512,
-        ) -> None:
-            logger.debug(
-                f"ToolRuntimeRouter.RagToolImpl.insert: {vector_db_id}, {len(documents)} documents, chunk_size={chunk_size_in_tokens}"
-            )
-            return await self.routing_table.get_provider_impl("insert_into_memory").insert(
-                documents, vector_db_id, chunk_size_in_tokens
-            )
-
-    def __init__(
-        self,
-        routing_table: RoutingTable,
-    ) -> None:
-        logger.debug("Initializing ToolRuntimeRouter")
-        self.routing_table = routing_table
-
-        # HACK ALERT this should be in sync with "get_all_api_endpoints()"
-        self.rag_tool = self.RagToolImpl(routing_table)
-        for method in ("query", "insert"):
-            setattr(self, f"rag_tool.{method}", getattr(self.rag_tool, method))
-
-    async def initialize(self) -> None:
-        logger.debug("ToolRuntimeRouter.initialize")
-        pass
-
-    async def shutdown(self) -> None:
-        logger.debug("ToolRuntimeRouter.shutdown")
-        pass
-
-    async def invoke_tool(self, tool_name: str, kwargs: dict[str, Any]) -> Any:
-        logger.debug(f"ToolRuntimeRouter.invoke_tool: {tool_name}")
-        return await self.routing_table.get_provider_impl(tool_name).invoke_tool(
-            tool_name=tool_name,
-            kwargs=kwargs,
-        )
-
-    async def list_runtime_tools(
-        self, tool_group_id: str | None = None, mcp_endpoint: URL | None = None
-    ) -> ListToolDefsResponse:
-        logger.debug(f"ToolRuntimeRouter.list_runtime_tools: {tool_group_id}")
-        return await self.routing_table.get_provider_impl(tool_group_id).list_tools(tool_group_id, mcp_endpoint)

llama_stack/distribution/routers/routing_tables.py

@@ -1,634 +0,0 @@
-# Copyright (c) Meta Platforms, Inc. and affiliates.
-# All rights reserved.
-#
-# This source code is licensed under the terms described in the LICENSE file in
-# the root directory of this source tree.
-
-import logging
-import time
-import uuid
-from typing import Any
-
-from pydantic import TypeAdapter
-
-from llama_stack.apis.benchmarks import Benchmark, Benchmarks, ListBenchmarksResponse
-from llama_stack.apis.common.content_types import URL
-from llama_stack.apis.common.type_system import ParamType
-from llama_stack.apis.datasets import (
-    Dataset,
-    DatasetPurpose,
-    Datasets,
-    DatasetType,
-    DataSource,
-    ListDatasetsResponse,
-    RowsDataSource,
-    URIDataSource,
-)
-from llama_stack.apis.models import ListModelsResponse, Model, Models, ModelType, OpenAIListModelsResponse, OpenAIModel
-from llama_stack.apis.resource import ResourceType
-from llama_stack.apis.scoring_functions import (
-    ListScoringFunctionsResponse,
-    ScoringFn,
-    ScoringFnParams,
-    ScoringFunctions,
-)
-from llama_stack.apis.shields import ListShieldsResponse, Shield, Shields
-from llama_stack.apis.tools import (
-    ListToolGroupsResponse,
-    ListToolsResponse,
-    Tool,
-    ToolGroup,
-    ToolGroups,
-    ToolHost,
-)
-from llama_stack.apis.vector_dbs import ListVectorDBsResponse, VectorDB, VectorDBs
-from llama_stack.distribution.access_control import check_access
-from llama_stack.distribution.datatypes import (
-    AccessAttributes,
-    BenchmarkWithACL,
-    DatasetWithACL,
-    ModelWithACL,
-    RoutableObject,
-    RoutableObjectWithProvider,
-    RoutedProtocol,
-    ScoringFnWithACL,
-    ShieldWithACL,
-    ToolGroupWithACL,
-    ToolWithACL,
-    VectorDBWithACL,
-)
-from llama_stack.distribution.request_headers import get_auth_attributes
-from llama_stack.distribution.store import DistributionRegistry
-from llama_stack.providers.datatypes import Api, RoutingTable
-
-logger = logging.getLogger(__name__)
-
-
-def get_impl_api(p: Any) -> Api:
-    return p.__provider_spec__.api
-
-
-# TODO: this should return the registered object for all APIs
-async def register_object_with_provider(obj: RoutableObject, p: Any) -> RoutableObject:
-    api = get_impl_api(p)
-
-    assert obj.provider_id != "remote", "Remote provider should not be registered"
-
-    if api == Api.inference:
-        return await p.register_model(obj)
-    elif api == Api.safety:
-        return await p.register_shield(obj)
-    elif api == Api.vector_io:
-        return await p.register_vector_db(obj)
-    elif api == Api.datasetio:
-        return await p.register_dataset(obj)
-    elif api == Api.scoring:
-        return await p.register_scoring_function(obj)
-    elif api == Api.eval:
-        return await p.register_benchmark(obj)
-    elif api == Api.tool_runtime:
-        return await p.register_tool(obj)
-    else:
-        raise ValueError(f"Unknown API {api} for registering object with provider")
-
-
-async def unregister_object_from_provider(obj: RoutableObject, p: Any) -> None:
-    api = get_impl_api(p)
-    if api == Api.vector_io:
-        return await p.unregister_vector_db(obj.identifier)
-    elif api == Api.inference:
-        return await p.unregister_model(obj.identifier)
-    elif api == Api.datasetio:
-        return await p.unregister_dataset(obj.identifier)
-    elif api == Api.tool_runtime:
-        return await p.unregister_tool(obj.identifier)
-    else:
-        raise ValueError(f"Unregister not supported for {api}")
-
-
-Registry = dict[str, list[RoutableObjectWithProvider]]
-
-
-class CommonRoutingTableImpl(RoutingTable):
-    def __init__(
-        self,
-        impls_by_provider_id: dict[str, RoutedProtocol],
-        dist_registry: DistributionRegistry,
-    ) -> None:
-        self.impls_by_provider_id = impls_by_provider_id
-        self.dist_registry = dist_registry
-
-    async def initialize(self) -> None:
-        async def add_objects(objs: list[RoutableObjectWithProvider], provider_id: str, cls) -> None:
-            for obj in objs:
-                if cls is None:
-                    obj.provider_id = provider_id
-                else:
-                    # Create a copy of the model data and explicitly set provider_id
-                    model_data = obj.model_dump()
-                    model_data["provider_id"] = provider_id
-                    obj = cls(**model_data)
-                await self.dist_registry.register(obj)
-
-        # Register all objects from providers
-        for pid, p in self.impls_by_provider_id.items():
-            api = get_impl_api(p)
-            if api == Api.inference:
-                p.model_store = self
-            elif api == Api.safety:
-                p.shield_store = self
-            elif api == Api.vector_io:
-                p.vector_db_store = self
-            elif api == Api.datasetio:
-                p.dataset_store = self
-            elif api == Api.scoring:
-                p.scoring_function_store = self
-                scoring_functions = await p.list_scoring_functions()
-                await add_objects(scoring_functions, pid, ScoringFn)
-            elif api == Api.eval:
-                p.benchmark_store = self
-            elif api == Api.tool_runtime:
-                p.tool_store = self
-
-    async def shutdown(self) -> None:
-        for p in self.impls_by_provider_id.values():
-            await p.shutdown()
-
-    def get_provider_impl(self, routing_key: str, provider_id: str | None = None) -> Any:
-        def apiname_object():
-            if isinstance(self, ModelsRoutingTable):
-                return ("Inference", "model")
-            elif isinstance(self, ShieldsRoutingTable):
-                return ("Safety", "shield")
-            elif isinstance(self, VectorDBsRoutingTable):
-                return ("VectorIO", "vector_db")
-            elif isinstance(self, DatasetsRoutingTable):
-                return ("DatasetIO", "dataset")
-            elif isinstance(self, ScoringFunctionsRoutingTable):
-                return ("Scoring", "scoring_function")
-            elif isinstance(self, BenchmarksRoutingTable):
-                return ("Eval", "benchmark")
-            elif isinstance(self, ToolGroupsRoutingTable):
-                return ("Tools", "tool")
-            else:
-                raise ValueError("Unknown routing table type")
-
-        apiname, objtype = apiname_object()
-
-        # Get objects from disk registry
-        obj = self.dist_registry.get_cached(objtype, routing_key)
-        if not obj:
-            provider_ids = list(self.impls_by_provider_id.keys())
-            if len(provider_ids) > 1:
-                provider_ids_str = f"any of the providers: {', '.join(provider_ids)}"
-            else:
-                provider_ids_str = f"provider: `{provider_ids[0]}`"
-            raise ValueError(
-                f"{objtype.capitalize()} `{routing_key}` not served by {provider_ids_str}. Make sure there is an {apiname} provider serving this {objtype}."
-            )
-
-        if not provider_id or provider_id == obj.provider_id:
-            return self.impls_by_provider_id[obj.provider_id]
-
-        raise ValueError(f"Provider not found for `{routing_key}`")
-
-    async def get_object_by_identifier(self, type: str, identifier: str) -> RoutableObjectWithProvider | None:
-        # Get from disk registry
-        obj = await self.dist_registry.get(type, identifier)
-        if not obj:
-            return None
-
-        # Check if user has permission to access this object
-        if not check_access(obj.identifier, getattr(obj, "access_attributes", None), get_auth_attributes()):
-            logger.debug(f"Access denied to {type} '{identifier}' based on attribute mismatch")
-            return None
-
-        return obj
-
-    async def unregister_object(self, obj: RoutableObjectWithProvider) -> None:
-        await self.dist_registry.delete(obj.type, obj.identifier)
-        await unregister_object_from_provider(obj, self.impls_by_provider_id[obj.provider_id])
-
-    async def register_object(self, obj: RoutableObjectWithProvider) -> RoutableObjectWithProvider:
-        # if provider_id is not specified, pick an arbitrary one from existing entries
-        if not obj.provider_id and len(self.impls_by_provider_id) > 0:
-            obj.provider_id = list(self.impls_by_provider_id.keys())[0]
-
-        if obj.provider_id not in self.impls_by_provider_id:
-            raise ValueError(f"Provider `{obj.provider_id}` not found")
-
-        p = self.impls_by_provider_id[obj.provider_id]
-
-        # If object supports access control but no attributes set, use creator's attributes
-        if not obj.access_attributes:
-            creator_attributes = get_auth_attributes()
-            if creator_attributes:
-                obj.access_attributes = AccessAttributes(**creator_attributes)
-                logger.info(f"Setting access attributes for {obj.type} '{obj.identifier}' based on creator's identity")
-
-        registered_obj = await register_object_with_provider(obj, p)
-        # TODO: This needs to be fixed for all APIs once they return the registered object
-        if obj.type == ResourceType.model.value:
-            await self.dist_registry.register(registered_obj)
-            return registered_obj
-
-        else:
-            await self.dist_registry.register(obj)
-            return obj
-
-    async def get_all_with_type(self, type: str) -> list[RoutableObjectWithProvider]:
-        objs = await self.dist_registry.get_all()
-        filtered_objs = [obj for obj in objs if obj.type == type]
-
-        # Apply attribute-based access control filtering
-        if filtered_objs:
-            filtered_objs = [
-                obj
-                for obj in filtered_objs
-                if check_access(obj.identifier, getattr(obj, "access_attributes", None), get_auth_attributes())
-            ]
-
-        return filtered_objs
-
-
-class ModelsRoutingTable(CommonRoutingTableImpl, Models):
-    async def list_models(self) -> ListModelsResponse:
-        return ListModelsResponse(data=await self.get_all_with_type("model"))
-
-    async def openai_list_models(self) -> OpenAIListModelsResponse:
-        models = await self.get_all_with_type("model")
-        openai_models = [
-            OpenAIModel(
-                id=model.identifier,
-                object="model",
-                created=int(time.time()),
-                owned_by="llama_stack",
-            )
-            for model in models
-        ]
-        return OpenAIListModelsResponse(data=openai_models)
-
-    async def get_model(self, model_id: str) -> Model:
-        model = await self.get_object_by_identifier("model", model_id)
-        if model is None:
-            raise ValueError(f"Model '{model_id}' not found")
-        return model
-
-    async def register_model(
-        self,
-        model_id: str,
-        provider_model_id: str | None = None,
-        provider_id: str | None = None,
-        metadata: dict[str, Any] | None = None,
-        model_type: ModelType | None = None,
-    ) -> Model:
-        if provider_model_id is None:
-            provider_model_id = model_id
-        if provider_id is None:
-            # If provider_id not specified, use the only provider if it supports this model
-            if len(self.impls_by_provider_id) == 1:
-                provider_id = list(self.impls_by_provider_id.keys())[0]
-            else:
-                raise ValueError(
-                    f"No provider specified and multiple providers available. Please specify a provider_id. Available providers: {self.impls_by_provider_id.keys()}"
-                )
-        if metadata is None:
-            metadata = {}
-        if model_type is None:
-            model_type = ModelType.llm
-        if "embedding_dimension" not in metadata and model_type == ModelType.embedding:
-            raise ValueError("Embedding model must have an embedding dimension in its metadata")
-        model = ModelWithACL(
-            identifier=model_id,
-            provider_resource_id=provider_model_id,
-            provider_id=provider_id,
-            metadata=metadata,
-            model_type=model_type,
-        )
-        registered_model = await self.register_object(model)
-        return registered_model
-
-    async def unregister_model(self, model_id: str) -> None:
-        existing_model = await self.get_model(model_id)
-        if existing_model is None:
-            raise ValueError(f"Model {model_id} not found")
-        await self.unregister_object(existing_model)
-
-
-class ShieldsRoutingTable(CommonRoutingTableImpl, Shields):
-    async def list_shields(self) -> ListShieldsResponse:
-        return ListShieldsResponse(data=await self.get_all_with_type(ResourceType.shield.value))
-
-    async def get_shield(self, identifier: str) -> Shield:
-        shield = await self.get_object_by_identifier("shield", identifier)
-        if shield is None:
-            raise ValueError(f"Shield '{identifier}' not found")
-        return shield
-
-    async def register_shield(
-        self,
-        shield_id: str,
-        provider_shield_id: str | None = None,
-        provider_id: str | None = None,
-        params: dict[str, Any] | None = None,
-    ) -> Shield:
-        if provider_shield_id is None:
-            provider_shield_id = shield_id
-        if provider_id is None:
-            # If provider_id not specified, use the only provider if it supports this shield type
-            if len(self.impls_by_provider_id) == 1:
-                provider_id = list(self.impls_by_provider_id.keys())[0]
-            else:
-                raise ValueError(
-                    "No provider specified and multiple providers available. Please specify a provider_id."
-                )
-        if params is None:
-            params = {}
-        shield = ShieldWithACL(
-            identifier=shield_id,
-            provider_resource_id=provider_shield_id,
-            provider_id=provider_id,
-            params=params,
-        )
-        await self.register_object(shield)
-        return shield
-
-
-class VectorDBsRoutingTable(CommonRoutingTableImpl, VectorDBs):
-    async def list_vector_dbs(self) -> ListVectorDBsResponse:
-        return ListVectorDBsResponse(data=await self.get_all_with_type("vector_db"))
-
-    async def get_vector_db(self, vector_db_id: str) -> VectorDB:
-        vector_db = await self.get_object_by_identifier("vector_db", vector_db_id)
-        if vector_db is None:
-            raise ValueError(f"Vector DB '{vector_db_id}' not found")
-        return vector_db
-
-    async def register_vector_db(
-        self,
-        vector_db_id: str,
-        embedding_model: str,
-        embedding_dimension: int | None = 384,
-        provider_id: str | None = None,
-        provider_vector_db_id: str | None = None,
-    ) -> VectorDB:
-        if provider_vector_db_id is None:
-            provider_vector_db_id = vector_db_id
-        if provider_id is None:
-            if len(self.impls_by_provider_id) > 0:
-                provider_id = list(self.impls_by_provider_id.keys())[0]
-                if len(self.impls_by_provider_id) > 1:
-                    logger.warning(
-                        f"No provider specified and multiple providers available. Arbitrarily selected the first provider {provider_id}."
-                    )
-            else:
-                raise ValueError("No provider available. Please configure a vector_io provider.")
-        model = await self.get_object_by_identifier("model", embedding_model)
-        if model is None:
-            raise ValueError(f"Model {embedding_model} not found")
-        if model.model_type != ModelType.embedding:
-            raise ValueError(f"Model {embedding_model} is not an embedding model")
-        if "embedding_dimension" not in model.metadata:
-            raise ValueError(f"Model {embedding_model} does not have an embedding dimension")
-        vector_db_data = {
-            "identifier": vector_db_id,
-            "type": ResourceType.vector_db.value,
-            "provider_id": provider_id,
-            "provider_resource_id": provider_vector_db_id,
-            "embedding_model": embedding_model,
-            "embedding_dimension": model.metadata["embedding_dimension"],
-        }
-        vector_db = TypeAdapter(VectorDBWithACL).validate_python(vector_db_data)
-        await self.register_object(vector_db)
-        return vector_db
-
-    async def unregister_vector_db(self, vector_db_id: str) -> None:
-        existing_vector_db = await self.get_vector_db(vector_db_id)
-        if existing_vector_db is None:
-            raise ValueError(f"Vector DB {vector_db_id} not found")
-        await self.unregister_object(existing_vector_db)
-
-
-class DatasetsRoutingTable(CommonRoutingTableImpl, Datasets):
-    async def list_datasets(self) -> ListDatasetsResponse:
-        return ListDatasetsResponse(data=await self.get_all_with_type(ResourceType.dataset.value))
-
-    async def get_dataset(self, dataset_id: str) -> Dataset:
-        dataset = await self.get_object_by_identifier("dataset", dataset_id)
-        if dataset is None:
-            raise ValueError(f"Dataset '{dataset_id}' not found")
-        return dataset
-
-    async def register_dataset(
-        self,
-        purpose: DatasetPurpose,
-        source: DataSource,
-        metadata: dict[str, Any] | None = None,
-        dataset_id: str | None = None,
-    ) -> Dataset:
-        if isinstance(source, dict):
-            if source["type"] == "uri":
-                source = URIDataSource.parse_obj(source)
-            elif source["type"] == "rows":
-                source = RowsDataSource.parse_obj(source)
-
-        if not dataset_id:
-            dataset_id = f"dataset-{str(uuid.uuid4())}"
-
-        provider_dataset_id = dataset_id
-
-        # infer provider from source
-        if metadata:
-            if metadata.get("provider_id"):
-                provider_id = metadata.get("provider_id")  # pass through from nvidia datasetio
-        elif source.type == DatasetType.rows.value:
-            provider_id = "localfs"
-        elif source.type == DatasetType.uri.value:
-            # infer provider from uri
-            if source.uri.startswith("huggingface"):
-                provider_id = "huggingface"
-            else:
-                provider_id = "localfs"
-        else:
-            raise ValueError(f"Unknown data source type: {source.type}")
-
-        if metadata is None:
-            metadata = {}
-
-        dataset = DatasetWithACL(
-            identifier=dataset_id,
-            provider_resource_id=provider_dataset_id,
-            provider_id=provider_id,
-            purpose=purpose,
-            source=source,
-            metadata=metadata,
-        )
-
-        await self.register_object(dataset)
-        return dataset
-
-    async def unregister_dataset(self, dataset_id: str) -> None:
-        dataset = await self.get_dataset(dataset_id)
-        if dataset is None:
-            raise ValueError(f"Dataset {dataset_id} not found")
-        await self.unregister_object(dataset)
-
-
-class ScoringFunctionsRoutingTable(CommonRoutingTableImpl, ScoringFunctions):
-    async def list_scoring_functions(self) -> ListScoringFunctionsResponse:
-        return ListScoringFunctionsResponse(data=await self.get_all_with_type(ResourceType.scoring_function.value))
-
-    async def get_scoring_function(self, scoring_fn_id: str) -> ScoringFn:
-        scoring_fn = await self.get_object_by_identifier("scoring_function", scoring_fn_id)
-        if scoring_fn is None:
-            raise ValueError(f"Scoring function '{scoring_fn_id}' not found")
-        return scoring_fn
-
-    async def register_scoring_function(
-        self,
-        scoring_fn_id: str,
-        description: str,
-        return_type: ParamType,
-        provider_scoring_fn_id: str | None = None,
-        provider_id: str | None = None,
-        params: ScoringFnParams | None = None,
-    ) -> None:
-        if provider_scoring_fn_id is None:
-            provider_scoring_fn_id = scoring_fn_id
-        if provider_id is None:
-            if len(self.impls_by_provider_id) == 1:
-                provider_id = list(self.impls_by_provider_id.keys())[0]
-            else:
-                raise ValueError(
-                    "No provider specified and multiple providers available. Please specify a provider_id."
-                )
-        scoring_fn = ScoringFnWithACL(
-            identifier=scoring_fn_id,
-            description=description,
-            return_type=return_type,
-            provider_resource_id=provider_scoring_fn_id,
-            provider_id=provider_id,
-            params=params,
-        )
-        scoring_fn.provider_id = provider_id
-        await self.register_object(scoring_fn)
-
-
-class BenchmarksRoutingTable(CommonRoutingTableImpl, Benchmarks):
-    async def list_benchmarks(self) -> ListBenchmarksResponse:
-        return ListBenchmarksResponse(data=await self.get_all_with_type("benchmark"))
-
-    async def get_benchmark(self, benchmark_id: str) -> Benchmark:
-        benchmark = await self.get_object_by_identifier("benchmark", benchmark_id)
-        if benchmark is None:
-            raise ValueError(f"Benchmark '{benchmark_id}' not found")
-        return benchmark
-
-    async def register_benchmark(
-        self,
-        benchmark_id: str,
-        dataset_id: str,
-        scoring_functions: list[str],
-        metadata: dict[str, Any] | None = None,
-        provider_benchmark_id: str | None = None,
-        provider_id: str | None = None,
-    ) -> None:
-        if metadata is None:
-            metadata = {}
-        if provider_id is None:
-            if len(self.impls_by_provider_id) == 1:
-                provider_id = list(self.impls_by_provider_id.keys())[0]
-            else:
-                raise ValueError(
-                    "No provider specified and multiple providers available. Please specify a provider_id."
-                )
-        if provider_benchmark_id is None:
-            provider_benchmark_id = benchmark_id
-        benchmark = BenchmarkWithACL(
-            identifier=benchmark_id,
-            dataset_id=dataset_id,
-            scoring_functions=scoring_functions,
-            metadata=metadata,
-            provider_id=provider_id,
-            provider_resource_id=provider_benchmark_id,
-        )
-        await self.register_object(benchmark)
-
-
-class ToolGroupsRoutingTable(CommonRoutingTableImpl, ToolGroups):
-    async def list_tools(self, toolgroup_id: str | None = None) -> ListToolsResponse:
-        tools = await self.get_all_with_type("tool")
-        if toolgroup_id:
-            tools = [tool for tool in tools if tool.toolgroup_id == toolgroup_id]
-        return ListToolsResponse(data=tools)
-
-    async def list_tool_groups(self) -> ListToolGroupsResponse:
-        return ListToolGroupsResponse(data=await self.get_all_with_type("tool_group"))
-
-    async def get_tool_group(self, toolgroup_id: str) -> ToolGroup:
-        tool_group = await self.get_object_by_identifier("tool_group", toolgroup_id)
-        if tool_group is None:
-            raise ValueError(f"Tool group '{toolgroup_id}' not found")
-        return tool_group
-
-    async def get_tool(self, tool_name: str) -> Tool:
-        return await self.get_object_by_identifier("tool", tool_name)
-
-    async def register_tool_group(
-        self,
-        toolgroup_id: str,
-        provider_id: str,
-        mcp_endpoint: URL | None = None,
-        args: dict[str, Any] | None = None,
-    ) -> None:
-        tools = []
-        tool_defs = await self.impls_by_provider_id[provider_id].list_runtime_tools(toolgroup_id, mcp_endpoint)
-        tool_host = ToolHost.model_context_protocol if mcp_endpoint else ToolHost.distribution
-
-        for tool_def in tool_defs.data:
-            tools.append(
-                ToolWithACL(
-                    identifier=tool_def.name,
-                    toolgroup_id=toolgroup_id,
-                    description=tool_def.description or "",
-                    parameters=tool_def.parameters or [],
-                    provider_id=provider_id,
-                    provider_resource_id=tool_def.name,
-                    metadata=tool_def.metadata,
-                    tool_host=tool_host,
-                )
-            )
-        for tool in tools:
-            existing_tool = await self.get_tool(tool.identifier)
-            # Compare existing and new object if one exists
-            if existing_tool:
-                existing_dict = existing_tool.model_dump()
-                new_dict = tool.model_dump()
-
-                if existing_dict != new_dict:
-                    raise ValueError(
-                        f"Object {tool.identifier} already exists in registry. Please use a different identifier."
-                    )
-            await self.register_object(tool)
-
-        await self.dist_registry.register(
-            ToolGroupWithACL(
-                identifier=toolgroup_id,
-                provider_id=provider_id,
-                provider_resource_id=toolgroup_id,
-                mcp_endpoint=mcp_endpoint,
-                args=args,
-            )
-        )
-
-    async def unregister_toolgroup(self, toolgroup_id: str) -> None:
-        tool_group = await self.get_tool_group(toolgroup_id)
-        if tool_group is None:
-            raise ValueError(f"Tool group {toolgroup_id} not found")
-        tools = await self.list_tools(toolgroup_id)
-        for tool in getattr(tools, "data", []):
-            await self.unregister_object(tool)
-        await self.unregister_object(tool_group)
-
-    async def shutdown(self) -> None:
-        pass

llama_stack/distribution/routers/safety.py

@@ -0,0 +1,57 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the terms described in the LICENSE file in
+# the root directory of this source tree.
+
+from typing import Any
+
+from llama_stack.apis.inference import (
+    Message,
+)
+from llama_stack.apis.safety import RunShieldResponse, Safety
+from llama_stack.apis.shields import Shield
+from llama_stack.log import get_logger
+from llama_stack.providers.datatypes import RoutingTable
+
+logger = get_logger(name=__name__, category="core")
+
+
+class SafetyRouter(Safety):
+    def __init__(
+        self,
+        routing_table: RoutingTable,
+    ) -> None:
+        logger.debug("Initializing SafetyRouter")
+        self.routing_table = routing_table
+
+    async def initialize(self) -> None:
+        logger.debug("SafetyRouter.initialize")
+        pass
+
+    async def shutdown(self) -> None:
+        logger.debug("SafetyRouter.shutdown")
+        pass
+
+    async def register_shield(
+        self,
+        shield_id: str,
+        provider_shield_id: str | None = None,
+        provider_id: str | None = None,
+        params: dict[str, Any] | None = None,
+    ) -> Shield:
+        logger.debug(f"SafetyRouter.register_shield: {shield_id}")
+        return await self.routing_table.register_shield(shield_id, provider_shield_id, provider_id, params)
+
+    async def run_shield(
+        self,
+        shield_id: str,
+        messages: list[Message],
+        params: dict[str, Any] = None,
+    ) -> RunShieldResponse:
+        logger.debug(f"SafetyRouter.run_shield: {shield_id}")
+        return await self.routing_table.get_provider_impl(shield_id).run_shield(
+            shield_id=shield_id,
+            messages=messages,
+            params=params,
+        )

llama_stack/distribution/routers/tool_runtime.py

@@ -0,0 +1,92 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the terms described in the LICENSE file in
+# the root directory of this source tree.
+
+from typing import Any
+
+from llama_stack.apis.common.content_types import (
+    URL,
+    InterleavedContent,
+)
+from llama_stack.apis.tools import (
+    ListToolsResponse,
+    RAGDocument,
+    RAGQueryConfig,
+    RAGQueryResult,
+    RAGToolRuntime,
+    ToolRuntime,
+)
+from llama_stack.log import get_logger
+
+from ..routing_tables.toolgroups import ToolGroupsRoutingTable
+
+logger = get_logger(name=__name__, category="core")
+
+
+class ToolRuntimeRouter(ToolRuntime):
+    class RagToolImpl(RAGToolRuntime):
+        def __init__(
+            self,
+            routing_table: ToolGroupsRoutingTable,
+        ) -> None:
+            logger.debug("Initializing ToolRuntimeRouter.RagToolImpl")
+            self.routing_table = routing_table
+
+        async def query(
+            self,
+            content: InterleavedContent,
+            vector_db_ids: list[str],
+            query_config: RAGQueryConfig | None = None,
+        ) -> RAGQueryResult:
+            logger.debug(f"ToolRuntimeRouter.RagToolImpl.query: {vector_db_ids}")
+            return await self.routing_table.get_provider_impl("knowledge_search").query(
+                content, vector_db_ids, query_config
+            )
+
+        async def insert(
+            self,
+            documents: list[RAGDocument],
+            vector_db_id: str,
+            chunk_size_in_tokens: int = 512,
+        ) -> None:
+            logger.debug(
+                f"ToolRuntimeRouter.RagToolImpl.insert: {vector_db_id}, {len(documents)} documents, chunk_size={chunk_size_in_tokens}"
+            )
+            return await self.routing_table.get_provider_impl("insert_into_memory").insert(
+                documents, vector_db_id, chunk_size_in_tokens
+            )
+
+    def __init__(
+        self,
+        routing_table: ToolGroupsRoutingTable,
+    ) -> None:
+        logger.debug("Initializing ToolRuntimeRouter")
+        self.routing_table = routing_table
+
+        # HACK ALERT this should be in sync with "get_all_api_endpoints()"
+        self.rag_tool = self.RagToolImpl(routing_table)
+        for method in ("query", "insert"):
+            setattr(self, f"rag_tool.{method}", getattr(self.rag_tool, method))
+
+    async def initialize(self) -> None:
+        logger.debug("ToolRuntimeRouter.initialize")
+        pass
+
+    async def shutdown(self) -> None:
+        logger.debug("ToolRuntimeRouter.shutdown")
+        pass
+
+    async def invoke_tool(self, tool_name: str, kwargs: dict[str, Any]) -> Any:
+        logger.debug(f"ToolRuntimeRouter.invoke_tool: {tool_name}")
+        return await self.routing_table.get_provider_impl(tool_name).invoke_tool(
+            tool_name=tool_name,
+            kwargs=kwargs,
+        )
+
+    async def list_runtime_tools(
+        self, tool_group_id: str | None = None, mcp_endpoint: URL | None = None
+    ) -> ListToolsResponse:
+        logger.debug(f"ToolRuntimeRouter.list_runtime_tools: {tool_group_id}")
+        return await self.routing_table.list_tools(tool_group_id)

llama_stack/distribution/routers/vector_io.py

@@ -0,0 +1,72 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the terms described in the LICENSE file in
+# the root directory of this source tree.
+
+from typing import Any
+
+from llama_stack.apis.common.content_types import (
+    InterleavedContent,
+)
+from llama_stack.apis.vector_io import Chunk, QueryChunksResponse, VectorIO
+from llama_stack.log import get_logger
+from llama_stack.providers.datatypes import RoutingTable
+
+logger = get_logger(name=__name__, category="core")
+
+
+class VectorIORouter(VectorIO):
+    """Routes to an provider based on the vector db identifier"""
+
+    def __init__(
+        self,
+        routing_table: RoutingTable,
+    ) -> None:
+        logger.debug("Initializing VectorIORouter")
+        self.routing_table = routing_table
+
+    async def initialize(self) -> None:
+        logger.debug("VectorIORouter.initialize")
+        pass
+
+    async def shutdown(self) -> None:
+        logger.debug("VectorIORouter.shutdown")
+        pass
+
+    async def register_vector_db(
+        self,
+        vector_db_id: str,
+        embedding_model: str,
+        embedding_dimension: int | None = 384,
+        provider_id: str | None = None,
+        provider_vector_db_id: str | None = None,
+    ) -> None:
+        logger.debug(f"VectorIORouter.register_vector_db: {vector_db_id}, {embedding_model}")
+        await self.routing_table.register_vector_db(
+            vector_db_id,
+            embedding_model,
+            embedding_dimension,
+            provider_id,
+            provider_vector_db_id,
+        )
+
+    async def insert_chunks(
+        self,
+        vector_db_id: str,
+        chunks: list[Chunk],
+        ttl_seconds: int | None = None,
+    ) -> None:
+        logger.debug(
+            f"VectorIORouter.insert_chunks: {vector_db_id}, {len(chunks)} chunks, ttl_seconds={ttl_seconds}, chunk_ids={[chunk.metadata['document_id'] for chunk in chunks[:3]]}{' and more...' if len(chunks) > 3 else ''}",
+        )
+        return await self.routing_table.get_provider_impl(vector_db_id).insert_chunks(vector_db_id, chunks, ttl_seconds)
+
+    async def query_chunks(
+        self,
+        vector_db_id: str,
+        query: InterleavedContent,
+        params: dict[str, Any] | None = None,
+    ) -> QueryChunksResponse:
+        logger.debug(f"VectorIORouter.query_chunks: {vector_db_id}")
+        return await self.routing_table.get_provider_impl(vector_db_id).query_chunks(vector_db_id, query, params)

llama_stack/distribution/routing_tables/__init__.py

@@ -0,0 +1,5 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the terms described in the LICENSE file in
+# the root directory of this source tree.

llama_stack/distribution/routing_tables/benchmarks.py

@@ -0,0 +1,58 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the terms described in the LICENSE file in
+# the root directory of this source tree.
+
+from typing import Any
+
+from llama_stack.apis.benchmarks import Benchmark, Benchmarks, ListBenchmarksResponse
+from llama_stack.distribution.datatypes import (
+    BenchmarkWithACL,
+)
+from llama_stack.log import get_logger
+
+from .common import CommonRoutingTableImpl
+
+logger = get_logger(name=__name__, category="core")
+
+
+class BenchmarksRoutingTable(CommonRoutingTableImpl, Benchmarks):
+    async def list_benchmarks(self) -> ListBenchmarksResponse:
+        return ListBenchmarksResponse(data=await self.get_all_with_type("benchmark"))
+
+    async def get_benchmark(self, benchmark_id: str) -> Benchmark:
+        benchmark = await self.get_object_by_identifier("benchmark", benchmark_id)
+        if benchmark is None:
+            raise ValueError(f"Benchmark '{benchmark_id}' not found")
+        return benchmark
+
+    async def register_benchmark(
+        self,
+        benchmark_id: str,
+        dataset_id: str,
+        scoring_functions: list[str],
+        metadata: dict[str, Any] | None = None,
+        provider_benchmark_id: str | None = None,
+        provider_id: str | None = None,
+    ) -> None:
+        if metadata is None:
+            metadata = {}
+        if provider_id is None:
+            if len(self.impls_by_provider_id) == 1:
+                provider_id = list(self.impls_by_provider_id.keys())[0]
+            else:
+                raise ValueError(
+                    "No provider specified and multiple providers available. Please specify a provider_id."
+                )
+        if provider_benchmark_id is None:
+            provider_benchmark_id = benchmark_id
+        benchmark = BenchmarkWithACL(
+            identifier=benchmark_id,
+            dataset_id=dataset_id,
+            scoring_functions=scoring_functions,
+            metadata=metadata,
+            provider_id=provider_id,
+            provider_resource_id=provider_benchmark_id,
+        )
+        await self.register_object(benchmark)

llama_stack/distribution/routing_tables/common.py

@@ -0,0 +1,218 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the terms described in the LICENSE file in
+# the root directory of this source tree.
+
+from typing import Any
+
+from llama_stack.apis.resource import ResourceType
+from llama_stack.apis.scoring_functions import ScoringFn
+from llama_stack.distribution.access_control import check_access
+from llama_stack.distribution.datatypes import (
+    AccessAttributes,
+    RoutableObject,
+    RoutableObjectWithProvider,
+    RoutedProtocol,
+)
+from llama_stack.distribution.request_headers import get_auth_attributes
+from llama_stack.distribution.store import DistributionRegistry
+from llama_stack.log import get_logger
+from llama_stack.providers.datatypes import Api, RoutingTable
+
+logger = get_logger(name=__name__, category="core")
+
+
+def get_impl_api(p: Any) -> Api:
+    return p.__provider_spec__.api
+
+
+# TODO: this should return the registered object for all APIs
+async def register_object_with_provider(obj: RoutableObject, p: Any) -> RoutableObject:
+    api = get_impl_api(p)
+
+    assert obj.provider_id != "remote", "Remote provider should not be registered"
+
+    if api == Api.inference:
+        return await p.register_model(obj)
+    elif api == Api.safety:
+        return await p.register_shield(obj)
+    elif api == Api.vector_io:
+        return await p.register_vector_db(obj)
+    elif api == Api.datasetio:
+        return await p.register_dataset(obj)
+    elif api == Api.scoring:
+        return await p.register_scoring_function(obj)
+    elif api == Api.eval:
+        return await p.register_benchmark(obj)
+    elif api == Api.tool_runtime:
+        return await p.register_toolgroup(obj)
+    else:
+        raise ValueError(f"Unknown API {api} for registering object with provider")
+
+
+async def unregister_object_from_provider(obj: RoutableObject, p: Any) -> None:
+    api = get_impl_api(p)
+    if api == Api.vector_io:
+        return await p.unregister_vector_db(obj.identifier)
+    elif api == Api.inference:
+        return await p.unregister_model(obj.identifier)
+    elif api == Api.datasetio:
+        return await p.unregister_dataset(obj.identifier)
+    elif api == Api.tool_runtime:
+        return await p.unregister_toolgroup(obj.identifier)
+    else:
+        raise ValueError(f"Unregister not supported for {api}")
+
+
+Registry = dict[str, list[RoutableObjectWithProvider]]
+
+
+class CommonRoutingTableImpl(RoutingTable):
+    def __init__(
+        self,
+        impls_by_provider_id: dict[str, RoutedProtocol],
+        dist_registry: DistributionRegistry,
+    ) -> None:
+        self.impls_by_provider_id = impls_by_provider_id
+        self.dist_registry = dist_registry
+
+    async def initialize(self) -> None:
+        async def add_objects(objs: list[RoutableObjectWithProvider], provider_id: str, cls) -> None:
+            for obj in objs:
+                if cls is None:
+                    obj.provider_id = provider_id
+                else:
+                    # Create a copy of the model data and explicitly set provider_id
+                    model_data = obj.model_dump()
+                    model_data["provider_id"] = provider_id
+                    obj = cls(**model_data)
+                await self.dist_registry.register(obj)
+
+        # Register all objects from providers
+        for pid, p in self.impls_by_provider_id.items():
+            api = get_impl_api(p)
+            if api == Api.inference:
+                p.model_store = self
+            elif api == Api.safety:
+                p.shield_store = self
+            elif api == Api.vector_io:
+                p.vector_db_store = self
+            elif api == Api.datasetio:
+                p.dataset_store = self
+            elif api == Api.scoring:
+                p.scoring_function_store = self
+                scoring_functions = await p.list_scoring_functions()
+                await add_objects(scoring_functions, pid, ScoringFn)
+            elif api == Api.eval:
+                p.benchmark_store = self
+            elif api == Api.tool_runtime:
+                p.tool_store = self
+
+    async def shutdown(self) -> None:
+        for p in self.impls_by_provider_id.values():
+            await p.shutdown()
+
+    def get_provider_impl(self, routing_key: str, provider_id: str | None = None) -> Any:
+        from .benchmarks import BenchmarksRoutingTable
+        from .datasets import DatasetsRoutingTable
+        from .models import ModelsRoutingTable
+        from .scoring_functions import ScoringFunctionsRoutingTable
+        from .shields import ShieldsRoutingTable
+        from .toolgroups import ToolGroupsRoutingTable
+        from .vector_dbs import VectorDBsRoutingTable
+
+        def apiname_object():
+            if isinstance(self, ModelsRoutingTable):
+                return ("Inference", "model")
+            elif isinstance(self, ShieldsRoutingTable):
+                return ("Safety", "shield")
+            elif isinstance(self, VectorDBsRoutingTable):
+                return ("VectorIO", "vector_db")
+            elif isinstance(self, DatasetsRoutingTable):
+                return ("DatasetIO", "dataset")
+            elif isinstance(self, ScoringFunctionsRoutingTable):
+                return ("Scoring", "scoring_function")
+            elif isinstance(self, BenchmarksRoutingTable):
+                return ("Eval", "benchmark")
+            elif isinstance(self, ToolGroupsRoutingTable):
+                return ("ToolGroups", "tool_group")
+            else:
+                raise ValueError("Unknown routing table type")
+
+        apiname, objtype = apiname_object()
+
+        # Get objects from disk registry
+        obj = self.dist_registry.get_cached(objtype, routing_key)
+        if not obj:
+            provider_ids = list(self.impls_by_provider_id.keys())
+            if len(provider_ids) > 1:
+                provider_ids_str = f"any of the providers: {', '.join(provider_ids)}"
+            else:
+                provider_ids_str = f"provider: `{provider_ids[0]}`"
+            raise ValueError(
+                f"{objtype.capitalize()} `{routing_key}` not served by {provider_ids_str}. Make sure there is an {apiname} provider serving this {objtype}."
+            )
+
+        if not provider_id or provider_id == obj.provider_id:
+            return self.impls_by_provider_id[obj.provider_id]
+
+        raise ValueError(f"Provider not found for `{routing_key}`")
+
+    async def get_object_by_identifier(self, type: str, identifier: str) -> RoutableObjectWithProvider | None:
+        # Get from disk registry
+        obj = await self.dist_registry.get(type, identifier)
+        if not obj:
+            return None
+
+        # Check if user has permission to access this object
+        if not check_access(obj.identifier, getattr(obj, "access_attributes", None), get_auth_attributes()):
+            logger.debug(f"Access denied to {type} '{identifier}' based on attribute mismatch")
+            return None
+
+        return obj
+
+    async def unregister_object(self, obj: RoutableObjectWithProvider) -> None:
+        await self.dist_registry.delete(obj.type, obj.identifier)
+        await unregister_object_from_provider(obj, self.impls_by_provider_id[obj.provider_id])
+
+    async def register_object(self, obj: RoutableObjectWithProvider) -> RoutableObjectWithProvider:
+        # if provider_id is not specified, pick an arbitrary one from existing entries
+        if not obj.provider_id and len(self.impls_by_provider_id) > 0:
+            obj.provider_id = list(self.impls_by_provider_id.keys())[0]
+
+        if obj.provider_id not in self.impls_by_provider_id:
+            raise ValueError(f"Provider `{obj.provider_id}` not found")
+
+        p = self.impls_by_provider_id[obj.provider_id]
+
+        # If object supports access control but no attributes set, use creator's attributes
+        if not obj.access_attributes:
+            creator_attributes = get_auth_attributes()
+            if creator_attributes:
+                obj.access_attributes = AccessAttributes(**creator_attributes)
+                logger.info(f"Setting access attributes for {obj.type} '{obj.identifier}' based on creator's identity")
+
+        registered_obj = await register_object_with_provider(obj, p)
+        # TODO: This needs to be fixed for all APIs once they return the registered object
+        if obj.type == ResourceType.model.value:
+            await self.dist_registry.register(registered_obj)
+            return registered_obj
+
+        else:
+            await self.dist_registry.register(obj)
+            return obj
+
+    async def get_all_with_type(self, type: str) -> list[RoutableObjectWithProvider]:
+        objs = await self.dist_registry.get_all()
+        filtered_objs = [obj for obj in objs if obj.type == type]
+
+        # Apply attribute-based access control filtering
+        if filtered_objs:
+            filtered_objs = [
+                obj
+                for obj in filtered_objs
+                if check_access(obj.identifier, getattr(obj, "access_attributes", None), get_auth_attributes())
+            ]
+
+        return filtered_objs

llama_stack/distribution/routing_tables/datasets.py

@@ -0,0 +1,93 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the terms described in the LICENSE file in
+# the root directory of this source tree.
+
+import uuid
+from typing import Any
+
+from llama_stack.apis.datasets import (
+    Dataset,
+    DatasetPurpose,
+    Datasets,
+    DatasetType,
+    DataSource,
+    ListDatasetsResponse,
+    RowsDataSource,
+    URIDataSource,
+)
+from llama_stack.apis.resource import ResourceType
+from llama_stack.distribution.datatypes import (
+    DatasetWithACL,
+)
+from llama_stack.log import get_logger
+
+from .common import CommonRoutingTableImpl
+
+logger = get_logger(name=__name__, category="core")
+
+
+class DatasetsRoutingTable(CommonRoutingTableImpl, Datasets):
+    async def list_datasets(self) -> ListDatasetsResponse:
+        return ListDatasetsResponse(data=await self.get_all_with_type(ResourceType.dataset.value))
+
+    async def get_dataset(self, dataset_id: str) -> Dataset:
+        dataset = await self.get_object_by_identifier("dataset", dataset_id)
+        if dataset is None:
+            raise ValueError(f"Dataset '{dataset_id}' not found")
+        return dataset
+
+    async def register_dataset(
+        self,
+        purpose: DatasetPurpose,
+        source: DataSource,
+        metadata: dict[str, Any] | None = None,
+        dataset_id: str | None = None,
+    ) -> Dataset:
+        if isinstance(source, dict):
+            if source["type"] == "uri":
+                source = URIDataSource.parse_obj(source)
+            elif source["type"] == "rows":
+                source = RowsDataSource.parse_obj(source)
+
+        if not dataset_id:
+            dataset_id = f"dataset-{str(uuid.uuid4())}"
+
+        provider_dataset_id = dataset_id
+
+        # infer provider from source
+        if metadata:
+            if metadata.get("provider_id"):
+                provider_id = metadata.get("provider_id")  # pass through from nvidia datasetio
+        elif source.type == DatasetType.rows.value:
+            provider_id = "localfs"
+        elif source.type == DatasetType.uri.value:
+            # infer provider from uri
+            if source.uri.startswith("huggingface"):
+                provider_id = "huggingface"
+            else:
+                provider_id = "localfs"
+        else:
+            raise ValueError(f"Unknown data source type: {source.type}")
+
+        if metadata is None:
+            metadata = {}
+
+        dataset = DatasetWithACL(
+            identifier=dataset_id,
+            provider_resource_id=provider_dataset_id,
+            provider_id=provider_id,
+            purpose=purpose,
+            source=source,
+            metadata=metadata,
+        )
+
+        await self.register_object(dataset)
+        return dataset
+
+    async def unregister_dataset(self, dataset_id: str) -> None:
+        dataset = await self.get_dataset(dataset_id)
+        if dataset is None:
+            raise ValueError(f"Dataset {dataset_id} not found")
+        await self.unregister_object(dataset)

llama_stack/distribution/routing_tables/models.py

@@ -0,0 +1,82 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the terms described in the LICENSE file in
+# the root directory of this source tree.
+
+import time
+from typing import Any
+
+from llama_stack.apis.models import ListModelsResponse, Model, Models, ModelType, OpenAIListModelsResponse, OpenAIModel
+from llama_stack.distribution.datatypes import (
+    ModelWithACL,
+)
+from llama_stack.log import get_logger
+
+from .common import CommonRoutingTableImpl
+
+logger = get_logger(name=__name__, category="core")
+
+
+class ModelsRoutingTable(CommonRoutingTableImpl, Models):
+    async def list_models(self) -> ListModelsResponse:
+        return ListModelsResponse(data=await self.get_all_with_type("model"))
+
+    async def openai_list_models(self) -> OpenAIListModelsResponse:
+        models = await self.get_all_with_type("model")
+        openai_models = [
+            OpenAIModel(
+                id=model.identifier,
+                object="model",
+                created=int(time.time()),
+                owned_by="llama_stack",
+            )
+            for model in models
+        ]
+        return OpenAIListModelsResponse(data=openai_models)
+
+    async def get_model(self, model_id: str) -> Model:
+        model = await self.get_object_by_identifier("model", model_id)
+        if model is None:
+            raise ValueError(f"Model '{model_id}' not found")
+        return model
+
+    async def register_model(
+        self,
+        model_id: str,
+        provider_model_id: str | None = None,
+        provider_id: str | None = None,
+        metadata: dict[str, Any] | None = None,
+        model_type: ModelType | None = None,
+    ) -> Model:
+        if provider_model_id is None:
+            provider_model_id = model_id
+        if provider_id is None:
+            # If provider_id not specified, use the only provider if it supports this model
+            if len(self.impls_by_provider_id) == 1:
+                provider_id = list(self.impls_by_provider_id.keys())[0]
+            else:
+                raise ValueError(
+                    f"No provider specified and multiple providers available. Please specify a provider_id. Available providers: {self.impls_by_provider_id.keys()}"
+                )
+        if metadata is None:
+            metadata = {}
+        if model_type is None:
+            model_type = ModelType.llm
+        if "embedding_dimension" not in metadata and model_type == ModelType.embedding:
+            raise ValueError("Embedding model must have an embedding dimension in its metadata")
+        model = ModelWithACL(
+            identifier=model_id,
+            provider_resource_id=provider_model_id,
+            provider_id=provider_id,
+            metadata=metadata,
+            model_type=model_type,
+        )
+        registered_model = await self.register_object(model)
+        return registered_model
+
+    async def unregister_model(self, model_id: str) -> None:
+        existing_model = await self.get_model(model_id)
+        if existing_model is None:
+            raise ValueError(f"Model {model_id} not found")
+        await self.unregister_object(existing_model)

llama_stack/distribution/routing_tables/scoring_functions.py

@@ -0,0 +1,62 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the terms described in the LICENSE file in
+# the root directory of this source tree.
+
+from llama_stack.apis.common.type_system import ParamType
+from llama_stack.apis.resource import ResourceType
+from llama_stack.apis.scoring_functions import (
+    ListScoringFunctionsResponse,
+    ScoringFn,
+    ScoringFnParams,
+    ScoringFunctions,
+)
+from llama_stack.distribution.datatypes import (
+    ScoringFnWithACL,
+)
+from llama_stack.log import get_logger
+
+from .common import CommonRoutingTableImpl
+
+logger = get_logger(name=__name__, category="core")
+
+
+class ScoringFunctionsRoutingTable(CommonRoutingTableImpl, ScoringFunctions):
+    async def list_scoring_functions(self) -> ListScoringFunctionsResponse:
+        return ListScoringFunctionsResponse(data=await self.get_all_with_type(ResourceType.scoring_function.value))
+
+    async def get_scoring_function(self, scoring_fn_id: str) -> ScoringFn:
+        scoring_fn = await self.get_object_by_identifier("scoring_function", scoring_fn_id)
+        if scoring_fn is None:
+            raise ValueError(f"Scoring function '{scoring_fn_id}' not found")
+        return scoring_fn
+
+    async def register_scoring_function(
+        self,
+        scoring_fn_id: str,
+        description: str,
+        return_type: ParamType,
+        provider_scoring_fn_id: str | None = None,
+        provider_id: str | None = None,
+        params: ScoringFnParams | None = None,
+    ) -> None:
+        if provider_scoring_fn_id is None:
+            provider_scoring_fn_id = scoring_fn_id
+        if provider_id is None:
+            if len(self.impls_by_provider_id) == 1:
+                provider_id = list(self.impls_by_provider_id.keys())[0]
+            else:
+                raise ValueError(
+                    "No provider specified and multiple providers available. Please specify a provider_id."
+                )
+        scoring_fn = ScoringFnWithACL(
+            identifier=scoring_fn_id,
+            description=description,
+            return_type=return_type,
+            provider_resource_id=provider_scoring_fn_id,
+            provider_id=provider_id,
+            params=params,
+        )
+        scoring_fn.provider_id = provider_id
+        await self.register_object(scoring_fn)

llama_stack/distribution/routing_tables/shields.py

@@ -0,0 +1,57 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the terms described in the LICENSE file in
+# the root directory of this source tree.
+
+from typing import Any
+
+from llama_stack.apis.resource import ResourceType
+from llama_stack.apis.shields import ListShieldsResponse, Shield, Shields
+from llama_stack.distribution.datatypes import (
+    ShieldWithACL,
+)
+from llama_stack.log import get_logger
+
+from .common import CommonRoutingTableImpl
+
+logger = get_logger(name=__name__, category="core")
+
+
+class ShieldsRoutingTable(CommonRoutingTableImpl, Shields):
+    async def list_shields(self) -> ListShieldsResponse:
+        return ListShieldsResponse(data=await self.get_all_with_type(ResourceType.shield.value))
+
+    async def get_shield(self, identifier: str) -> Shield:
+        shield = await self.get_object_by_identifier("shield", identifier)
+        if shield is None:
+            raise ValueError(f"Shield '{identifier}' not found")
+        return shield
+
+    async def register_shield(
+        self,
+        shield_id: str,
+        provider_shield_id: str | None = None,
+        provider_id: str | None = None,
+        params: dict[str, Any] | None = None,
+    ) -> Shield:
+        if provider_shield_id is None:
+            provider_shield_id = shield_id
+        if provider_id is None:
+            # If provider_id not specified, use the only provider if it supports this shield type
+            if len(self.impls_by_provider_id) == 1:
+                provider_id = list(self.impls_by_provider_id.keys())[0]
+            else:
+                raise ValueError(
+                    "No provider specified and multiple providers available. Please specify a provider_id."
+                )
+        if params is None:
+            params = {}
+        shield = ShieldWithACL(
+            identifier=shield_id,
+            provider_resource_id=provider_shield_id,
+            provider_id=provider_id,
+            params=params,
+        )
+        await self.register_object(shield)
+        return shield

llama_stack/distribution/routing_tables/toolgroups.py

@@ -0,0 +1,132 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the terms described in the LICENSE file in
+# the root directory of this source tree.
+
+from typing import Any
+
+from llama_stack.apis.common.content_types import URL
+from llama_stack.apis.tools import ListToolGroupsResponse, ListToolsResponse, Tool, ToolGroup, ToolGroups
+from llama_stack.distribution.datatypes import ToolGroupWithACL
+from llama_stack.log import get_logger
+
+from .common import CommonRoutingTableImpl
+
+logger = get_logger(name=__name__, category="core")
+
+
+def parse_toolgroup_from_toolgroup_name_pair(toolgroup_name_with_maybe_tool_name: str) -> str | None:
+    # handle the funny case like "builtin::rag/knowledge_search"
+    parts = toolgroup_name_with_maybe_tool_name.split("/")
+    if len(parts) == 2:
+        return parts[0]
+    else:
+        return None
+
+
+class ToolGroupsRoutingTable(CommonRoutingTableImpl, ToolGroups):
+    toolgroups_to_tools: dict[str, list[Tool]] = {}
+    tool_to_toolgroup: dict[str, str] = {}
+
+    # overridden
+    def get_provider_impl(self, routing_key: str, provider_id: str | None = None) -> Any:
+        # we don't index tools in the registry anymore, but only keep a cache of them by toolgroup_id
+        # TODO: we may want to invalidate the cache (for a given toolgroup_id) every once in a while?
+
+        toolgroup_id = parse_toolgroup_from_toolgroup_name_pair(routing_key)
+        if toolgroup_id:
+            routing_key = toolgroup_id
+
+        if routing_key in self.tool_to_toolgroup:
+            routing_key = self.tool_to_toolgroup[routing_key]
+        return super().get_provider_impl(routing_key, provider_id)
+
+    async def list_tools(self, toolgroup_id: str | None = None) -> ListToolsResponse:
+        if toolgroup_id:
+            if group_id := parse_toolgroup_from_toolgroup_name_pair(toolgroup_id):
+                toolgroup_id = group_id
+            toolgroups = [await self.get_tool_group(toolgroup_id)]
+        else:
+            toolgroups = await self.get_all_with_type("tool_group")
+
+        all_tools = []
+        for toolgroup in toolgroups:
+            if toolgroup.identifier not in self.toolgroups_to_tools:
+                await self._index_tools(toolgroup)
+            all_tools.extend(self.toolgroups_to_tools[toolgroup.identifier])
+
+        return ListToolsResponse(data=all_tools)
+
+    async def _index_tools(self, toolgroup: ToolGroup):
+        provider_impl = super().get_provider_impl(toolgroup.identifier, toolgroup.provider_id)
+        tooldefs_response = await provider_impl.list_runtime_tools(toolgroup.identifier, toolgroup.mcp_endpoint)
+
+        # TODO: kill this Tool vs ToolDef distinction
+        tooldefs = tooldefs_response.data
+        tools = []
+        for t in tooldefs:
+            tools.append(
+                Tool(
+                    identifier=t.name,
+                    toolgroup_id=toolgroup.identifier,
+                    description=t.description or "",
+                    parameters=t.parameters or [],
+                    metadata=t.metadata,
+                    provider_id=toolgroup.provider_id,
+                )
+            )
+
+        self.toolgroups_to_tools[toolgroup.identifier] = tools
+        for tool in tools:
+            self.tool_to_toolgroup[tool.identifier] = toolgroup.identifier
+
+    async def list_tool_groups(self) -> ListToolGroupsResponse:
+        return ListToolGroupsResponse(data=await self.get_all_with_type("tool_group"))
+
+    async def get_tool_group(self, toolgroup_id: str) -> ToolGroup:
+        tool_group = await self.get_object_by_identifier("tool_group", toolgroup_id)
+        if tool_group is None:
+            raise ValueError(f"Tool group '{toolgroup_id}' not found")
+        return tool_group
+
+    async def get_tool(self, tool_name: str) -> Tool:
+        if tool_name in self.tool_to_toolgroup:
+            toolgroup_id = self.tool_to_toolgroup[tool_name]
+            tools = self.toolgroups_to_tools[toolgroup_id]
+            for tool in tools:
+                if tool.identifier == tool_name:
+                    return tool
+        raise ValueError(f"Tool '{tool_name}' not found")
+
+    async def register_tool_group(
+        self,
+        toolgroup_id: str,
+        provider_id: str,
+        mcp_endpoint: URL | None = None,
+        args: dict[str, Any] | None = None,
+    ) -> None:
+        toolgroup = ToolGroupWithACL(
+            identifier=toolgroup_id,
+            provider_id=provider_id,
+            provider_resource_id=toolgroup_id,
+            mcp_endpoint=mcp_endpoint,
+            args=args,
+        )
+        await self.register_object(toolgroup)
+
+        # ideally, indexing of the tools should not be necessary because anyone using
+        # the tools should first list the tools and then use them. but there are assumptions
+        # baked in some of the code and tests right now.
+        if not toolgroup.mcp_endpoint:
+            await self._index_tools(toolgroup)
+        return toolgroup
+
+    async def unregister_toolgroup(self, toolgroup_id: str) -> None:
+        tool_group = await self.get_tool_group(toolgroup_id)
+        if tool_group is None:
+            raise ValueError(f"Tool group {toolgroup_id} not found")
+        await self.unregister_object(tool_group)
+
+    async def shutdown(self) -> None:
+        pass

llama_stack/distribution/routing_tables/vector_dbs.py

@@ -0,0 +1,74 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the terms described in the LICENSE file in
+# the root directory of this source tree.
+
+from pydantic import TypeAdapter
+
+from llama_stack.apis.models import ModelType
+from llama_stack.apis.resource import ResourceType
+from llama_stack.apis.vector_dbs import ListVectorDBsResponse, VectorDB, VectorDBs
+from llama_stack.distribution.datatypes import (
+    VectorDBWithACL,
+)
+from llama_stack.log import get_logger
+
+from .common import CommonRoutingTableImpl
+
+logger = get_logger(name=__name__, category="core")
+
+
+class VectorDBsRoutingTable(CommonRoutingTableImpl, VectorDBs):
+    async def list_vector_dbs(self) -> ListVectorDBsResponse:
+        return ListVectorDBsResponse(data=await self.get_all_with_type("vector_db"))
+
+    async def get_vector_db(self, vector_db_id: str) -> VectorDB:
+        vector_db = await self.get_object_by_identifier("vector_db", vector_db_id)
+        if vector_db is None:
+            raise ValueError(f"Vector DB '{vector_db_id}' not found")
+        return vector_db
+
+    async def register_vector_db(
+        self,
+        vector_db_id: str,
+        embedding_model: str,
+        embedding_dimension: int | None = 384,
+        provider_id: str | None = None,
+        provider_vector_db_id: str | None = None,
+    ) -> VectorDB:
+        if provider_vector_db_id is None:
+            provider_vector_db_id = vector_db_id
+        if provider_id is None:
+            if len(self.impls_by_provider_id) > 0:
+                provider_id = list(self.impls_by_provider_id.keys())[0]
+                if len(self.impls_by_provider_id) > 1:
+                    logger.warning(
+                        f"No provider specified and multiple providers available. Arbitrarily selected the first provider {provider_id}."
+                    )
+            else:
+                raise ValueError("No provider available. Please configure a vector_io provider.")
+        model = await self.get_object_by_identifier("model", embedding_model)
+        if model is None:
+            raise ValueError(f"Model {embedding_model} not found")
+        if model.model_type != ModelType.embedding:
+            raise ValueError(f"Model {embedding_model} is not an embedding model")
+        if "embedding_dimension" not in model.metadata:
+            raise ValueError(f"Model {embedding_model} does not have an embedding dimension")
+        vector_db_data = {
+            "identifier": vector_db_id,
+            "type": ResourceType.vector_db.value,
+            "provider_id": provider_id,
+            "provider_resource_id": provider_vector_db_id,
+            "embedding_model": embedding_model,
+            "embedding_dimension": model.metadata["embedding_dimension"],
+        }
+        vector_db = TypeAdapter(VectorDBWithACL).validate_python(vector_db_data)
+        await self.register_object(vector_db)
+        return vector_db
+
+    async def unregister_vector_db(self, vector_db_id: str) -> None:
+        existing_vector_db = await self.get_vector_db(vector_db_id)
+        if existing_vector_db is None:
+            raise ValueError(f"Vector DB {vector_db_id} not found")
+        await self.unregister_object(existing_vector_db)

llama_stack/distribution/server/auth.py

@@ -8,7 +8,8 @@ import json
 
 import httpx
 
-from llama_stack.distribution.server.auth_providers import AuthProviderConfig, create_auth_provider
+from llama_stack.distribution.datatypes import AuthenticationConfig
+from llama_stack.distribution.server.auth_providers import create_auth_provider
 from llama_stack.log import get_logger
 
 logger = get_logger(name=__name__, category="auth")
@@ -77,7 +78,7 @@ class AuthenticationMiddleware:
     access resources that don't have access_attributes defined.
     """
 
-    def __init__(self, app, auth_config: AuthProviderConfig):
+    def __init__(self, app, auth_config: AuthenticationConfig):
         self.app = app
         self.auth_provider = create_auth_provider(auth_config)
 
@@ -113,6 +114,10 @@ class AuthenticationMiddleware:
                     "roles": [token],
                 }
 
+            # Store the client ID in the request scope so that downstream middleware (like QuotaMiddleware)
+            # can identify the requester and enforce per-client rate limits.
+            scope["authenticated_client_id"] = token
+
             # Store attributes in request scope
             scope["user_attributes"] = user_attributes
             scope["principal"] = validation_result.principal

llama_stack/distribution/server/auth_providers.py

@@ -4,17 +4,19 @@
 # This source code is licensed under the terms described in the LICENSE file in
 # the root directory of this source tree.
 
-import json
+import ssl
 import time
 from abc import ABC, abstractmethod
-from enum import Enum
+from asyncio import Lock
+from pathlib import Path
 from urllib.parse import parse_qs
 
 import httpx
 from jose import jwt
-from pydantic import BaseModel, Field, field_validator
+from pydantic import BaseModel, Field, field_validator, model_validator
+from typing_extensions import Self
 
-from llama_stack.distribution.datatypes import AccessAttributes
+from llama_stack.distribution.datatypes import AccessAttributes, AuthenticationConfig, AuthProviderType
 from llama_stack.log import get_logger
 
 logger = get_logger(name=__name__, category="auth")
@@ -72,21 +74,6 @@ class AuthRequest(BaseModel):
     request: AuthRequestContext = Field(description="Context information about the request being authenticated")
 
 
-class AuthProviderType(str, Enum):
-    """Supported authentication provider types."""
-
-    KUBERNETES = "kubernetes"
-    CUSTOM = "custom"
-    OAUTH2_TOKEN = "oauth2_token"
-
-
-class AuthProviderConfig(BaseModel):
-    """Base configuration for authentication providers."""
-
-    provider_type: AuthProviderType = Field(..., description="Type of authentication provider")
-    config: dict[str, str] = Field(..., description="Provider-specific configuration")
-
-
 class AuthProvider(ABC):
     """Abstract base class for authentication providers."""
 
@@ -101,83 +88,6 @@ class AuthProvider(ABC):
         pass
 
 
-class KubernetesAuthProviderConfig(BaseModel):
-    api_server_url: str
-    ca_cert_path: str | None = None
-
-
-class KubernetesAuthProvider(AuthProvider):
-    """Kubernetes authentication provider that validates tokens against the Kubernetes API server."""
-
-    def __init__(self, config: KubernetesAuthProviderConfig):
-        self.config = config
-        self._client = None
-
-    async def _get_client(self):
-        """Get or create a Kubernetes client."""
-        if self._client is None:
-            # kubernetes-client has not async support, see:
-            # https://github.com/kubernetes-client/python/issues/323
-            from kubernetes import client
-            from kubernetes.client import ApiClient
-
-            # Configure the client
-            configuration = client.Configuration()
-            configuration.host = self.config.api_server_url
-            if self.config.ca_cert_path:
-                configuration.ssl_ca_cert = self.config.ca_cert_path
-            configuration.verify_ssl = bool(self.config.ca_cert_path)
-
-            # Create API client
-            self._client = ApiClient(configuration)
-        return self._client
-
-    async def validate_token(self, token: str, scope: dict | None = None) -> TokenValidationResult:
-        """Validate a Kubernetes token and return access attributes."""
-        try:
-            client = await self._get_client()
-
-            # Set the token in the client
-            client.set_default_header("Authorization", f"Bearer {token}")
-
-            # Make a request to validate the token
-            # We use the /api endpoint which requires authentication
-            from kubernetes.client import CoreV1Api
-
-            api = CoreV1Api(client)
-            api.get_api_resources(_request_timeout=3.0)  # Set timeout for this specific request
-
-            # If we get here, the token is valid
-            # Extract user info from the token claims
-            import base64
-
-            # Decode the token (without verification since we've already validated it)
-            token_parts = token.split(".")
-            payload = json.loads(base64.b64decode(token_parts[1] + "=" * (-len(token_parts[1]) % 4)))
-
-            # Extract user information from the token
-            username = payload.get("sub", "")
-            groups = payload.get("groups", [])
-
-            return TokenValidationResult(
-                principal=username,
-                access_attributes=AccessAttributes(
-                    roles=[username],  # Use username as a role
-                    teams=groups,  # Use Kubernetes groups as teams
-                ),
-            )
-
-        except Exception as e:
-            logger.exception("Failed to validate Kubernetes token")
-            raise ValueError("Invalid or expired token") from e
-
-    async def close(self):
-        """Close the HTTP client."""
-        if self._client:
-            self._client.close()
-            self._client = None
-
-
 def get_attributes_from_claims(claims: dict[str, str], mapping: dict[str, str]) -> AccessAttributes:
     attributes = AccessAttributes()
     for claim_key, attribute_key in mapping.items():
@@ -197,11 +107,24 @@ def get_attributes_from_claims(claims: dict[str, str], mapping: dict[str, str])
     return attributes
 
 
-class OAuth2TokenAuthProviderConfig(BaseModel):
+class OAuth2JWKSConfig(BaseModel):
     # The JWKS URI for collecting public keys
-    jwks_uri: str
-    cache_ttl: int = 3600
+    uri: str
+    key_recheck_period: int = Field(default=3600, description="The period to recheck the JWKS URI for key updates")
+
+
+class OAuth2IntrospectionConfig(BaseModel):
+    url: str
+    client_id: str
+    client_secret: str
+    send_secret_in_body: bool = False
+
+
+class OAuth2TokenAuthProviderConfig(BaseModel):
     audience: str = "llama-stack"
+    verify_tls: bool = True
+    tls_cafile: Path | None = None
+    issuer: str | None = Field(default=None, description="The OIDC issuer URL.")
     claims_mapping: dict[str, str] = Field(
         default_factory=lambda: {
             "sub": "roles",
@@ -213,6 +136,8 @@ class OAuth2TokenAuthProviderConfig(BaseModel):
             "namespace": "namespaces",
         },
     )
+    jwks: OAuth2JWKSConfig | None
+    introspection: OAuth2IntrospectionConfig | None = None
 
     @classmethod
     @field_validator("claims_mapping")
@@ -224,6 +149,14 @@ class OAuth2TokenAuthProviderConfig(BaseModel):
                 raise ValueError(f"claims_mapping value is not a valid attribute: {value}")
         return v
 
+    @model_validator(mode="after")
+    def validate_mode(self) -> Self:
+        if not self.jwks and not self.introspection:
+            raise ValueError("One of jwks or introspection must be configured")
+        if self.jwks and self.introspection:
+            raise ValueError("At present only one of jwks or introspection should be configured")
+        return self
+
 
 class OAuth2TokenAuthProvider(AuthProvider):
     """
@@ -236,8 +169,16 @@ class OAuth2TokenAuthProvider(AuthProvider):
         self.config = config
         self._jwks_at: float = 0.0
         self._jwks: dict[str, str] = {}
+        self._jwks_lock = Lock()
 
     async def validate_token(self, token: str, scope: dict | None = None) -> TokenValidationResult:
+        if self.config.jwks:
+            return await self.validate_jwt_token(token, scope)
+        if self.config.introspection:
+            return await self.introspect_token(token, scope)
+        raise ValueError("One of jwks or introspection must be configured")
+
+    async def validate_jwt_token(self, token: str, scope: dict | None = None) -> TokenValidationResult:
         """Validate a token using the JWT token."""
         await self._refresh_jwks()
 
@@ -253,7 +194,7 @@ class OAuth2TokenAuthProvider(AuthProvider):
                 key_data,
                 algorithms=[algorithm],
                 audience=self.config.audience,
-                options={"verify_exp": True},
+                issuer=self.config.issuer,
             )
         except Exception as exc:
             raise ValueError(f"Invalid JWT token: {token}") from exc
@@ -267,21 +208,84 @@ class OAuth2TokenAuthProvider(AuthProvider):
             access_attributes=access_attributes,
         )
 
+    async def introspect_token(self, token: str, scope: dict | None = None) -> TokenValidationResult:
+        """Validate a token using token introspection as defined by RFC 7662."""
+        form = {
+            "token": token,
+        }
+        if self.config.introspection is None:
+            raise ValueError("Introspection is not configured")
+
+        if self.config.introspection.send_secret_in_body:
+            form["client_id"] = self.config.introspection.client_id
+            form["client_secret"] = self.config.introspection.client_secret
+            auth = None
+        else:
+            auth = (self.config.introspection.client_id, self.config.introspection.client_secret)
+        ssl_ctxt = None
+        if self.config.tls_cafile:
+            ssl_ctxt = ssl.create_default_context(cafile=self.config.tls_cafile.as_posix())
+        try:
+            async with httpx.AsyncClient(verify=ssl_ctxt) as client:
+                response = await client.post(
+                    self.config.introspection.url,
+                    data=form,
+                    auth=auth,
+                    timeout=10.0,  # Add a reasonable timeout
+                )
+                if response.status_code != 200:
+                    logger.warning(f"Token introspection failed with status code: {response.status_code}")
+                    raise ValueError(f"Token introspection failed: {response.status_code}")
+
+                fields = response.json()
+                if not fields["active"]:
+                    raise ValueError("Token not active")
+                principal = fields["sub"] or fields["username"]
+                access_attributes = get_attributes_from_claims(fields, self.config.claims_mapping)
+                return TokenValidationResult(
+                    principal=principal,
+                    access_attributes=access_attributes,
+                )
+        except httpx.TimeoutException:
+            logger.exception("Token introspection request timed out")
+            raise
+        except ValueError:
+            # Re-raise ValueError exceptions to preserve their message
+            raise
+        except Exception as e:
+            logger.exception("Error during token introspection")
+            raise ValueError("Token introspection error") from e
+
     async def close(self):
-        """Close the HTTP client."""
+        pass
 
     async def _refresh_jwks(self) -> None:
-        if time.time() - self._jwks_at > self.config.cache_ttl:
-            async with httpx.AsyncClient() as client:
-                res = await client.get(self.config.jwks_uri, timeout=5)
-                res.raise_for_status()
-                jwks_data = res.json()["keys"]
-                self._jwks = {}
-                for k in jwks_data:
-                    kid = k["kid"]
-                    # Store the entire key object as it may be needed for different algorithms
-                    self._jwks[kid] = k
-                self._jwks_at = time.time()
+        """
+        Refresh the JWKS cache.
+
+        This is a simple cache that expires after a certain amount of time (defined by `key_recheck_period`).
+        If the cache is expired, we refresh the JWKS from the JWKS URI.
+
+        Notes: for Kubernetes which doesn't fully implement the OIDC protocol:
+            * It doesn't have user authentication flows
+            * It doesn't have refresh tokens
+        """
+        async with self._jwks_lock:
+            if self.config.jwks is None:
+                raise ValueError("JWKS is not configured")
+            if time.time() - self._jwks_at > self.config.jwks.key_recheck_period:
+                verify = self.config.tls_cafile.as_posix() if self.config.tls_cafile else self.config.verify_tls
+                async with httpx.AsyncClient(verify=verify) as client:
+                    res = await client.get(self.config.jwks.uri, timeout=5)
+                    res.raise_for_status()
+                    jwks_data = res.json()["keys"]
+                    updated = {}
+                    for k in jwks_data:
+                        kid = k["kid"]
+                        # Store the entire key object as it may be needed for different algorithms
+                        updated[kid] = k
+                    self._jwks = updated
+                    self._jwks_at = time.time()
 
 
 class CustomAuthProviderConfig(BaseModel):
@@ -359,13 +363,11 @@ class CustomAuthProvider(AuthProvider):
             self._client = None
 
 
-def create_auth_provider(config: AuthProviderConfig) -> AuthProvider:
+def create_auth_provider(config: AuthenticationConfig) -> AuthProvider:
     """Factory function to create the appropriate auth provider."""
     provider_type = config.provider_type.lower()
 
-    if provider_type == "kubernetes":
-        return KubernetesAuthProvider(KubernetesAuthProviderConfig.model_validate(config.config))
-    elif provider_type == "custom":
+    if provider_type == "custom":
         return CustomAuthProvider(CustomAuthProviderConfig.model_validate(config.config))
     elif provider_type == "oauth2_token":
         return OAuth2TokenAuthProvider(OAuth2TokenAuthProviderConfig.model_validate(config.config))

llama_stack/distribution/server/quota.py

@@ -0,0 +1,110 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the terms described in the LICENSE file in
+# the root directory of this source tree.
+
+import json
+import time
+from datetime import datetime, timedelta, timezone
+
+from starlette.types import ASGIApp, Receive, Scope, Send
+
+from llama_stack.log import get_logger
+from llama_stack.providers.utils.kvstore.api import KVStore
+from llama_stack.providers.utils.kvstore.config import KVStoreConfig, SqliteKVStoreConfig
+from llama_stack.providers.utils.kvstore.kvstore import kvstore_impl
+
+logger = get_logger(name=__name__, category="quota")
+
+
+class QuotaMiddleware:
+    """
+    ASGI middleware that enforces separate quotas for authenticated and anonymous clients
+    within a configurable time window.
+
+    - For authenticated requests, it reads the client ID from the
+      `Authorization: Bearer <client_id>` header.
+    - For anonymous requests, it falls back to the IP address of the client.
+    Requests are counted in a KV store (e.g., SQLite), and HTTP 429 is returned
+    once a client exceeds its quota.
+    """
+
+    def __init__(
+        self,
+        app: ASGIApp,
+        kv_config: KVStoreConfig,
+        anonymous_max_requests: int,
+        authenticated_max_requests: int,
+        window_seconds: int = 86400,
+    ):
+        self.app = app
+        self.kv_config = kv_config
+        self.kv: KVStore | None = None
+        self.anonymous_max_requests = anonymous_max_requests
+        self.authenticated_max_requests = authenticated_max_requests
+        self.window_seconds = window_seconds
+
+        if isinstance(self.kv_config, SqliteKVStoreConfig):
+            logger.warning(
+                "QuotaMiddleware: Using SQLite backend. Expiry/TTL is not enforced; cleanup is manual. "
+                f"window_seconds={self.window_seconds}"
+            )
+
+    async def _get_kv(self) -> KVStore:
+        if self.kv is None:
+            self.kv = await kvstore_impl(self.kv_config)
+        return self.kv
+
+    async def __call__(self, scope: Scope, receive: Receive, send: Send):
+        if scope["type"] == "http":
+            # pick key & limit based on auth
+            auth_id = scope.get("authenticated_client_id")
+            if auth_id:
+                key_id = auth_id
+                limit = self.authenticated_max_requests
+            else:
+                # fallback to IP
+                client = scope.get("client")
+                key_id = client[0] if client else "anonymous"
+                limit = self.anonymous_max_requests
+
+            current_window = int(time.time() // self.window_seconds)
+            key = f"quota:{key_id}:{current_window}"
+
+            try:
+                kv = await self._get_kv()
+                prev = await kv.get(key) or "0"
+                count = int(prev) + 1
+
+                if int(prev) == 0:
+                    # Set with expiration datetime when it is the first request in the window.
+                    expiration = datetime.now(timezone.utc) + timedelta(seconds=self.window_seconds)
+                    await kv.set(key, str(count), expiration=expiration)
+                else:
+                    await kv.set(key, str(count))
+            except Exception:
+                logger.exception("Failed to access KV store for quota")
+                return await self._send_error(send, 500, "Quota service error")
+
+            if count > limit:
+                logger.warning(
+                    "Quota exceeded for client %s: %d/%d",
+                    key_id,
+                    count,
+                    limit,
+                )
+                return await self._send_error(send, 429, "Quota exceeded")
+
+        return await self.app(scope, receive, send)
+
+    async def _send_error(self, send: Send, status: int, message: str):
+        await send(
+            {
+                "type": "http.response.start",
+                "status": status,
+                "headers": [[b"content-type", b"application/json"]],
+            }
+        )
+        body = json.dumps({"error": {"message": message}}).encode()
+        await send({"type": "http.response.body", "body": body})

llama_stack/distribution/server/server.py

@@ -23,11 +23,12 @@ import yaml
 from fastapi import Body, FastAPI, HTTPException, Request
 from fastapi import Path as FastapiPath
 from fastapi.exceptions import RequestValidationError
+from fastapi.middleware.cors import CORSMiddleware
 from fastapi.responses import JSONResponse, StreamingResponse
 from openai import BadRequestError
 from pydantic import BaseModel, ValidationError
 
-from llama_stack.distribution.datatypes import LoggingConfig, StackRunConfig
+from llama_stack.distribution.datatypes import AuthenticationRequiredError, LoggingConfig, StackRunConfig
 from llama_stack.distribution.distribution import builtin_automatically_routed_apis
 from llama_stack.distribution.request_headers import (
     PROVIDER_DATA_VAR,
@@ -60,6 +61,7 @@ from llama_stack.providers.utils.telemetry.tracing import (
 
 from .auth import AuthenticationMiddleware
 from .endpoints import get_all_api_endpoints
+from .quota import QuotaMiddleware
 
 REPO_ROOT = Path(__file__).parent.parent.parent.parent
 
@@ -120,6 +122,8 @@ def translate_exception(exc: Exception) -> HTTPException | RequestValidationErro
         return HTTPException(status_code=504, detail=f"Operation timed out: {str(exc)}")
     elif isinstance(exc, NotImplementedError):
         return HTTPException(status_code=501, detail=f"Not implemented: {str(exc)}")
+    elif isinstance(exc, AuthenticationRequiredError):
+        return HTTPException(status_code=401, detail=f"Authentication required: {str(exc)}")
     else:
         return HTTPException(
             status_code=500,
@@ -280,7 +284,18 @@ class TracingMiddleware:
             logger.debug(f"No matching endpoint found for path: {path}, falling back to FastAPI")
             return await self.app(scope, receive, send)
 
-        trace_context = await start_trace(trace_path, {"__location__": "server", "raw_path": path})
+        trace_attributes = {"__location__": "server", "raw_path": path}
+
+        # Extract W3C trace context headers and store as trace attributes
+        headers = dict(scope.get("headers", []))
+        traceparent = headers.get(b"traceparent", b"").decode()
+        if traceparent:
+            trace_attributes["traceparent"] = traceparent
+        tracestate = headers.get(b"tracestate", b"").decode()
+        if tracestate:
+            trace_attributes["tracestate"] = tracestate
+
+        trace_context = await start_trace(trace_path, trace_attributes)
 
         async def send_with_trace_id(message):
             if message["type"] == "http.response.start":
@@ -370,14 +385,6 @@ def main(args: argparse.Namespace | None = None):
     if args is None:
         args = parser.parse_args()
 
-    # Check for deprecated argument usage
-    if "--config" in sys.argv:
-        warnings.warn(
-            "The '--config' argument is deprecated and will be removed in a future version. Use '--config' instead.",
-            DeprecationWarning,
-            stacklevel=2,
-        )
-
     log_line = ""
     if args.config:
         # if the user provided a config file, use it, even if template was specified
@@ -431,6 +438,46 @@ def main(args: argparse.Namespace | None = None):
     if config.server.auth:
         logger.info(f"Enabling authentication with provider: {config.server.auth.provider_type.value}")
         app.add_middleware(AuthenticationMiddleware, auth_config=config.server.auth)
+    else:
+        if config.server.quota:
+            quota = config.server.quota
+            logger.warning(
+                "Configured authenticated_max_requests (%d) but no auth is enabled; "
+                "falling back to anonymous_max_requests (%d) for all the requests",
+                quota.authenticated_max_requests,
+                quota.anonymous_max_requests,
+            )
+
+    if config.server.quota:
+        logger.info("Enabling quota middleware for authenticated and anonymous clients")
+
+        quota = config.server.quota
+        anonymous_max_requests = quota.anonymous_max_requests
+        # if auth is disabled, use the anonymous max requests
+        authenticated_max_requests = quota.authenticated_max_requests if config.server.auth else anonymous_max_requests
+
+        kv_config = quota.kvstore
+        window_map = {"day": 86400}
+        window_seconds = window_map[quota.period.value]
+
+        app.add_middleware(
+            QuotaMiddleware,
+            kv_config=kv_config,
+            anonymous_max_requests=anonymous_max_requests,
+            authenticated_max_requests=authenticated_max_requests,
+            window_seconds=window_seconds,
+        )
+
+    # --- CORS middleware for local development ---
+    # TODO: move to reverse proxy
+    ui_port = os.environ.get("LLAMA_STACK_UI_PORT", 8322)
+    app.add_middleware(
+        CORSMiddleware,
+        allow_origins=[f"http://localhost:{ui_port}"],
+        allow_credentials=True,
+        allow_methods=["*"],
+        allow_headers=["*"],
+    )
 
     try:
         impls = asyncio.run(construct_stack(config))

llama_stack/distribution/store/registry.py

@@ -36,7 +36,7 @@ class DistributionRegistry(Protocol):
 
 
 REGISTER_PREFIX = "distributions:registry"
-KEY_VERSION = "v8"
+KEY_VERSION = "v9"
 KEY_FORMAT = f"{REGISTER_PREFIX}:{KEY_VERSION}::" + "{type}:{identifier}"
 
 

llama_stack/distribution/utils/exec.py

@@ -8,6 +8,7 @@ import logging
 import os
 import signal
 import subprocess
+import sys
 
 from termcolor import cprint
 
@@ -33,6 +34,7 @@ def formulate_run_args(image_type, image_name, config, template_name) -> list:
             cprint(
                 "No current conda environment detected, please specify a conda environment name with --image-name",
                 color="red",
+                file=sys.stderr,
             )
             return
 
@@ -49,12 +51,13 @@ def formulate_run_args(image_type, image_name, config, template_name) -> list:
                     return envpath
             return None
 
-        print(f"Using conda environment: {env_name}")
+        cprint(f"Using conda environment: {env_name}", color="green", file=sys.stderr)
         conda_prefix = get_conda_prefix(env_name)
         if not conda_prefix:
             cprint(
                 f"Conda environment {env_name} does not exist.",
                 color="red",
+                file=sys.stderr,
             )
             return
 
@@ -63,6 +66,7 @@ def formulate_run_args(image_type, image_name, config, template_name) -> list:
             cprint(
                 f"Build file {build_file} does not exist.\n\nPlease run `llama stack build` or specify the correct conda environment name with --image-name",
                 color="red",
+                file=sys.stderr,
             )
             return
     else:
@@ -73,9 +77,10 @@ def formulate_run_args(image_type, image_name, config, template_name) -> list:
             cprint(
                 "No current virtual environment detected, please specify a virtual environment name with --image-name",
                 color="red",
+                file=sys.stderr,
             )
             return
-        print(f"Using virtual environment: {env_name}")
+        cprint(f"Using virtual environment: {env_name}", file=sys.stderr)
 
     script = importlib.resources.files("llama_stack") / "distribution/start_stack.sh"
     run_args = [

llama_stack/log.py

@@ -6,6 +6,7 @@
 
 import logging
 import os
+import sys
 from logging.config import dictConfig
 
 from rich.console import Console
@@ -234,7 +235,7 @@ def get_logger(
 
 env_config = os.environ.get("LLAMA_STACK_LOGGING", "")
 if env_config:
-    cprint(f"Environment variable LLAMA_STACK_LOGGING found: {env_config}", "yellow")
+    cprint(f"Environment variable LLAMA_STACK_LOGGING found: {env_config}", color="yellow", file=sys.stderr)
     _category_levels.update(parse_environment_config(env_config))
 
 log_file = os.environ.get("LLAMA_STACK_LOG_FILE")

llama_stack/models/llama/llama3/generation.py

@@ -174,6 +174,7 @@ class Llama3:
                 cprint(
                     "Input to model:\n" + self.tokenizer.decode(tokens_to_print) + "\n",
                     "red",
+                    file=sys.stderr,
                 )
         prompt_tokens = [inp.tokens for inp in llm_inputs]
 
@@ -184,7 +185,11 @@ class Llama3:
         max_prompt_len = max(len(t) for t in prompt_tokens)
 
         if max_prompt_len >= params.max_seq_len:
-            cprint(f"Out of token budget {max_prompt_len} vs {params.max_seq_len}", "red")
+            cprint(
+                f"Out of token budget {max_prompt_len} vs {params.max_seq_len}",
+                color="red",
+                file=sys.stderr,
+            )
             return
 
         total_len = min(max_gen_len + max_prompt_len, params.max_seq_len)

llama_stack/models/llama/llama4/generation.py

@@ -133,9 +133,9 @@ class Llama4:
 
         print_model_input = print_model_input or os.environ.get("LLAMA_MODELS_DEBUG", "0") == "1"
         if print_model_input:
-            cprint("Input to model:\n", "yellow")
+            cprint("Input to model:\n", color="yellow", file=sys.stderr)
             for inp in llm_inputs:
-                cprint(self.tokenizer.decode(inp.tokens), "grey")
+                cprint(self.tokenizer.decode(inp.tokens), color="grey", file=sys.stderr)
         prompt_tokens = [inp.tokens for inp in llm_inputs]
 
         bsz = len(llm_inputs)
@@ -145,7 +145,7 @@ class Llama4:
         max_prompt_len = max(len(t) for t in prompt_tokens)
 
         if max_prompt_len >= params.max_seq_len:
-            cprint(f"Out of token budget {max_prompt_len} vs {params.max_seq_len}", "red")
+            cprint(f"Out of token budget {max_prompt_len} vs {params.max_seq_len}", color="red", file=sys.stderr)
             return
 
         total_len = min(max_gen_len + max_prompt_len, params.max_seq_len)

llama_stack/providers/datatypes.py

@@ -16,7 +16,7 @@ from llama_stack.apis.datatypes import Api
 from llama_stack.apis.models import Model
 from llama_stack.apis.scoring_functions import ScoringFn
 from llama_stack.apis.shields import Shield
-from llama_stack.apis.tools import Tool
+from llama_stack.apis.tools import ToolGroup
 from llama_stack.apis.vector_dbs import VectorDB
 from llama_stack.schema_utils import json_schema_type
 
@@ -74,10 +74,10 @@ class BenchmarksProtocolPrivate(Protocol):
     async def register_benchmark(self, benchmark: Benchmark) -> None: ...
 
 
-class ToolsProtocolPrivate(Protocol):
-    async def register_tool(self, tool: Tool) -> None: ...
+class ToolGroupsProtocolPrivate(Protocol):
+    async def register_toolgroup(self, toolgroup: ToolGroup) -> None: ...
 
-    async def unregister_tool(self, tool_id: str) -> None: ...
+    async def unregister_toolgroup(self, toolgroup_id: str) -> None: ...
 
 
 @json_schema_type

llama_stack/providers/inline/agents/meta_reference/agents.py

@@ -20,9 +20,12 @@ from llama_stack.apis.agents import (
     AgentTurnCreateRequest,
     AgentTurnResumeRequest,
     Document,
+    ListOpenAIResponseInputItem,
+    ListOpenAIResponseObject,
     OpenAIResponseInput,
     OpenAIResponseInputTool,
     OpenAIResponseObject,
+    Order,
     Session,
     Turn,
 )
@@ -39,6 +42,7 @@ from llama_stack.apis.tools import ToolGroups, ToolRuntime
 from llama_stack.apis.vector_io import VectorIO
 from llama_stack.providers.utils.kvstore import InmemoryKVStoreImpl, kvstore_impl
 from llama_stack.providers.utils.pagination import paginate_records
+from llama_stack.providers.utils.responses.responses_store import ResponsesStore
 
 from .agent_instance import ChatAgent
 from .config import MetaReferenceAgentsImplConfig
@@ -66,15 +70,17 @@ class MetaReferenceAgentsImpl(Agents):
         self.tool_groups_api = tool_groups_api
 
         self.in_memory_store = InmemoryKVStoreImpl()
-        self.openai_responses_impl = None
+        self.openai_responses_impl: OpenAIResponsesImpl | None = None
 
     async def initialize(self) -> None:
         self.persistence_store = await kvstore_impl(self.config.persistence_store)
+        self.responses_store = ResponsesStore(self.config.responses_store)
+        await self.responses_store.initialize()
         self.openai_responses_impl = OpenAIResponsesImpl(
-            self.persistence_store,
             inference_api=self.inference_api,
             tool_groups_api=self.tool_groups_api,
             tool_runtime_api=self.tool_runtime_api,
+            responses_store=self.responses_store,
         )
 
     async def create_agent(
@@ -305,14 +311,15 @@ class MetaReferenceAgentsImpl(Agents):
     # OpenAI responses
     async def get_openai_response(
         self,
-        id: str,
+        response_id: str,
     ) -> OpenAIResponseObject:
-        return await self.openai_responses_impl.get_openai_response(id)
+        return await self.openai_responses_impl.get_openai_response(response_id)
 
     async def create_openai_response(
         self,
         input: str | list[OpenAIResponseInput],
         model: str,
+        instructions: str | None = None,
         previous_response_id: str | None = None,
         store: bool | None = True,
         stream: bool | None = False,
@@ -320,5 +327,27 @@ class MetaReferenceAgentsImpl(Agents):
         tools: list[OpenAIResponseInputTool] | None = None,
     ) -> OpenAIResponseObject:
         return await self.openai_responses_impl.create_openai_response(
-            input, model, previous_response_id, store, stream, temperature, tools
+            input, model, instructions, previous_response_id, store, stream, temperature, tools
+        )
+
+    async def list_openai_responses(
+        self,
+        after: str | None = None,
+        limit: int | None = 50,
+        model: str | None = None,
+        order: Order | None = Order.desc,
+    ) -> ListOpenAIResponseObject:
+        return await self.openai_responses_impl.list_openai_responses(after, limit, model, order)
+
+    async def list_openai_response_input_items(
+        self,
+        response_id: str,
+        after: str | None = None,
+        before: str | None = None,
+        include: list[str] | None = None,
+        limit: int | None = 20,
+        order: Order | None = Order.desc,
+    ) -> ListOpenAIResponseInputItem:
+        return await self.openai_responses_impl.list_openai_response_input_items(
+            response_id, after, before, include, limit, order
         )

llama_stack/providers/inline/agents/meta_reference/config.py

@@ -10,10 +10,12 @@ from pydantic import BaseModel
 
 from llama_stack.providers.utils.kvstore import KVStoreConfig
 from llama_stack.providers.utils.kvstore.config import SqliteKVStoreConfig
+from llama_stack.providers.utils.sqlstore.sqlstore import SqliteSqlStoreConfig, SqlStoreConfig
 
 
 class MetaReferenceAgentsImplConfig(BaseModel):
     persistence_store: KVStoreConfig
+    responses_store: SqlStoreConfig
 
     @classmethod
     def sample_run_config(cls, __distro_dir__: str) -> dict[str, Any]:
@@ -21,5 +23,9 @@ class MetaReferenceAgentsImplConfig(BaseModel):
             "persistence_store": SqliteKVStoreConfig.sample_run_config(
                 __distro_dir__=__distro_dir__,
                 db_name="agents_store.db",
-            )
+            ),
+            "responses_store": SqliteSqlStoreConfig.sample_run_config(
+                __distro_dir__=__distro_dir__,
+                db_name="responses_store.db",
+            ),
         }

llama_stack/providers/inline/agents/meta_reference/openai_responses.py

@@ -5,6 +5,7 @@
 # the root directory of this source tree.
 
 import json
+import time
 import uuid
 from collections.abc import AsyncIterator
 from typing import Any, cast
@@ -12,24 +13,29 @@ from typing import Any, cast
 from openai.types.chat import ChatCompletionToolParam
 from pydantic import BaseModel
 
+from llama_stack.apis.agents import Order
 from llama_stack.apis.agents.openai_responses import (
+    AllowedToolsFilter,
+    ListOpenAIResponseInputItem,
+    ListOpenAIResponseObject,
     OpenAIResponseInput,
     OpenAIResponseInputFunctionToolCallOutput,
-    OpenAIResponseInputItemList,
     OpenAIResponseInputMessageContent,
     OpenAIResponseInputMessageContentImage,
     OpenAIResponseInputMessageContentText,
     OpenAIResponseInputTool,
-    OpenAIResponseInputToolFunction,
+    OpenAIResponseInputToolMCP,
     OpenAIResponseMessage,
     OpenAIResponseObject,
     OpenAIResponseObjectStream,
     OpenAIResponseObjectStreamResponseCompleted,
     OpenAIResponseObjectStreamResponseCreated,
+    OpenAIResponseObjectStreamResponseOutputTextDelta,
     OpenAIResponseOutput,
     OpenAIResponseOutputMessageContent,
     OpenAIResponseOutputMessageContentOutputText,
     OpenAIResponseOutputMessageFunctionToolCall,
+    OpenAIResponseOutputMessageMCPListTools,
     OpenAIResponseOutputMessageWebSearchToolCall,
 )
 from llama_stack.apis.inference.inference import (
@@ -49,11 +55,12 @@ from llama_stack.apis.inference.inference import (
     OpenAIToolMessageParam,
     OpenAIUserMessageParam,
 )
-from llama_stack.apis.tools.tools import ToolGroups, ToolInvocationResult, ToolRuntime
+from llama_stack.apis.tools.tools import ToolGroups, ToolRuntime
 from llama_stack.log import get_logger
 from llama_stack.models.llama.datatypes import ToolDefinition, ToolParamDefinition
 from llama_stack.providers.utils.inference.openai_compat import convert_tooldef_to_openai_tool
-from llama_stack.providers.utils.kvstore import KVStore
+from llama_stack.providers.utils.responses.responses_store import ResponsesStore
+from llama_stack.providers.utils.tools.mcp import invoke_mcp_tool, list_mcp_tools
 
 logger = get_logger(name=__name__, category="openai_responses")
 
@@ -162,41 +169,43 @@ async def _get_message_type_by_role(role: str):
 
 
 class OpenAIResponsePreviousResponseWithInputItems(BaseModel):
-    input_items: OpenAIResponseInputItemList
+    input_items: ListOpenAIResponseInputItem
     response: OpenAIResponseObject
 
 
+class ChatCompletionContext(BaseModel):
+    model: str
+    messages: list[OpenAIMessageParam]
+    tools: list[ChatCompletionToolParam] | None = None
+    mcp_tool_to_server: dict[str, OpenAIResponseInputToolMCP]
+    stream: bool
+    temperature: float | None
+
+
 class OpenAIResponsesImpl:
     def __init__(
         self,
-        persistence_store: KVStore,
         inference_api: Inference,
         tool_groups_api: ToolGroups,
         tool_runtime_api: ToolRuntime,
+        responses_store: ResponsesStore,
     ):
-        self.persistence_store = persistence_store
         self.inference_api = inference_api
         self.tool_groups_api = tool_groups_api
         self.tool_runtime_api = tool_runtime_api
-
-    async def _get_previous_response_with_input(self, id: str) -> OpenAIResponsePreviousResponseWithInputItems:
-        key = f"{OPENAI_RESPONSES_PREFIX}{id}"
-        response_json = await self.persistence_store.get(key=key)
-        if response_json is None:
-            raise ValueError(f"OpenAI response with id '{id}' not found")
-        return OpenAIResponsePreviousResponseWithInputItems.model_validate_json(response_json)
+        self.responses_store = responses_store
 
     async def _prepend_previous_response(
         self, input: str | list[OpenAIResponseInput], previous_response_id: str | None = None
     ):
         if previous_response_id:
-            previous_response_with_input = await self._get_previous_response_with_input(previous_response_id)
+            previous_response_with_input = await self.responses_store.get_response_object(previous_response_id)
 
             # previous response input items
-            new_input_items = previous_response_with_input.input_items.data
+            new_input_items = previous_response_with_input.input
 
             # previous response output items
-            new_input_items.extend(previous_response_with_input.response.output)
+            new_input_items.extend(previous_response_with_input.output)
 
             # new input items from the current request
             if isinstance(input, str):
@@ -208,99 +217,60 @@ class OpenAIResponsesImpl:
 
         return input
 
+    async def _prepend_instructions(self, messages, instructions):
+        if instructions:
+            messages.insert(0, OpenAISystemMessageParam(content=instructions))
+
     async def get_openai_response(
         self,
-        id: str,
+        response_id: str,
     ) -> OpenAIResponseObject:
-        response_with_input = await self._get_previous_response_with_input(id)
-        return response_with_input.response
+        response_with_input = await self.responses_store.get_response_object(response_id)
+        return OpenAIResponseObject(**{k: v for k, v in response_with_input.model_dump().items() if k != "input"})
 
-    async def create_openai_response(
+    async def list_openai_responses(
         self,
-        input: str | list[OpenAIResponseInput],
-        model: str,
-        previous_response_id: str | None = None,
-        store: bool | None = True,
-        stream: bool | None = False,
-        temperature: float | None = None,
-        tools: list[OpenAIResponseInputTool] | None = None,
-    ):
-        stream = False if stream is None else stream
+        after: str | None = None,
+        limit: int | None = 50,
+        model: str | None = None,
+        order: Order | None = Order.desc,
+    ) -> ListOpenAIResponseObject:
+        return await self.responses_store.list_responses(after, limit, model, order)
 
-        input = await self._prepend_previous_response(input, previous_response_id)
-        messages = await _convert_response_input_to_chat_messages(input)
-        chat_tools = await self._convert_response_tools_to_chat_tools(tools) if tools else None
-        chat_response = await self.inference_api.openai_chat_completion(
-            model=model,
-            messages=messages,
-            tools=chat_tools,
-            stream=stream,
-            temperature=temperature,
-        )
+    async def list_openai_response_input_items(
+        self,
+        response_id: str,
+        after: str | None = None,
+        before: str | None = None,
+        include: list[str] | None = None,
+        limit: int | None = 20,
+        order: Order | None = Order.desc,
+    ) -> ListOpenAIResponseInputItem:
+        """List input items for a given OpenAI response.
 
-        if stream:
-            # TODO: refactor this into a separate method that handles streaming
-            chat_response_id = ""
-            chat_response_content = []
-            chat_response_tool_calls: dict[int, OpenAIChatCompletionToolCall] = {}
-            # TODO: these chunk_ fields are hacky and only take the last chunk into account
-            chunk_created = 0
-            chunk_model = ""
-            chunk_finish_reason = ""
-            async for chunk in chat_response:
-                chat_response_id = chunk.id
-                chunk_created = chunk.created
-                chunk_model = chunk.model
-                for chunk_choice in chunk.choices:
-                    # TODO: this only works for text content
-                    chat_response_content.append(chunk_choice.delta.content or "")
-                    if chunk_choice.finish_reason:
-                        chunk_finish_reason = chunk_choice.finish_reason
-
-                    # Aggregate tool call arguments across chunks, using their index as the aggregation key
-                    if chunk_choice.delta.tool_calls:
-                        for tool_call in chunk_choice.delta.tool_calls:
-                            response_tool_call = chat_response_tool_calls.get(tool_call.index, None)
-                            if response_tool_call:
-                                response_tool_call.function.arguments += tool_call.function.arguments
-                            else:
-                                tool_call_dict: dict[str, Any] = tool_call.model_dump()
-                                # Ensure we don't have any empty type field in the tool call dict.
-                                # The OpenAI client used by providers often returns a type=None here.
-                                tool_call_dict.pop("type", None)
-                                response_tool_call = OpenAIChatCompletionToolCall(**tool_call_dict)
-                            chat_response_tool_calls[tool_call.index] = response_tool_call
-
-            # Convert the dict of tool calls by index to a list of tool calls to pass back in our response
-            if chat_response_tool_calls:
-                tool_calls = [chat_response_tool_calls[i] for i in sorted(chat_response_tool_calls.keys())]
-            else:
-                tool_calls = None
-            assistant_message = OpenAIAssistantMessageParam(
-                content="".join(chat_response_content),
-                tool_calls=tool_calls,
-            )
-            chat_response = OpenAIChatCompletion(
-                id=chat_response_id,
-                choices=[
-                    OpenAIChoice(
-                        message=assistant_message,
-                        finish_reason=chunk_finish_reason,
-                        index=0,
-                    )
-                ],
-                created=chunk_created,
-                model=chunk_model,
-            )
-        else:
-            # dump and reload to map to our pydantic types
-            chat_response = OpenAIChatCompletion(**chat_response.model_dump())
+        :param response_id: The ID of the response to retrieve input items for.
+        :param after: An item ID to list items after, used for pagination.
+        :param before: An item ID to list items before, used for pagination.
+        :param include: Additional fields to include in the response.
+        :param limit: A limit on the number of objects to be returned.
+        :param order: The order to return the input items in.
+        :returns: An ListOpenAIResponseInputItem.
+        """
+        return await self.responses_store.list_response_input_items(response_id, after, before, include, limit, order)
 
+    async def _process_response_choices(
+        self,
+        chat_response: OpenAIChatCompletion,
+        ctx: ChatCompletionContext,
+        tools: list[OpenAIResponseInputTool] | None,
+    ) -> list[OpenAIResponseOutput]:
+        """Handle tool execution and response message creation."""
         output_messages: list[OpenAIResponseOutput] = []
+        # Execute tool calls if any
         for choice in chat_response.choices:
             if choice.message.tool_calls and tools:
                 # Assume if the first tool is a function, all tools are functions
-                if isinstance(tools[0], OpenAIResponseInputToolFunction):
+                if tools[0].type == "function":
                     for tool_call in choice.message.tool_calls:
                         output_messages.append(
                             OpenAIResponseOutputMessageFunctionToolCall(
@@ -312,11 +282,133 @@ class OpenAIResponsesImpl:
                             )
                         )
                 else:
-                    output_messages.extend(
-                        await self._execute_tool_and_return_final_output(model, stream, choice, messages, temperature)
-                    )
+                    tool_messages = await self._execute_tool_and_return_final_output(choice, ctx)
+                    output_messages.extend(tool_messages)
             else:
                 output_messages.append(await _convert_chat_choice_to_response_message(choice))
+
+        return output_messages
+
+    async def _store_response(
+        self,
+        response: OpenAIResponseObject,
+        original_input: str | list[OpenAIResponseInput],
+    ) -> None:
+        new_input_id = f"msg_{uuid.uuid4()}"
+        if isinstance(original_input, str):
+            # synthesize a message from the input string
+            input_content = OpenAIResponseInputMessageContentText(text=original_input)
+            input_content_item = OpenAIResponseMessage(
+                role="user",
+                content=[input_content],
+                id=new_input_id,
+            )
+            input_items_data = [input_content_item]
+        else:
+            # we already have a list of messages
+            input_items_data = []
+            for input_item in original_input:
+                if isinstance(input_item, OpenAIResponseMessage):
+                    # These may or may not already have an id, so dump to dict, check for id, and add if missing
+                    input_item_dict = input_item.model_dump()
+                    if "id" not in input_item_dict:
+                        input_item_dict["id"] = new_input_id
+                    input_items_data.append(OpenAIResponseMessage(**input_item_dict))
+                else:
+                    input_items_data.append(input_item)
+
+        await self.responses_store.store_response_object(
+            response_object=response,
+            input=input_items_data,
+        )
+
+    async def create_openai_response(
+        self,
+        input: str | list[OpenAIResponseInput],
+        model: str,
+        instructions: str | None = None,
+        previous_response_id: str | None = None,
+        store: bool | None = True,
+        stream: bool | None = False,
+        temperature: float | None = None,
+        tools: list[OpenAIResponseInputTool] | None = None,
+    ):
+        stream = False if stream is None else stream
+        original_input = input  # Keep reference for storage
+
+        output_messages: list[OpenAIResponseOutput] = []
+
+        # Input preprocessing
+        input = await self._prepend_previous_response(input, previous_response_id)
+        messages = await _convert_response_input_to_chat_messages(input)
+        await self._prepend_instructions(messages, instructions)
+
+        # Tool setup
+        chat_tools, mcp_tool_to_server, mcp_list_message = (
+            await self._convert_response_tools_to_chat_tools(tools) if tools else (None, {}, None)
+        )
+        if mcp_list_message:
+            output_messages.append(mcp_list_message)
+
+        ctx = ChatCompletionContext(
+            model=model,
+            messages=messages,
+            tools=chat_tools,
+            mcp_tool_to_server=mcp_tool_to_server,
+            stream=stream,
+            temperature=temperature,
+        )
+
+        inference_result = await self.inference_api.openai_chat_completion(
+            model=model,
+            messages=messages,
+            tools=chat_tools,
+            stream=stream,
+            temperature=temperature,
+        )
+
+        if stream:
+            return self._create_streaming_response(
+                inference_result=inference_result,
+                ctx=ctx,
+                output_messages=output_messages,
+                original_input=original_input,
+                model=model,
+                store=store,
+                tools=tools,
+            )
+        else:
+            return await self._create_non_streaming_response(
+                inference_result=inference_result,
+                ctx=ctx,
+                output_messages=output_messages,
+                original_input=original_input,
+                model=model,
+                store=store,
+                tools=tools,
+            )
+
+    async def _create_non_streaming_response(
+        self,
+        inference_result: Any,
+        ctx: ChatCompletionContext,
+        output_messages: list[OpenAIResponseOutput],
+        original_input: str | list[OpenAIResponseInput],
+        model: str,
+        store: bool | None,
+        tools: list[OpenAIResponseInputTool] | None,
+    ) -> OpenAIResponseObject:
+        chat_response = OpenAIChatCompletion(**inference_result.model_dump())
+
+        # Process response choices (tool execution and message creation)
+        output_messages.extend(
+            await self._process_response_choices(
+                chat_response=chat_response,
+                ctx=ctx,
+                tools=tools,
+            )
+        )
+
         response = OpenAIResponseObject(
             created_at=chat_response.created,
             id=f"resp-{uuid.uuid4()}",
@@ -327,57 +419,168 @@ class OpenAIResponsesImpl:
         )
         logger.debug(f"OpenAI Responses response: {response}")
 
+        # Store response if requested
         if store:
-            # Store in kvstore
-
-            new_input_id = f"msg_{uuid.uuid4()}"
-            if isinstance(input, str):
-                # synthesize a message from the input string
-                input_content = OpenAIResponseInputMessageContentText(text=input)
-                input_content_item = OpenAIResponseMessage(
-                    role="user",
-                    content=[input_content],
-                    id=new_input_id,
-                )
-                input_items_data = [input_content_item]
-            else:
-                # we already have a list of messages
-                input_items_data = []
-                for input_item in input:
-                    if isinstance(input_item, OpenAIResponseMessage):
-                        # These may or may not already have an id, so dump to dict, check for id, and add if missing
-                        input_item_dict = input_item.model_dump()
-                        if "id" not in input_item_dict:
-                            input_item_dict["id"] = new_input_id
-                        input_items_data.append(OpenAIResponseMessage(**input_item_dict))
-                    else:
-                        input_items_data.append(input_item)
-
-            input_items = OpenAIResponseInputItemList(data=input_items_data)
-            prev_response = OpenAIResponsePreviousResponseWithInputItems(
-                input_items=input_items,
+            await self._store_response(
                 response=response,
+                original_input=original_input,
             )
-            key = f"{OPENAI_RESPONSES_PREFIX}{response.id}"
-            await self.persistence_store.set(
-                key=key,
-                value=prev_response.model_dump_json(),
-            )
-
-        if stream:
-
-            async def async_response() -> AsyncIterator[OpenAIResponseObjectStream]:
-                # TODO: response created should actually get emitted much earlier in the process
-                yield OpenAIResponseObjectStreamResponseCreated(response=response)
-                yield OpenAIResponseObjectStreamResponseCompleted(response=response)
-
-            return async_response()
 
         return response
 
+    async def _create_streaming_response(
+        self,
+        inference_result: Any,
+        ctx: ChatCompletionContext,
+        output_messages: list[OpenAIResponseOutput],
+        original_input: str | list[OpenAIResponseInput],
+        model: str,
+        store: bool | None,
+        tools: list[OpenAIResponseInputTool] | None,
+    ) -> AsyncIterator[OpenAIResponseObjectStream]:
+        # Create initial response and emit response.created immediately
+        response_id = f"resp-{uuid.uuid4()}"
+        created_at = int(time.time())
+
+        initial_response = OpenAIResponseObject(
+            created_at=created_at,
+            id=response_id,
+            model=model,
+            object="response",
+            status="in_progress",
+            output=output_messages.copy(),
+        )
+
+        # Emit response.created immediately
+        yield OpenAIResponseObjectStreamResponseCreated(response=initial_response)
+
+        # For streaming, inference_result is an async iterator of chunks
+        # Stream chunks and emit delta events as they arrive
+        chat_response_id = ""
+        chat_response_content = []
+        chat_response_tool_calls: dict[int, OpenAIChatCompletionToolCall] = {}
+        chunk_created = 0
+        chunk_model = ""
+        chunk_finish_reason = ""
+        sequence_number = 0
+
+        # Create a placeholder message item for delta events
+        message_item_id = f"msg_{uuid.uuid4()}"
+
+        async for chunk in inference_result:
+            chat_response_id = chunk.id
+            chunk_created = chunk.created
+            chunk_model = chunk.model
+            for chunk_choice in chunk.choices:
+                # Emit incremental text content as delta events
+                if chunk_choice.delta.content:
+                    sequence_number += 1
+                    yield OpenAIResponseObjectStreamResponseOutputTextDelta(
+                        content_index=0,
+                        delta=chunk_choice.delta.content,
+                        item_id=message_item_id,
+                        output_index=0,
+                        sequence_number=sequence_number,
+                    )
+
+                # Collect content for final response
+                chat_response_content.append(chunk_choice.delta.content or "")
+                if chunk_choice.finish_reason:
+                    chunk_finish_reason = chunk_choice.finish_reason
+
+                # Aggregate tool call arguments across chunks, using their index as the aggregation key
+                if chunk_choice.delta.tool_calls:
+                    for tool_call in chunk_choice.delta.tool_calls:
+                        response_tool_call = chat_response_tool_calls.get(tool_call.index, None)
+                        if response_tool_call:
+                            response_tool_call.function.arguments += tool_call.function.arguments
+                        else:
+                            tool_call_dict: dict[str, Any] = tool_call.model_dump()
+                            tool_call_dict.pop("type", None)
+                            response_tool_call = OpenAIChatCompletionToolCall(**tool_call_dict)
+                        chat_response_tool_calls[tool_call.index] = response_tool_call
+
+        # Convert collected chunks to complete response
+        if chat_response_tool_calls:
+            tool_calls = [chat_response_tool_calls[i] for i in sorted(chat_response_tool_calls.keys())]
+        else:
+            tool_calls = None
+        assistant_message = OpenAIAssistantMessageParam(
+            content="".join(chat_response_content),
+            tool_calls=tool_calls,
+        )
+        chat_response_obj = OpenAIChatCompletion(
+            id=chat_response_id,
+            choices=[
+                OpenAIChoice(
+                    message=assistant_message,
+                    finish_reason=chunk_finish_reason,
+                    index=0,
+                )
+            ],
+            created=chunk_created,
+            model=chunk_model,
+        )
+
+        # Process response choices (tool execution and message creation)
+        output_messages.extend(
+            await self._process_response_choices(
+                chat_response=chat_response_obj,
+                ctx=ctx,
+                tools=tools,
+            )
+        )
+
+        # Create final response
+        final_response = OpenAIResponseObject(
+            created_at=created_at,
+            id=response_id,
+            model=model,
+            object="response",
+            status="completed",
+            output=output_messages,
+        )
+
+        if store:
+            await self._store_response(
+                response=final_response,
+                original_input=original_input,
+            )
+
+        # Emit response.completed
+        yield OpenAIResponseObjectStreamResponseCompleted(response=final_response)
+
     async def _convert_response_tools_to_chat_tools(
         self, tools: list[OpenAIResponseInputTool]
-    ) -> list[ChatCompletionToolParam]:
+    ) -> tuple[
+        list[ChatCompletionToolParam],
+        dict[str, OpenAIResponseInputToolMCP],
+        OpenAIResponseOutput | None,
+    ]:
+        from llama_stack.apis.agents.openai_responses import (
+            MCPListToolsTool,
+        )
+        from llama_stack.apis.tools.tools import Tool
+
+        mcp_tool_to_server = {}
+
+        def make_openai_tool(tool_name: str, tool: Tool) -> ChatCompletionToolParam:
+            tool_def = ToolDefinition(
+                tool_name=tool_name,
+                description=tool.description,
+                parameters={
+                    param.name: ToolParamDefinition(
+                        param_type=param.parameter_type,
+                        description=param.description,
+                        required=param.required,
+                        default=param.default,
+                    )
+                    for param in tool.parameters
+                },
+            )
+            return convert_tooldef_to_openai_tool(tool_def)
+
+        mcp_list_message = None
         chat_tools: list[ChatCompletionToolParam] = []
         for input_tool in tools:
             # TODO: Handle other tool types
@@ -386,91 +589,95 @@ class OpenAIResponsesImpl:
             elif input_tool.type == "web_search":
                 tool_name = "web_search"
                 tool = await self.tool_groups_api.get_tool(tool_name)
-                tool_def = ToolDefinition(
-                    tool_name=tool_name,
-                    description=tool.description,
-                    parameters={
-                        param.name: ToolParamDefinition(
-                            param_type=param.parameter_type,
-                            description=param.description,
-                            required=param.required,
-                            default=param.default,
-                        )
-                        for param in tool.parameters
-                    },
+                if not tool:
+                    raise ValueError(f"Tool {tool_name} not found")
+                chat_tools.append(make_openai_tool(tool_name, tool))
+            elif input_tool.type == "mcp":
+                always_allowed = None
+                never_allowed = None
+                if input_tool.allowed_tools:
+                    if isinstance(input_tool.allowed_tools, list):
+                        always_allowed = input_tool.allowed_tools
+                    elif isinstance(input_tool.allowed_tools, AllowedToolsFilter):
+                        always_allowed = input_tool.allowed_tools.always
+                        never_allowed = input_tool.allowed_tools.never
+
+                tool_defs = await list_mcp_tools(
+                    endpoint=input_tool.server_url,
+                    headers=input_tool.headers or {},
                 )
-                chat_tool = convert_tooldef_to_openai_tool(tool_def)
-                chat_tools.append(chat_tool)
+
+                mcp_list_message = OpenAIResponseOutputMessageMCPListTools(
+                    id=f"mcp_list_{uuid.uuid4()}",
+                    status="completed",
+                    server_label=input_tool.server_label,
+                    tools=[],
+                )
+                for t in tool_defs.data:
+                    if never_allowed and t.name in never_allowed:
+                        continue
+                    if not always_allowed or t.name in always_allowed:
+                        chat_tools.append(make_openai_tool(t.name, t))
+                        if t.name in mcp_tool_to_server:
+                            raise ValueError(f"Duplicate tool name {t.name} found for server {input_tool.server_label}")
+                        mcp_tool_to_server[t.name] = input_tool
+                        mcp_list_message.tools.append(
+                            MCPListToolsTool(
+                                name=t.name,
+                                description=t.description,
+                                input_schema={
+                                    "type": "object",
+                                    "properties": {
+                                        p.name: {
+                                            "type": p.parameter_type,
+                                            "description": p.description,
+                                        }
+                                        for p in t.parameters
+                                    },
+                                    "required": [p.name for p in t.parameters if p.required],
+                                },
+                            )
+                        )
             else:
                 raise ValueError(f"Llama Stack OpenAI Responses does not yet support tool type: {input_tool.type}")
-        return chat_tools
+        return chat_tools, mcp_tool_to_server, mcp_list_message
 
     async def _execute_tool_and_return_final_output(
         self,
-        model_id: str,
-        stream: bool,
         choice: OpenAIChoice,
-        messages: list[OpenAIMessageParam],
-        temperature: float,
+        ctx: ChatCompletionContext,
     ) -> list[OpenAIResponseOutput]:
         output_messages: list[OpenAIResponseOutput] = []
 
-        # If the choice is not an assistant message, we don't need to execute any tools
         if not isinstance(choice.message, OpenAIAssistantMessageParam):
             return output_messages
 
-        # If the assistant message doesn't have any tool calls, we don't need to execute any tools
         if not choice.message.tool_calls:
             return output_messages
 
-        # Copy the messages list to avoid mutating the original list
-        messages = messages.copy()
+        next_turn_messages = ctx.messages.copy()
 
         # Add the assistant message with tool_calls response to the messages list
-        messages.append(choice.message)
+        next_turn_messages.append(choice.message)
 
         for tool_call in choice.message.tool_calls:
-            tool_call_id = tool_call.id
-            function = tool_call.function
-
-            # If for some reason the tool call doesn't have a function or id, we can't execute it
-            if not function or not tool_call_id:
-                continue
-
             # TODO: telemetry spans for tool calls
-            result = await self._execute_tool_call(function)
-
-            # Handle tool call failure
-            if not result:
-                output_messages.append(
-                    OpenAIResponseOutputMessageWebSearchToolCall(
-                        id=tool_call_id,
-                        status="failed",
-                    )
-                )
-                continue
-
-            output_messages.append(
-                OpenAIResponseOutputMessageWebSearchToolCall(
-                    id=tool_call_id,
-                    status="completed",
-                ),
-            )
-
-            result_content = ""
-            # TODO: handle other result content types and lists
-            if isinstance(result.content, str):
-                result_content = result.content
-            messages.append(OpenAIToolMessageParam(content=result_content, tool_call_id=tool_call_id))
+            tool_call_log, further_input = await self._execute_tool_call(tool_call, ctx)
+            if tool_call_log:
+                output_messages.append(tool_call_log)
+            if further_input:
+                next_turn_messages.append(further_input)
 
         tool_results_chat_response = await self.inference_api.openai_chat_completion(
-            model=model_id,
-            messages=messages,
-            stream=stream,
-            temperature=temperature,
+            model=ctx.model,
+            messages=next_turn_messages,
+            stream=ctx.stream,
+            temperature=ctx.temperature,
         )
-        # type cast to appease mypy
+        # type cast to appease mypy: this is needed because we don't handle streaming properly :)
         tool_results_chat_response = cast(OpenAIChatCompletion, tool_results_chat_response)
+
+        # Huge TODO: these are NOT the final outputs, we must keep the loop going
         tool_final_outputs = [
             await _convert_chat_choice_to_response_message(choice) for choice in tool_results_chat_response.choices
         ]
@@ -480,15 +687,86 @@ class OpenAIResponsesImpl:
 
     async def _execute_tool_call(
         self,
-        function: OpenAIChatCompletionToolCallFunction,
-    ) -> ToolInvocationResult | None:
-        if not function.name:
-            return None
-        function_args = json.loads(function.arguments) if function.arguments else {}
-        logger.info(f"executing tool call: {function.name} with args: {function_args}")
-        result = await self.tool_runtime_api.invoke_tool(
-            tool_name=function.name,
-            kwargs=function_args,
+        tool_call: OpenAIChatCompletionToolCall,
+        ctx: ChatCompletionContext,
+    ) -> tuple[OpenAIResponseOutput | None, OpenAIMessageParam | None]:
+        from llama_stack.providers.utils.inference.prompt_adapter import (
+            interleaved_content_as_str,
         )
-        logger.debug(f"tool call {function.name} completed with result: {result}")
-        return result
+
+        tool_call_id = tool_call.id
+        function = tool_call.function
+
+        if not function or not tool_call_id or not function.name:
+            return None, None
+
+        error_exc = None
+        result = None
+        try:
+            if function.name in ctx.mcp_tool_to_server:
+                mcp_tool = ctx.mcp_tool_to_server[function.name]
+                result = await invoke_mcp_tool(
+                    endpoint=mcp_tool.server_url,
+                    headers=mcp_tool.headers or {},
+                    tool_name=function.name,
+                    kwargs=json.loads(function.arguments) if function.arguments else {},
+                )
+            else:
+                result = await self.tool_runtime_api.invoke_tool(
+                    tool_name=function.name,
+                    kwargs=json.loads(function.arguments) if function.arguments else {},
+                )
+        except Exception as e:
+            error_exc = e
+
+        if function.name in ctx.mcp_tool_to_server:
+            from llama_stack.apis.agents.openai_responses import OpenAIResponseOutputMessageMCPCall
+
+            message = OpenAIResponseOutputMessageMCPCall(
+                id=tool_call_id,
+                arguments=function.arguments,
+                name=function.name,
+                server_label=ctx.mcp_tool_to_server[function.name].server_label,
+            )
+            if error_exc:
+                message.error = str(error_exc)
+            elif (result.error_code and result.error_code > 0) or result.error_message:
+                message.error = f"Error (code {result.error_code}): {result.error_message}"
+            elif result.content:
+                message.output = interleaved_content_as_str(result.content)
+        else:
+            if function.name == "web_search":
+                message = OpenAIResponseOutputMessageWebSearchToolCall(
+                    id=tool_call_id,
+                    status="completed",
+                )
+                if error_exc or (result.error_code and result.error_code > 0) or result.error_message:
+                    message.status = "failed"
+            else:
+                raise ValueError(f"Unknown tool {function.name} called")
+
+        input_message = None
+        if result and result.content:
+            if isinstance(result.content, str):
+                content = result.content
+            elif isinstance(result.content, list):
+                from llama_stack.apis.common.content_types import ImageContentItem, TextContentItem
+
+                content = []
+                for item in result.content:
+                    if isinstance(item, TextContentItem):
+                        part = OpenAIChatCompletionContentPartTextParam(text=item.text)
+                    elif isinstance(item, ImageContentItem):
+                        if item.image.data:
+                            url = f"data:image;base64,{item.image.data}"
+                        else:
+                            url = item.image.url
+                        part = OpenAIChatCompletionContentPartImageParam(image_url=OpenAIImageURL(url=url))
+                    else:
+                        raise ValueError(f"Unknown result content type: {type(item)}")
+                    content.append(part)
+            else:
+                raise ValueError(f"Unknown result content type: {type(result.content)}")
+            input_message = OpenAIToolMessageParam(content=content, tool_call_id=tool_call_id)
+
+        return message, input_message

llama_stack/providers/inline/inference/meta_reference/inference.py

@@ -6,6 +6,7 @@
 
 import asyncio
 import os
+import sys
 from collections.abc import AsyncGenerator
 
 from pydantic import BaseModel
@@ -455,9 +456,9 @@ class MetaReferenceInferenceImpl(
                 first = token_results[0]
                 if not first.finished and not first.ignore_token:
                     if os.environ.get("LLAMA_MODELS_DEBUG", "0") in ("1", "2"):
-                        cprint(first.text, "cyan", end="")
+                        cprint(first.text, color="cyan", end="", file=sys.stderr)
                     if os.environ.get("LLAMA_MODELS_DEBUG", "0") == "2":
-                        cprint(f"<{first.token}>", "magenta", end="")
+                        cprint(f"<{first.token}>", color="magenta", end="", file=sys.stderr)
 
                 for result in token_results:
                     idx = result.batch_idx
@@ -519,9 +520,9 @@ class MetaReferenceInferenceImpl(
             for token_results in self.generator.chat_completion([request]):
                 token_result = token_results[0]
                 if os.environ.get("LLAMA_MODELS_DEBUG", "0") == "1":
-                    cprint(token_result.text, "cyan", end="")
+                    cprint(token_result.text, color="cyan", end="", file=sys.stderr)
                 if os.environ.get("LLAMA_MODELS_DEBUG", "0") == "2":
-                    cprint(f"<{token_result.token}>", "magenta", end="")
+                    cprint(f"<{token_result.token}>", color="magenta", end="", file=sys.stderr)
 
                 if token_result.token == tokenizer.eot_id:
                     stop_reason = StopReason.end_of_turn

llama_stack/providers/inline/telemetry/meta_reference/telemetry.py

@@ -16,6 +16,7 @@ from opentelemetry.sdk.resources import Resource
 from opentelemetry.sdk.trace import TracerProvider
 from opentelemetry.sdk.trace.export import BatchSpanProcessor
 from opentelemetry.semconv.resource import ResourceAttributes
+from opentelemetry.trace.propagation.tracecontext import TraceContextTextMapPropagator
 
 from llama_stack.apis.telemetry import (
     Event,
@@ -44,6 +45,7 @@ from llama_stack.providers.inline.telemetry.meta_reference.sqlite_span_processor
 )
 from llama_stack.providers.utils.telemetry.dataset_mixin import TelemetryDatasetMixin
 from llama_stack.providers.utils.telemetry.sqlite_trace_store import SQLiteTraceStore
+from llama_stack.providers.utils.telemetry.tracing import ROOT_SPAN_MARKERS
 
 from .config import TelemetryConfig, TelemetrySink
 
@@ -146,7 +148,7 @@ class TelemetryAdapter(TelemetryDatasetMixin, Telemetry):
             if span:
                 timestamp_ns = int(event.timestamp.timestamp() * 1e9)
                 span.add_event(
-                    name=event.type,
+                    name=event.type.value,
                     attributes={
                         "message": event.message,
                         "severity": event.severity.value,
@@ -206,6 +208,15 @@ class TelemetryAdapter(TelemetryDatasetMixin, Telemetry):
                 event.attributes = {}
             event.attributes["__ttl__"] = ttl_seconds
 
+            # Extract these W3C trace context attributes so they are not written to
+            # underlying storage, as we just need them to propagate the trace context.
+            traceparent = event.attributes.pop("traceparent", None)
+            tracestate = event.attributes.pop("tracestate", None)
+            if traceparent:
+                # If we have a traceparent header value, we're not the root span.
+                for root_attribute in ROOT_SPAN_MARKERS:
+                    event.attributes.pop(root_attribute, None)
+
             if isinstance(event.payload, SpanStartPayload):
                 # Check if span already exists to prevent duplicates
                 if span_id in _GLOBAL_STORAGE["active_spans"]:
@@ -216,8 +227,12 @@ class TelemetryAdapter(TelemetryDatasetMixin, Telemetry):
                     parent_span_id = int(event.payload.parent_span_id, 16)
                     parent_span = _GLOBAL_STORAGE["active_spans"].get(parent_span_id)
                     context = trace.set_span_in_context(parent_span)
-                else:
-                    event.attributes["__root_span__"] = "true"
+                elif traceparent:
+                    carrier = {
+                        "traceparent": traceparent,
+                        "tracestate": tracestate,
+                    }
+                    context = TraceContextTextMapPropagator().extract(carrier=carrier)
 
                 span = tracer.start_span(
                     name=event.payload.name,

llama_stack/providers/inline/tool_runtime/rag/memory.py

@@ -25,14 +25,14 @@ from llama_stack.apis.tools import (
     RAGQueryConfig,
     RAGQueryResult,
     RAGToolRuntime,
-    Tool,
     ToolDef,
+    ToolGroup,
     ToolInvocationResult,
     ToolParameter,
     ToolRuntime,
 )
 from llama_stack.apis.vector_io import QueryChunksResponse, VectorIO
-from llama_stack.providers.datatypes import ToolsProtocolPrivate
+from llama_stack.providers.datatypes import ToolGroupsProtocolPrivate
 from llama_stack.providers.utils.inference.prompt_adapter import interleaved_content_as_str
 from llama_stack.providers.utils.memory.vector_store import (
     content_from_doc,
@@ -49,7 +49,7 @@ def make_random_string(length: int = 8):
     return "".join(secrets.choice(string.ascii_letters + string.digits) for _ in range(length))
 
 
-class MemoryToolRuntimeImpl(ToolsProtocolPrivate, ToolRuntime, RAGToolRuntime):
+class MemoryToolRuntimeImpl(ToolGroupsProtocolPrivate, ToolRuntime, RAGToolRuntime):
     def __init__(
         self,
         config: RagToolRuntimeConfig,
@@ -66,10 +66,10 @@ class MemoryToolRuntimeImpl(ToolsProtocolPrivate, ToolRuntime, RAGToolRuntime):
     async def shutdown(self):
         pass
 
-    async def register_tool(self, tool: Tool) -> None:
+    async def register_toolgroup(self, toolgroup: ToolGroup) -> None:
         pass
 
-    async def unregister_tool(self, tool_id: str) -> None:
+    async def unregister_toolgroup(self, toolgroup_id: str) -> None:
         return
 
     async def insert(
@@ -122,6 +122,7 @@ class MemoryToolRuntimeImpl(ToolsProtocolPrivate, ToolRuntime, RAGToolRuntime):
                 query=query,
                 params={
                     "max_chunks": query_config.max_chunks,
+                    "mode": query_config.mode,
                 },
             )
             for vector_db_id in vector_db_ids
@@ -146,7 +147,7 @@ class MemoryToolRuntimeImpl(ToolsProtocolPrivate, ToolRuntime, RAGToolRuntime):
         for i, chunk in enumerate(chunks):
             metadata = chunk.metadata
             tokens += metadata["token_count"]
-            tokens += metadata["metadata_token_count"]
+            tokens += metadata.get("metadata_token_count", 0)
 
             if tokens > query_config.max_tokens_in_context:
                 log.error(

llama_stack/providers/inline/vector_io/faiss/faiss.py

@@ -99,9 +99,13 @@ class FaissIndex(EmbeddingIndex):
         # Save updated index
         await self._save_index()
 
-    async def query(self, embedding: NDArray, k: int, score_threshold: float) -> QueryChunksResponse:
+    async def query_vector(
+        self,
+        embedding: NDArray,
+        k: int,
+        score_threshold: float,
+    ) -> QueryChunksResponse:
         distances, indices = await asyncio.to_thread(self.index.search, embedding.reshape(1, -1).astype(np.float32), k)
-
         chunks = []
         scores = []
         for d, i in zip(distances[0], indices[0], strict=False):
@@ -112,6 +116,14 @@ class FaissIndex(EmbeddingIndex):
 
         return QueryChunksResponse(chunks=chunks, scores=scores)
 
+    async def query_keyword(
+        self,
+        query_string: str,
+        k: int,
+        score_threshold: float,
+    ) -> QueryChunksResponse:
+        raise NotImplementedError("Keyword search is not supported in FAISS")
+
 
 class FaissVectorIOAdapter(VectorIO, VectorDBsProtocolPrivate):
     def __init__(self, config: FaissVectorIOConfig, inference_api: Inference) -> None:

llama_stack/providers/inline/vector_io/sqlite_vec/sqlite_vec.py

@@ -24,6 +24,11 @@ from llama_stack.providers.utils.memory.vector_store import EmbeddingIndex, Vect
 
 logger = logging.getLogger(__name__)
 
+# Specifying search mode is dependent on the VectorIO provider.
+VECTOR_SEARCH = "vector"
+KEYWORD_SEARCH = "keyword"
+SEARCH_MODES = {VECTOR_SEARCH, KEYWORD_SEARCH}
+
 
 def serialize_vector(vector: list[float]) -> bytes:
     """Serialize a list of floats into a compact binary representation."""
@@ -45,6 +50,7 @@ class SQLiteVecIndex(EmbeddingIndex):
     Two tables are used:
       - A metadata table (chunks_{bank_id}) that holds the chunk JSON.
       - A virtual table (vec_chunks_{bank_id}) that holds the serialized vector.
+      - An FTS5 table (fts_chunks_{bank_id}) for full-text keyword search.
     """
 
     def __init__(self, dimension: int, db_path: str, bank_id: str):
@@ -53,6 +59,7 @@ class SQLiteVecIndex(EmbeddingIndex):
         self.bank_id = bank_id
         self.metadata_table = f"chunks_{bank_id}".replace("-", "_")
         self.vector_table = f"vec_chunks_{bank_id}".replace("-", "_")
+        self.fts_table = f"fts_chunks_{bank_id}".replace("-", "_")
 
     @classmethod
     async def create(cls, dimension: int, db_path: str, bank_id: str):
@@ -78,6 +85,14 @@ class SQLiteVecIndex(EmbeddingIndex):
                     USING vec0(embedding FLOAT[{self.dimension}], id TEXT);
                 """)
                 connection.commit()
+                # FTS5 table (for keyword search) - creating both the tables by default. Will use the relevant one
+                # based on query. Implementation of the change on client side will allow passing the search_mode option
+                # during initialization to make it easier to create the table that is required.
+                cur.execute(f"""
+                            CREATE VIRTUAL TABLE IF NOT EXISTS {self.fts_table}
+                            USING fts5(id, content);
+                        """)
+                connection.commit()
             finally:
                 cur.close()
                 connection.close()
@@ -91,6 +106,7 @@ class SQLiteVecIndex(EmbeddingIndex):
             try:
                 cur.execute(f"DROP TABLE IF EXISTS {self.metadata_table};")
                 cur.execute(f"DROP TABLE IF EXISTS {self.vector_table};")
+                cur.execute(f"DROP TABLE IF EXISTS {self.fts_table};")
                 connection.commit()
             finally:
                 cur.close()
@@ -104,6 +120,7 @@ class SQLiteVecIndex(EmbeddingIndex):
         For each chunk, we insert its JSON into the metadata table and then insert its
         embedding (serialized to raw bytes) into the virtual table using the assigned rowid.
         If any insert fails, the transaction is rolled back to maintain consistency.
+        Also inserts chunk content into FTS table for keyword search support.
         """
         assert all(isinstance(chunk.content, str) for chunk in chunks), "SQLiteVecIndex only supports text chunks"
 
@@ -112,18 +129,16 @@ class SQLiteVecIndex(EmbeddingIndex):
             cur = connection.cursor()
 
             try:
-                # Start transaction a single transcation for all batches
                 cur.execute("BEGIN TRANSACTION")
                 for i in range(0, len(chunks), batch_size):
                     batch_chunks = chunks[i : i + batch_size]
                     batch_embeddings = embeddings[i : i + batch_size]
-                    # Prepare metadata inserts
+
+                    # Insert metadata
                     metadata_data = [
                         (generate_chunk_id(chunk.metadata["document_id"], chunk.content), chunk.model_dump_json())
                         for chunk in batch_chunks
-                        if isinstance(chunk.content, str)
                     ]
-                    # Insert metadata (ON CONFLICT to avoid duplicates)
                     cur.executemany(
                         f"""
                         INSERT INTO {self.metadata_table} (id, chunk)
@@ -132,21 +147,43 @@ class SQLiteVecIndex(EmbeddingIndex):
                         """,
                         metadata_data,
                     )
-                    # Prepare embeddings inserts
+
+                    # Insert vector embeddings
                     embedding_data = [
                         (
-                            generate_chunk_id(chunk.metadata["document_id"], chunk.content),
-                            serialize_vector(emb.tolist()),
+                            (
+                                generate_chunk_id(chunk.metadata["document_id"], chunk.content),
+                                serialize_vector(emb.tolist()),
+                            )
                         )
                         for chunk, emb in zip(batch_chunks, batch_embeddings, strict=True)
-                        if isinstance(chunk.content, str)
                     ]
-                    # Insert embeddings in batch
-                    cur.executemany(f"INSERT INTO {self.vector_table} (id, embedding) VALUES (?, ?);", embedding_data)
+                    cur.executemany(
+                        f"INSERT INTO {self.vector_table} (id, embedding) VALUES (?, ?);",
+                        embedding_data,
+                    )
+
+                    # Insert FTS content
+                    fts_data = [
+                        (generate_chunk_id(chunk.metadata["document_id"], chunk.content), chunk.content)
+                        for chunk in batch_chunks
+                    ]
+                    # DELETE existing entries with same IDs (FTS5 doesn't support ON CONFLICT)
+                    cur.executemany(
+                        f"DELETE FROM {self.fts_table} WHERE id = ?;",
+                        [(row[0],) for row in fts_data],
+                    )
+
+                    # INSERT new entries
+                    cur.executemany(
+                        f"INSERT INTO {self.fts_table} (id, content) VALUES (?, ?);",
+                        fts_data,
+                    )
+
                 connection.commit()
 
             except sqlite3.Error as e:
-                connection.rollback()  # Rollback on failure
+                connection.rollback()
                 logger.error(f"Error inserting into {self.vector_table}: {e}")
                 raise
 
@@ -154,22 +191,25 @@ class SQLiteVecIndex(EmbeddingIndex):
                 cur.close()
                 connection.close()
 
-        # Process all batches in a single thread
+        # Run batch insertion in a background thread
         await asyncio.to_thread(_execute_all_batch_inserts)
 
-    async def query(self, embedding: NDArray, k: int, score_threshold: float) -> QueryChunksResponse:
+    async def query_vector(
+        self,
+        embedding: NDArray,
+        k: int,
+        score_threshold: float,
+    ) -> QueryChunksResponse:
         """
-        Query for the k most similar chunks. We convert the query embedding to a blob and run a SQL query
-        against the virtual table. The SQL joins the metadata table to recover the chunk JSON.
+        Performs vector-based search using a virtual table for vector similarity.
         """
-        emb_list = embedding.tolist() if isinstance(embedding, np.ndarray) else list(embedding)
-        emb_blob = serialize_vector(emb_list)
 
         def _execute_query():
             connection = _create_sqlite_connection(self.db_path)
             cur = connection.cursor()
-
             try:
+                emb_list = embedding.tolist() if isinstance(embedding, np.ndarray) else list(embedding)
+                emb_blob = serialize_vector(emb_list)
                 query_sql = f"""
                     SELECT m.id, m.chunk, v.distance
                     FROM {self.vector_table} AS v
@@ -184,17 +224,66 @@ class SQLiteVecIndex(EmbeddingIndex):
                 connection.close()
 
         rows = await asyncio.to_thread(_execute_query)
-
         chunks, scores = [], []
-        for _id, chunk_json, distance in rows:
+        for row in rows:
+            _id, chunk_json, distance = row
+            score = 1.0 / distance if distance != 0 else float("inf")
+            if score < score_threshold:
+                continue
+            try:
+                chunk = Chunk.model_validate_json(chunk_json)
+            except Exception as e:
+                logger.error(f"Error parsing chunk JSON for id {_id}: {e}")
+                continue
+            chunks.append(chunk)
+            scores.append(score)
+        return QueryChunksResponse(chunks=chunks, scores=scores)
+
+    async def query_keyword(
+        self,
+        query_string: str,
+        k: int,
+        score_threshold: float,
+    ) -> QueryChunksResponse:
+        """
+        Performs keyword-based search using SQLite FTS5 for relevance-ranked full-text search.
+        """
+        if query_string is None:
+            raise ValueError("query_string is required for keyword search.")
+
+        def _execute_query():
+            connection = _create_sqlite_connection(self.db_path)
+            cur = connection.cursor()
+            try:
+                query_sql = f"""
+                    SELECT DISTINCT m.id, m.chunk, bm25({self.fts_table}) AS score
+                    FROM {self.fts_table} AS f
+                    JOIN {self.metadata_table} AS m ON m.id = f.id
+                    WHERE f.content MATCH ?
+                    ORDER BY score ASC
+                    LIMIT ?;
+                """
+                cur.execute(query_sql, (query_string, k))
+                return cur.fetchall()
+            finally:
+                cur.close()
+                connection.close()
+
+        rows = await asyncio.to_thread(_execute_query)
+        chunks, scores = [], []
+        for row in rows:
+            _id, chunk_json, score = row
+            # BM25 scores returned by sqlite-vec are NEGATED (i.e., more relevant = more negative).
+            # This design is intentional to simplify sorting by ascending score.
+            # Reference: https://alexgarcia.xyz/blog/2024/sqlite-vec-hybrid-search/index.html
+            if score > -score_threshold:
+                continue
             try:
                 chunk = Chunk.model_validate_json(chunk_json)
             except Exception as e:
                 logger.error(f"Error parsing chunk JSON for id {_id}: {e}")
                 continue
             chunks.append(chunk)
-            # Mimic the Faiss scoring: score = 1/distance (avoid division by zero)
-            score = 1.0 / distance if distance != 0 else float("inf")
             scores.append(score)
         return QueryChunksResponse(chunks=chunks, scores=scores)
 

llama_stack/providers/registry/safety.py

@@ -63,4 +63,14 @@ def available_providers() -> list[ProviderSpec]:
                 config_class="llama_stack.providers.remote.safety.nvidia.NVIDIASafetyConfig",
             ),
         ),
+        remote_provider_spec(
+            api=Api.safety,
+            adapter=AdapterSpec(
+                adapter_type="sambanova",
+                pip_packages=["litellm"],
+                module="llama_stack.providers.remote.safety.sambanova",
+                config_class="llama_stack.providers.remote.safety.sambanova.SambaNovaSafetyConfig",
+                provider_data_validator="llama_stack.providers.remote.safety.sambanova.config.SambaNovaProviderDataValidator",
+            ),
+        ),
     ]

llama_stack/providers/registry/tool_runtime.py

@@ -80,8 +80,9 @@ def available_providers() -> list[ProviderSpec]:
             adapter=AdapterSpec(
                 adapter_type="model-context-protocol",
                 module="llama_stack.providers.remote.tool_runtime.model_context_protocol",
-                config_class="llama_stack.providers.remote.tool_runtime.model_context_protocol.config.ModelContextProtocolConfig",
+                config_class="llama_stack.providers.remote.tool_runtime.model_context_protocol.config.MCPProviderConfig",
                 pip_packages=["mcp"],
+                provider_data_validator="llama_stack.providers.remote.tool_runtime.model_context_protocol.config.MCPProviderDataValidator",
             ),
         ),
     ]

llama_stack/providers/remote/inference/openai/openai.py

@@ -92,8 +92,11 @@ class OpenAIInferenceAdapter(LiteLLMOpenAIMixin):
         if prompt_logprobs is not None:
             logging.warning("prompt_logprobs is not supported by the OpenAI API. Ignoring.")
 
+        model_id = (await self.model_store.get_model(model)).provider_resource_id
+        if model_id.startswith("openai/"):
+            model_id = model_id[len("openai/") :]
         params = await prepare_openai_completion_params(
-            model=(await self.model_store.get_model(model)).provider_resource_id,
+            model=model_id,
             prompt=prompt,
             best_of=best_of,
             echo=echo,
@@ -139,8 +142,11 @@ class OpenAIInferenceAdapter(LiteLLMOpenAIMixin):
         top_p: float | None = None,
         user: str | None = None,
     ) -> OpenAIChatCompletion | AsyncIterator[OpenAIChatCompletionChunk]:
+        model_id = (await self.model_store.get_model(model)).provider_resource_id
+        if model_id.startswith("openai/"):
+            model_id = model_id[len("openai/") :]
         params = await prepare_openai_completion_params(
-            model=(await self.model_store.get_model(model)).provider_resource_id,
+            model=model_id,
             messages=messages,
             frequency_penalty=frequency_penalty,
             function_call=function_call,

llama_stack/providers/remote/inference/vllm/config.py

@@ -4,8 +4,9 @@
 # This source code is licensed under the terms described in the LICENSE file in
 # the root directory of this source tree.
 
+from pathlib import Path
 
-from pydantic import BaseModel, Field
+from pydantic import BaseModel, Field, field_validator
 
 from llama_stack.schema_utils import json_schema_type
 
@@ -24,11 +25,27 @@ class VLLMInferenceAdapterConfig(BaseModel):
         default="fake",
         description="The API token",
     )
-    tls_verify: bool = Field(
+    tls_verify: bool | str = Field(
         default=True,
-        description="Whether to verify TLS certificates",
+        description="Whether to verify TLS certificates. Can be a boolean or a path to a CA certificate file.",
     )
 
+    @field_validator("tls_verify")
+    @classmethod
+    def validate_tls_verify(cls, v):
+        if isinstance(v, str):
+            # Check if it's a boolean string
+            if v.lower() in ("true", "false"):
+                return v.lower() == "true"
+            # Otherwise, treat it as a cert path
+            cert_path = Path(v).expanduser().resolve()
+            if not cert_path.exists():
+                raise ValueError(f"TLS certificate file does not exist: {v}")
+            if not cert_path.is_file():
+                raise ValueError(f"TLS certificate path is not a file: {v}")
+            return v
+        return v
+
     @classmethod
     def sample_run_config(
         cls,

llama_stack/providers/remote/inference/vllm/vllm.py

@@ -313,7 +313,7 @@ class VLLMInferenceAdapter(Inference, ModelsProtocolPrivate):
         return AsyncOpenAI(
             base_url=self.config.url,
             api_key=self.config.api_token,
-            http_client=None if self.config.tls_verify else httpx.AsyncClient(verify=False),
+            http_client=httpx.AsyncClient(verify=self.config.tls_verify),
         )
 
     async def completion(

llama_stack/providers/remote/post_training/nvidia/post_training.py

@@ -224,7 +224,7 @@ class NvidiaPostTrainingAdapter(ModelRegistryHelper):
 
         Parameters:
             training_config: TrainingConfig - Configuration for training
-            model: str - Model identifier
+            model: str - NeMo Customizer configuration name
             algorithm_config: Optional[AlgorithmConfig] - Algorithm-specific configuration
             checkpoint_dir: Optional[str] - Directory containing model checkpoints, ignored atm
             job_uuid: str - Unique identifier for the job, ignored atm
@@ -299,9 +299,6 @@ class NvidiaPostTrainingAdapter(ModelRegistryHelper):
 
             User is informed about unsupported parameters via warnings.
         """
-        # Map model to nvidia model name
-        # See `_MODEL_ENTRIES` for supported models
-        nvidia_model = self.get_provider_model_id(model)
 
         # Check for unsupported method parameters
         unsupported_method_params = []
@@ -347,7 +344,7 @@ class NvidiaPostTrainingAdapter(ModelRegistryHelper):
 
         # Prepare base job configuration
         job_config = {
-            "config": nvidia_model,
+            "config": model,
             "dataset": {
                 "name": training_config["data_config"]["dataset_id"],
                 "namespace": self.config.dataset_namespace,

llama_stack/providers/remote/safety/sambanova/__init__.py

@@ -0,0 +1,18 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the terms described in the LICENSE file in
+# the root directory of this source tree.
+
+
+from typing import Any
+
+from .config import SambaNovaSafetyConfig
+
+
+async def get_adapter_impl(config: SambaNovaSafetyConfig, _deps) -> Any:
+    from .sambanova import SambaNovaSafetyAdapter
+
+    impl = SambaNovaSafetyAdapter(config)
+    await impl.initialize()
+    return impl

llama_stack/providers/remote/safety/sambanova/config.py

@@ -0,0 +1,37 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the terms described in the LICENSE file in
+# the root directory of this source tree.
+
+from typing import Any
+
+from pydantic import BaseModel, Field, SecretStr
+
+from llama_stack.schema_utils import json_schema_type
+
+
+class SambaNovaProviderDataValidator(BaseModel):
+    sambanova_api_key: str | None = Field(
+        default=None,
+        description="Sambanova Cloud API key",
+    )
+
+
+@json_schema_type
+class SambaNovaSafetyConfig(BaseModel):
+    url: str = Field(
+        default="https://api.sambanova.ai/v1",
+        description="The URL for the SambaNova AI server",
+    )
+    api_key: SecretStr | None = Field(
+        default=None,
+        description="The SambaNova cloud API Key",
+    )
+
+    @classmethod
+    def sample_run_config(cls, api_key: str = "${env.SAMBANOVA_API_KEY}", **kwargs) -> dict[str, Any]:
+        return {
+            "url": "https://api.sambanova.ai/v1",
+            "api_key": api_key,
+        }

llama_stack/providers/remote/safety/sambanova/sambanova.py

@@ -0,0 +1,100 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the terms described in the LICENSE file in
+# the root directory of this source tree.
+
+import json
+import logging
+from typing import Any
+
+import litellm
+import requests
+
+from llama_stack.apis.inference import Message
+from llama_stack.apis.safety import (
+    RunShieldResponse,
+    Safety,
+    SafetyViolation,
+    ViolationLevel,
+)
+from llama_stack.apis.shields import Shield
+from llama_stack.distribution.request_headers import NeedsRequestProviderData
+from llama_stack.providers.datatypes import ShieldsProtocolPrivate
+from llama_stack.providers.utils.inference.openai_compat import convert_message_to_openai_dict_new
+
+from .config import SambaNovaSafetyConfig
+
+logger = logging.getLogger(__name__)
+
+CANNED_RESPONSE_TEXT = "I can't answer that. Can I help with something else?"
+
+
+class SambaNovaSafetyAdapter(Safety, ShieldsProtocolPrivate, NeedsRequestProviderData):
+    def __init__(self, config: SambaNovaSafetyConfig) -> None:
+        self.config = config
+
+    async def initialize(self) -> None:
+        pass
+
+    async def shutdown(self) -> None:
+        pass
+
+    def _get_api_key(self) -> str:
+        config_api_key = self.config.api_key if self.config.api_key else None
+        if config_api_key:
+            return config_api_key.get_secret_value()
+        else:
+            provider_data = self.get_request_provider_data()
+            if provider_data is None or not provider_data.sambanova_api_key:
+                raise ValueError(
+                    'Pass Sambanova API Key in the header X-LlamaStack-Provider-Data as { "sambanova_api_key": <your api key> }'
+                )
+            return provider_data.sambanova_api_key
+
+    async def register_shield(self, shield: Shield) -> None:
+        list_models_url = self.config.url + "/models"
+        try:
+            response = requests.get(list_models_url)
+            response.raise_for_status()
+        except requests.exceptions.RequestException as e:
+            raise RuntimeError(f"Request to {list_models_url} failed") from e
+        available_models = [model.get("id") for model in response.json().get("data", {})]
+        if (
+            len(available_models) == 0
+            or "guard" not in shield.provider_resource_id.lower()
+            or shield.provider_resource_id.split("sambanova/")[-1] not in available_models
+        ):
+            raise ValueError(f"Shield {shield.provider_resource_id} not found in SambaNova")
+
+    async def run_shield(
+        self, shield_id: str, messages: list[Message], params: dict[str, Any] | None = None
+    ) -> RunShieldResponse:
+        shield = await self.shield_store.get_shield(shield_id)
+        if not shield:
+            raise ValueError(f"Shield {shield_id} not found")
+
+        shield_params = shield.params
+        logger.debug(f"run_shield::{shield_params}::messages={messages}")
+        content_messages = [await convert_message_to_openai_dict_new(m) for m in messages]
+        logger.debug(f"run_shield::final:messages::{json.dumps(content_messages, indent=2)}:")
+
+        response = litellm.completion(
+            model=shield.provider_resource_id, messages=content_messages, api_key=self._get_api_key()
+        )
+        shield_message = response.choices[0].message.content
+
+        if "unsafe" in shield_message.lower():
+            user_message = CANNED_RESPONSE_TEXT
+            violation_type = shield_message.split("\n")[-1]
+            metadata = {"violation_type": violation_type}
+
+            return RunShieldResponse(
+                violation=SafetyViolation(
+                    user_message=user_message,
+                    violation_level=ViolationLevel.ERROR,
+                    metadata=metadata,
+                )
+            )
+
+        return RunShieldResponse()

llama_stack/providers/remote/tool_runtime/bing_search/bing_search.py

@@ -12,19 +12,19 @@ import httpx
 from llama_stack.apis.common.content_types import URL
 from llama_stack.apis.tools import (
     ListToolDefsResponse,
-    Tool,
     ToolDef,
+    ToolGroup,
     ToolInvocationResult,
     ToolParameter,
     ToolRuntime,
 )
 from llama_stack.distribution.request_headers import NeedsRequestProviderData
-from llama_stack.providers.datatypes import ToolsProtocolPrivate
+from llama_stack.providers.datatypes import ToolGroupsProtocolPrivate
 
 from .config import BingSearchToolConfig
 
 
-class BingSearchToolRuntimeImpl(ToolsProtocolPrivate, ToolRuntime, NeedsRequestProviderData):
+class BingSearchToolRuntimeImpl(ToolGroupsProtocolPrivate, ToolRuntime, NeedsRequestProviderData):
     def __init__(self, config: BingSearchToolConfig):
         self.config = config
         self.url = "https://api.bing.microsoft.com/v7.0/search"
@@ -32,10 +32,10 @@ class BingSearchToolRuntimeImpl(ToolsProtocolPrivate, ToolRuntime, NeedsRequestP
     async def initialize(self):
         pass
 
-    async def register_tool(self, tool: Tool) -> None:
+    async def register_toolgroup(self, toolgroup: ToolGroup) -> None:
         pass
 
-    async def unregister_tool(self, tool_id: str) -> None:
+    async def unregister_toolgroup(self, toolgroup_id: str) -> None:
         return
 
     def _get_api_key(self) -> str:

llama_stack/providers/remote/tool_runtime/brave_search/brave_search.py

@@ -11,30 +11,30 @@ import httpx
 from llama_stack.apis.common.content_types import URL
 from llama_stack.apis.tools import (
     ListToolDefsResponse,
-    Tool,
     ToolDef,
+    ToolGroup,
     ToolInvocationResult,
     ToolParameter,
     ToolRuntime,
 )
 from llama_stack.distribution.request_headers import NeedsRequestProviderData
 from llama_stack.models.llama.datatypes import BuiltinTool
-from llama_stack.providers.datatypes import ToolsProtocolPrivate
+from llama_stack.providers.datatypes import ToolGroupsProtocolPrivate
 
 from .config import BraveSearchToolConfig
 
 
-class BraveSearchToolRuntimeImpl(ToolsProtocolPrivate, ToolRuntime, NeedsRequestProviderData):
+class BraveSearchToolRuntimeImpl(ToolGroupsProtocolPrivate, ToolRuntime, NeedsRequestProviderData):
     def __init__(self, config: BraveSearchToolConfig):
         self.config = config
 
     async def initialize(self):
         pass
 
-    async def register_tool(self, tool: Tool) -> None:
+    async def register_toolgroup(self, toolgroup: ToolGroup) -> None:
         pass
 
-    async def unregister_tool(self, tool_id: str) -> None:
+    async def unregister_toolgroup(self, toolgroup_id: str) -> None:
         return
 
     def _get_api_key(self) -> str:

llama_stack/providers/remote/tool_runtime/model_context_protocol/__init__.py

@@ -4,18 +4,12 @@
 # This source code is licensed under the terms described in the LICENSE file in
 # the root directory of this source tree.
 
-from pydantic import BaseModel
-
-from .config import ModelContextProtocolConfig
+from .config import MCPProviderConfig
 
 
-class ModelContextProtocolToolProviderDataValidator(BaseModel):
-    api_key: str
-
-
-async def get_adapter_impl(config: ModelContextProtocolConfig, _deps):
+async def get_adapter_impl(config: MCPProviderConfig, _deps):
     from .model_context_protocol import ModelContextProtocolToolRuntimeImpl
 
-    impl = ModelContextProtocolToolRuntimeImpl(config)
+    impl = ModelContextProtocolToolRuntimeImpl(config, _deps)
     await impl.initialize()
     return impl

llama_stack/providers/remote/tool_runtime/model_context_protocol/config.py

@@ -9,7 +9,12 @@ from typing import Any
 from pydantic import BaseModel
 
 
-class ModelContextProtocolConfig(BaseModel):
+class MCPProviderDataValidator(BaseModel):
+    # mcp_endpoint => dict of headers to send
+    mcp_headers: dict[str, dict[str, str]] | None = None
+
+
+class MCPProviderConfig(BaseModel):
     @classmethod
     def sample_run_config(cls, __distro_dir__: str, **kwargs: Any) -> dict[str, Any]:
         return {}

llama_stack/providers/remote/tool_runtime/model_context_protocol/model_context_protocol.py

@@ -7,61 +7,45 @@
 from typing import Any
 from urllib.parse import urlparse
 
-from mcp import ClientSession
-from mcp.client.sse import sse_client
-
 from llama_stack.apis.common.content_types import URL
+from llama_stack.apis.datatypes import Api
 from llama_stack.apis.tools import (
     ListToolDefsResponse,
-    ToolDef,
+    ToolGroup,
     ToolInvocationResult,
-    ToolParameter,
     ToolRuntime,
 )
-from llama_stack.providers.datatypes import ToolsProtocolPrivate
+from llama_stack.distribution.request_headers import NeedsRequestProviderData
+from llama_stack.log import get_logger
+from llama_stack.providers.datatypes import ToolGroupsProtocolPrivate
+from llama_stack.providers.utils.tools.mcp import invoke_mcp_tool, list_mcp_tools
 
-from .config import ModelContextProtocolConfig
+from .config import MCPProviderConfig
+
+logger = get_logger(__name__, category="tools")
 
 
-class ModelContextProtocolToolRuntimeImpl(ToolsProtocolPrivate, ToolRuntime):
-    def __init__(self, config: ModelContextProtocolConfig):
+class ModelContextProtocolToolRuntimeImpl(ToolGroupsProtocolPrivate, ToolRuntime, NeedsRequestProviderData):
+    def __init__(self, config: MCPProviderConfig, _deps: dict[Api, Any]):
         self.config = config
 
     async def initialize(self):
         pass
 
+    async def register_toolgroup(self, toolgroup: ToolGroup) -> None:
+        pass
+
+    async def unregister_toolgroup(self, toolgroup_id: str) -> None:
+        return
+
     async def list_runtime_tools(
         self, tool_group_id: str | None = None, mcp_endpoint: URL | None = None
     ) -> ListToolDefsResponse:
+        # this endpoint should be retrieved by getting the tool group right?
         if mcp_endpoint is None:
             raise ValueError("mcp_endpoint is required")
-
-        tools = []
-        async with sse_client(mcp_endpoint.uri) as streams:
-            async with ClientSession(*streams) as session:
-                await session.initialize()
-                tools_result = await session.list_tools()
-                for tool in tools_result.tools:
-                    parameters = []
-                    for param_name, param_schema in tool.inputSchema.get("properties", {}).items():
-                        parameters.append(
-                            ToolParameter(
-                                name=param_name,
-                                parameter_type=param_schema.get("type", "string"),
-                                description=param_schema.get("description", ""),
-                            )
-                        )
-                    tools.append(
-                        ToolDef(
-                            name=tool.name,
-                            description=tool.description,
-                            parameters=parameters,
-                            metadata={
-                                "endpoint": mcp_endpoint.uri,
-                            },
-                        )
-                    )
-        return ListToolDefsResponse(data=tools)
+        headers = await self.get_headers_from_request(mcp_endpoint.uri)
+        return await list_mcp_tools(mcp_endpoint.uri, headers)
 
     async def invoke_tool(self, tool_name: str, kwargs: dict[str, Any]) -> ToolInvocationResult:
         tool = await self.tool_store.get_tool(tool_name)
@@ -71,12 +55,19 @@ class ModelContextProtocolToolRuntimeImpl(ToolsProtocolPrivate, ToolRuntime):
         if urlparse(endpoint).scheme not in ("http", "https"):
             raise ValueError(f"Endpoint {endpoint} is not a valid HTTP(S) URL")
 
-        async with sse_client(endpoint) as streams:
-            async with ClientSession(*streams) as session:
-                await session.initialize()
-                result = await session.call_tool(tool.identifier, kwargs)
+        headers = await self.get_headers_from_request(endpoint)
+        return await invoke_mcp_tool(endpoint, headers, tool_name, kwargs)
 
-        return ToolInvocationResult(
-            content="\n".join([result.model_dump_json() for result in result.content]),
-            error_code=1 if result.isError else 0,
-        )
+    async def get_headers_from_request(self, mcp_endpoint_uri: str) -> dict[str, str]:
+        def canonicalize_uri(uri: str) -> str:
+            return f"{urlparse(uri).netloc or ''}/{urlparse(uri).path or ''}"
+
+        headers = {}
+
+        provider_data = self.get_request_provider_data()
+        if provider_data and provider_data.mcp_headers:
+            for uri, values in provider_data.mcp_headers.items():
+                if canonicalize_uri(uri) != canonicalize_uri(mcp_endpoint_uri):
+                    continue
+                headers.update(values)
+        return headers

llama_stack/providers/remote/tool_runtime/tavily_search/tavily_search.py

@@ -12,29 +12,29 @@ import httpx
 from llama_stack.apis.common.content_types import URL
 from llama_stack.apis.tools import (
     ListToolDefsResponse,
-    Tool,
     ToolDef,
+    ToolGroup,
     ToolInvocationResult,
     ToolParameter,
     ToolRuntime,
 )
 from llama_stack.distribution.request_headers import NeedsRequestProviderData
-from llama_stack.providers.datatypes import ToolsProtocolPrivate
+from llama_stack.providers.datatypes import ToolGroupsProtocolPrivate
 
 from .config import TavilySearchToolConfig
 
 
-class TavilySearchToolRuntimeImpl(ToolsProtocolPrivate, ToolRuntime, NeedsRequestProviderData):
+class TavilySearchToolRuntimeImpl(ToolGroupsProtocolPrivate, ToolRuntime, NeedsRequestProviderData):
     def __init__(self, config: TavilySearchToolConfig):
         self.config = config
 
     async def initialize(self):
         pass
 
-    async def register_tool(self, tool: Tool) -> None:
+    async def register_toolgroup(self, toolgroup: ToolGroup) -> None:
         pass
 
-    async def unregister_tool(self, tool_id: str) -> None:
+    async def unregister_toolgroup(self, toolgroup_id: str) -> None:
         return
 
     def _get_api_key(self) -> str:

llama_stack/providers/remote/tool_runtime/wolfram_alpha/wolfram_alpha.py

@@ -12,19 +12,19 @@ import httpx
 from llama_stack.apis.common.content_types import URL
 from llama_stack.apis.tools import (
     ListToolDefsResponse,
-    Tool,
     ToolDef,
+    ToolGroup,
     ToolInvocationResult,
     ToolParameter,
     ToolRuntime,
 )
 from llama_stack.distribution.request_headers import NeedsRequestProviderData
-from llama_stack.providers.datatypes import ToolsProtocolPrivate
+from llama_stack.providers.datatypes import ToolGroupsProtocolPrivate
 
 from .config import WolframAlphaToolConfig
 
 
-class WolframAlphaToolRuntimeImpl(ToolsProtocolPrivate, ToolRuntime, NeedsRequestProviderData):
+class WolframAlphaToolRuntimeImpl(ToolGroupsProtocolPrivate, ToolRuntime, NeedsRequestProviderData):
     def __init__(self, config: WolframAlphaToolConfig):
         self.config = config
         self.url = "https://api.wolframalpha.com/v2/query"
@@ -32,10 +32,10 @@ class WolframAlphaToolRuntimeImpl(ToolsProtocolPrivate, ToolRuntime, NeedsReques
     async def initialize(self):
         pass
 
-    async def register_tool(self, tool: Tool) -> None:
+    async def register_toolgroup(self, toolgroup: ToolGroup) -> None:
         pass
 
-    async def unregister_tool(self, tool_id: str) -> None:
+    async def unregister_toolgroup(self, toolgroup_id: str) -> None:
         return
 
     def _get_api_key(self) -> str:

llama_stack/providers/remote/vector_io/chroma/chroma.py

@@ -84,6 +84,14 @@ class ChromaIndex(EmbeddingIndex):
     async def delete(self):
         await maybe_await(self.client.delete_collection(self.collection.name))
 
+    async def query_keyword(
+        self,
+        query_string: str,
+        k: int,
+        score_threshold: float,
+    ) -> QueryChunksResponse:
+        raise NotImplementedError("Keyword search is not supported in Chroma")
+
 
 class ChromaVectorIOAdapter(VectorIO, VectorDBsProtocolPrivate):
     def __init__(

llama_stack/providers/remote/vector_io/milvus/milvus.py

@@ -73,7 +73,7 @@ class MilvusIndex(EmbeddingIndex):
             logger.error(f"Error inserting chunks into Milvus collection {self.collection_name}: {e}")
             raise e
 
-    async def query(self, embedding: NDArray, k: int, score_threshold: float) -> QueryChunksResponse:
+    async def query_vector(self, embedding: NDArray, k: int, score_threshold: float) -> QueryChunksResponse:
         search_res = await asyncio.to_thread(
             self.client.search,
             collection_name=self.collection_name,
@@ -86,6 +86,14 @@ class MilvusIndex(EmbeddingIndex):
         scores = [res["distance"] for res in search_res[0]]
         return QueryChunksResponse(chunks=chunks, scores=scores)
 
+    async def query_keyword(
+        self,
+        query_string: str,
+        k: int,
+        score_threshold: float,
+    ) -> QueryChunksResponse:
+        raise NotImplementedError("Keyword search is not supported in Milvus")
+
 
 class MilvusVectorIOAdapter(VectorIO, VectorDBsProtocolPrivate):
     def __init__(

llama_stack/providers/remote/vector_io/pgvector/pgvector.py

@@ -99,7 +99,7 @@ class PGVectorIndex(EmbeddingIndex):
         with self.conn.cursor(cursor_factory=psycopg2.extras.DictCursor) as cur:
             execute_values(cur, query, values, template="(%s, %s, %s::vector)")
 
-    async def query(self, embedding: NDArray, k: int, score_threshold: float) -> QueryChunksResponse:
+    async def query_vector(self, embedding: NDArray, k: int, score_threshold: float) -> QueryChunksResponse:
         with self.conn.cursor(cursor_factory=psycopg2.extras.DictCursor) as cur:
             cur.execute(
                 f"""
@@ -120,6 +120,14 @@ class PGVectorIndex(EmbeddingIndex):
 
             return QueryChunksResponse(chunks=chunks, scores=scores)
 
+    async def query_keyword(
+        self,
+        query_string: str,
+        k: int,
+        score_threshold: float,
+    ) -> QueryChunksResponse:
+        raise NotImplementedError("Keyword search is not supported in PGVector")
+
     async def delete(self):
         with self.conn.cursor(cursor_factory=psycopg2.extras.DictCursor) as cur:
             cur.execute(f"DROP TABLE IF EXISTS {self.table_name}")

llama_stack/providers/remote/vector_io/qdrant/qdrant.py

@@ -68,7 +68,7 @@ class QdrantIndex(EmbeddingIndex):
 
         await self.client.upsert(collection_name=self.collection_name, points=points)
 
-    async def query(self, embedding: NDArray, k: int, score_threshold: float) -> QueryChunksResponse:
+    async def query_vector(self, embedding: NDArray, k: int, score_threshold: float) -> QueryChunksResponse:
         results = (
             await self.client.query_points(
                 collection_name=self.collection_name,
@@ -95,6 +95,14 @@ class QdrantIndex(EmbeddingIndex):
 
         return QueryChunksResponse(chunks=chunks, scores=scores)
 
+    async def query_keyword(
+        self,
+        query_string: str,
+        k: int,
+        score_threshold: float,
+    ) -> QueryChunksResponse:
+        raise NotImplementedError("Keyword search is not supported in Qdrant")
+
     async def delete(self):
         await self.client.delete_collection(collection_name=self.collection_name)
 

llama_stack/providers/remote/vector_io/weaviate/weaviate.py

@@ -55,7 +55,7 @@ class WeaviateIndex(EmbeddingIndex):
         # TODO: make this async friendly
         collection.data.insert_many(data_objects)
 
-    async def query(self, embedding: NDArray, k: int, score_threshold: float) -> QueryChunksResponse:
+    async def query_vector(self, embedding: NDArray, k: int, score_threshold: float) -> QueryChunksResponse:
         collection = self.client.collections.get(self.collection_name)
 
         results = collection.query.near_vector(
@@ -84,6 +84,14 @@ class WeaviateIndex(EmbeddingIndex):
         collection = self.client.collections.get(self.collection_name)
         collection.data.delete_many(where=Filter.by_property("id").contains_any(chunk_ids))
 
+    async def query_keyword(
+        self,
+        query_string: str,
+        k: int,
+        score_threshold: float,
+    ) -> QueryChunksResponse:
+        raise NotImplementedError("Keyword search is not supported in Weaviate")
+
 
 class WeaviateVectorIOAdapter(
     VectorIO,

llama_stack/providers/utils/inference/inference_store.py

@@ -0,0 +1,123 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the terms described in the LICENSE file in
+# the root directory of this source tree.
+from llama_stack.apis.inference import (
+    ListOpenAIChatCompletionResponse,
+    OpenAIChatCompletion,
+    OpenAICompletionWithInputMessages,
+    OpenAIMessageParam,
+    Order,
+)
+from llama_stack.distribution.utils.config_dirs import RUNTIME_BASE_DIR
+
+from ..sqlstore.api import ColumnDefinition, ColumnType
+from ..sqlstore.sqlstore import SqliteSqlStoreConfig, SqlStoreConfig, sqlstore_impl
+
+
+class InferenceStore:
+    def __init__(self, sql_store_config: SqlStoreConfig):
+        if not sql_store_config:
+            sql_store_config = SqliteSqlStoreConfig(
+                db_path=(RUNTIME_BASE_DIR / "sqlstore.db").as_posix(),
+            )
+        self.sql_store_config = sql_store_config
+        self.sql_store = None
+
+    async def initialize(self):
+        """Create the necessary tables if they don't exist."""
+        self.sql_store = sqlstore_impl(self.sql_store_config)
+        await self.sql_store.create_table(
+            "chat_completions",
+            {
+                "id": ColumnDefinition(type=ColumnType.STRING, primary_key=True),
+                "created": ColumnType.INTEGER,
+                "model": ColumnType.STRING,
+                "choices": ColumnType.JSON,
+                "input_messages": ColumnType.JSON,
+            },
+        )
+
+    async def store_chat_completion(
+        self, chat_completion: OpenAIChatCompletion, input_messages: list[OpenAIMessageParam]
+    ) -> None:
+        if not self.sql_store:
+            raise ValueError("Inference store is not initialized")
+
+        data = chat_completion.model_dump()
+
+        await self.sql_store.insert(
+            "chat_completions",
+            {
+                "id": data["id"],
+                "created": data["created"],
+                "model": data["model"],
+                "choices": data["choices"],
+                "input_messages": [message.model_dump() for message in input_messages],
+            },
+        )
+
+    async def list_chat_completions(
+        self,
+        after: str | None = None,
+        limit: int | None = 50,
+        model: str | None = None,
+        order: Order | None = Order.desc,
+    ) -> ListOpenAIChatCompletionResponse:
+        """
+        List chat completions from the database.
+
+        :param after: The ID of the last chat completion to return.
+        :param limit: The maximum number of chat completions to return.
+        :param model: The model to filter by.
+        :param order: The order to sort the chat completions by.
+        """
+        if not self.sql_store:
+            raise ValueError("Inference store is not initialized")
+
+        # TODO: support after
+        if after:
+            raise NotImplementedError("After is not supported for SQLite")
+        if not order:
+            order = Order.desc
+
+        rows = await self.sql_store.fetch_all(
+            "chat_completions",
+            where={"model": model} if model else None,
+            order_by=[("created", order.value)],
+            limit=limit,
+        )
+
+        data = [
+            OpenAICompletionWithInputMessages(
+                id=row["id"],
+                created=row["created"],
+                model=row["model"],
+                choices=row["choices"],
+                input_messages=row["input_messages"],
+            )
+            for row in rows
+        ]
+        return ListOpenAIChatCompletionResponse(
+            data=data,
+            # TODO: implement has_more
+            has_more=False,
+            first_id=data[0].id if data else "",
+            last_id=data[-1].id if data else "",
+        )
+
+    async def get_chat_completion(self, completion_id: str) -> OpenAICompletionWithInputMessages:
+        if not self.sql_store:
+            raise ValueError("Inference store is not initialized")
+
+        row = await self.sql_store.fetch_one("chat_completions", where={"id": completion_id})
+        if not row:
+            raise ValueError(f"Chat completion with id {completion_id} not found") from None
+        return OpenAICompletionWithInputMessages(
+            id=row["id"],
+            created=row["created"],
+            model=row["model"],
+            choices=row["choices"],
+            input_messages=row["input_messages"],
+        )

llama_stack/providers/utils/inference/openai_compat.py

@@ -1402,9 +1402,8 @@ class OpenAIChatCompletionToLlamaStackMixin:
         outstanding_responses: list[Awaitable[AsyncIterator[ChatCompletionResponseStreamChunk]]],
     ):
         id = f"chatcmpl-{uuid.uuid4()}"
-        for outstanding_response in outstanding_responses:
+        for i, outstanding_response in enumerate(outstanding_responses):
             response = await outstanding_response
-            i = 0
             async for chunk in response:
                 event = chunk.event
                 finish_reason = _convert_stop_reason_to_openai_finish_reason(event.stop_reason)
@@ -1459,7 +1458,6 @@ class OpenAIChatCompletionToLlamaStackMixin:
                             model=model,
                             object="chat.completion.chunk",
                         )
-                i = i + 1
 
     async def _process_non_stream_response(
         self, model: str, outstanding_responses: list[Awaitable[ChatCompletionResponse]]

llama_stack/providers/utils/inference/stream_utils.py

@@ -0,0 +1,129 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the terms described in the LICENSE file in
+# the root directory of this source tree.
+
+from collections.abc import AsyncIterator
+from datetime import datetime, timezone
+from typing import Any
+
+from llama_stack.apis.inference import (
+    OpenAIAssistantMessageParam,
+    OpenAIChatCompletion,
+    OpenAIChatCompletionChunk,
+    OpenAIChatCompletionToolCall,
+    OpenAIChatCompletionToolCallFunction,
+    OpenAIChoice,
+    OpenAIChoiceLogprobs,
+    OpenAIMessageParam,
+)
+from llama_stack.providers.utils.inference.inference_store import InferenceStore
+
+
+async def stream_and_store_openai_completion(
+    provider_stream: AsyncIterator[OpenAIChatCompletionChunk],
+    model: str,
+    store: InferenceStore,
+    input_messages: list[OpenAIMessageParam],
+) -> AsyncIterator[OpenAIChatCompletionChunk]:
+    """
+    Wraps a provider's stream, yields chunks, and stores the full completion at the end.
+    """
+    id = None
+    created = None
+    choices_data: dict[int, dict[str, Any]] = {}
+
+    try:
+        async for chunk in provider_stream:
+            if id is None and chunk.id:
+                id = chunk.id
+            if created is None and chunk.created:
+                created = chunk.created
+
+            if chunk.choices:
+                for choice_delta in chunk.choices:
+                    idx = choice_delta.index
+                    if idx not in choices_data:
+                        choices_data[idx] = {
+                            "content_parts": [],
+                            "tool_calls_builder": {},
+                            "finish_reason": None,
+                            "logprobs_content_parts": [],
+                        }
+                    current_choice_data = choices_data[idx]
+
+                    if choice_delta.delta:
+                        delta = choice_delta.delta
+                        if delta.content:
+                            current_choice_data["content_parts"].append(delta.content)
+                        if delta.tool_calls:
+                            for tool_call_delta in delta.tool_calls:
+                                tc_idx = tool_call_delta.index
+                                if tc_idx not in current_choice_data["tool_calls_builder"]:
+                                    # Initialize with correct structure for _ToolCallBuilderData
+                                    current_choice_data["tool_calls_builder"][tc_idx] = {
+                                        "id": None,
+                                        "type": "function",
+                                        "function_name_parts": [],
+                                        "function_arguments_parts": [],
+                                    }
+                                builder = current_choice_data["tool_calls_builder"][tc_idx]
+                                if tool_call_delta.id:
+                                    builder["id"] = tool_call_delta.id
+                                if tool_call_delta.type:
+                                    builder["type"] = tool_call_delta.type
+                                if tool_call_delta.function:
+                                    if tool_call_delta.function.name:
+                                        builder["function_name_parts"].append(tool_call_delta.function.name)
+                                    if tool_call_delta.function.arguments:
+                                        builder["function_arguments_parts"].append(tool_call_delta.function.arguments)
+                    if choice_delta.finish_reason:
+                        current_choice_data["finish_reason"] = choice_delta.finish_reason
+                    if choice_delta.logprobs and choice_delta.logprobs.content:
+                        # Ensure that we are extending with the correct type
+                        current_choice_data["logprobs_content_parts"].extend(choice_delta.logprobs.content)
+            yield chunk
+    finally:
+        if id:
+            assembled_choices: list[OpenAIChoice] = []
+            for choice_idx, choice_data in choices_data.items():
+                content_str = "".join(choice_data["content_parts"])
+                assembled_tool_calls: list[OpenAIChatCompletionToolCall] = []
+                if choice_data["tool_calls_builder"]:
+                    for tc_build_data in choice_data["tool_calls_builder"].values():
+                        if tc_build_data["id"]:
+                            func_name = "".join(tc_build_data["function_name_parts"])
+                            func_args = "".join(tc_build_data["function_arguments_parts"])
+                            assembled_tool_calls.append(
+                                OpenAIChatCompletionToolCall(
+                                    id=tc_build_data["id"],
+                                    type=tc_build_data["type"],  # No or "function" needed, already set
+                                    function=OpenAIChatCompletionToolCallFunction(name=func_name, arguments=func_args),
+                                )
+                            )
+                message = OpenAIAssistantMessageParam(
+                    role="assistant",
+                    content=content_str if content_str else None,
+                    tool_calls=assembled_tool_calls if assembled_tool_calls else None,
+                )
+                logprobs_content = choice_data["logprobs_content_parts"]
+                final_logprobs = OpenAIChoiceLogprobs(content=logprobs_content) if logprobs_content else None
+
+                assembled_choices.append(
+                    OpenAIChoice(
+                        finish_reason=choice_data["finish_reason"],
+                        index=choice_idx,
+                        message=message,
+                        logprobs=final_logprobs,
+                    )
+                )
+
+            final_response = OpenAIChatCompletion(
+                id=id,
+                choices=assembled_choices,
+                created=created or int(datetime.now(timezone.utc).timestamp()),
+                model=model,
+                object="chat.completion",
+            )
+            await store.store_chat_completion(final_response, input_messages)

llama_stack/providers/utils/memory/vector_store.py

@@ -177,7 +177,11 @@ class EmbeddingIndex(ABC):
         raise NotImplementedError()
 
     @abstractmethod
-    async def query(self, embedding: NDArray, k: int, score_threshold: float) -> QueryChunksResponse:
+    async def query_vector(self, embedding: NDArray, k: int, score_threshold: float) -> QueryChunksResponse:
+        raise NotImplementedError()
+
+    @abstractmethod
+    async def query_keyword(self, query_string: str, k: int, score_threshold: float) -> QueryChunksResponse:
         raise NotImplementedError()
 
     @abstractmethod
@@ -210,9 +214,12 @@ class VectorDBWithIndex:
         if params is None:
             params = {}
         k = params.get("max_chunks", 3)
+        mode = params.get("mode")
         score_threshold = params.get("score_threshold", 0.0)
-
-        query_str = interleaved_content_as_str(query)
-        embeddings_response = await self.inference_api.embeddings(self.vector_db.embedding_model, [query_str])
-        query_vector = np.array(embeddings_response.embeddings[0], dtype=np.float32)
-        return await self.index.query(query_vector, k, score_threshold)
+        query_string = interleaved_content_as_str(query)
+        if mode == "keyword":
+            return await self.index.query_keyword(query_string, k, score_threshold)
+        else:
+            embeddings_response = await self.inference_api.embeddings(self.vector_db.embedding_model, [query_string])
+            query_vector = np.array(embeddings_response.embeddings[0], dtype=np.float32)
+            return await self.index.query_vector(query_vector, k, score_threshold)

llama_stack/providers/utils/responses/responses_store.py

@@ -0,0 +1,135 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the terms described in the LICENSE file in
+# the root directory of this source tree.
+from llama_stack.apis.agents import (
+    Order,
+)
+from llama_stack.apis.agents.openai_responses import (
+    ListOpenAIResponseInputItem,
+    ListOpenAIResponseObject,
+    OpenAIResponseInput,
+    OpenAIResponseObject,
+    OpenAIResponseObjectWithInput,
+)
+from llama_stack.distribution.utils.config_dirs import RUNTIME_BASE_DIR
+
+from ..sqlstore.api import ColumnDefinition, ColumnType
+from ..sqlstore.sqlstore import SqliteSqlStoreConfig, SqlStoreConfig, sqlstore_impl
+
+
+class ResponsesStore:
+    def __init__(self, sql_store_config: SqlStoreConfig):
+        if not sql_store_config:
+            sql_store_config = SqliteSqlStoreConfig(
+                db_path=(RUNTIME_BASE_DIR / "sqlstore.db").as_posix(),
+            )
+        self.sql_store = sqlstore_impl(sql_store_config)
+
+    async def initialize(self):
+        """Create the necessary tables if they don't exist."""
+        await self.sql_store.create_table(
+            "openai_responses",
+            {
+                "id": ColumnDefinition(type=ColumnType.STRING, primary_key=True),
+                "created_at": ColumnType.INTEGER,
+                "response_object": ColumnType.JSON,
+                "model": ColumnType.STRING,
+            },
+        )
+
+    async def store_response_object(
+        self, response_object: OpenAIResponseObject, input: list[OpenAIResponseInput]
+    ) -> None:
+        data = response_object.model_dump()
+        data["input"] = [input_item.model_dump() for input_item in input]
+
+        await self.sql_store.insert(
+            "openai_responses",
+            {
+                "id": data["id"],
+                "created_at": data["created_at"],
+                "model": data["model"],
+                "response_object": data,
+            },
+        )
+
+    async def list_responses(
+        self,
+        after: str | None = None,
+        limit: int | None = 50,
+        model: str | None = None,
+        order: Order | None = Order.desc,
+    ) -> ListOpenAIResponseObject:
+        """
+        List responses from the database.
+
+        :param after: The ID of the last response to return.
+        :param limit: The maximum number of responses to return.
+        :param model: The model to filter by.
+        :param order: The order to sort the responses by.
+        """
+        # TODO: support after
+        if after:
+            raise NotImplementedError("After is not supported for SQLite")
+        if not order:
+            order = Order.desc
+
+        rows = await self.sql_store.fetch_all(
+            "openai_responses",
+            where={"model": model} if model else None,
+            order_by=[("created_at", order.value)],
+            limit=limit,
+        )
+
+        data = [OpenAIResponseObjectWithInput(**row["response_object"]) for row in rows]
+        return ListOpenAIResponseObject(
+            data=data,
+            # TODO: implement has_more
+            has_more=False,
+            first_id=data[0].id if data else "",
+            last_id=data[-1].id if data else "",
+        )
+
+    async def get_response_object(self, response_id: str) -> OpenAIResponseObjectWithInput:
+        row = await self.sql_store.fetch_one("openai_responses", where={"id": response_id})
+        if not row:
+            raise ValueError(f"Response with id {response_id} not found") from None
+        return OpenAIResponseObjectWithInput(**row["response_object"])
+
+    async def list_response_input_items(
+        self,
+        response_id: str,
+        after: str | None = None,
+        before: str | None = None,
+        include: list[str] | None = None,
+        limit: int | None = 20,
+        order: Order | None = Order.desc,
+    ) -> ListOpenAIResponseInputItem:
+        """
+        List input items for a given response.
+
+        :param response_id: The ID of the response to retrieve input items for.
+        :param after: An item ID to list items after, used for pagination.
+        :param before: An item ID to list items before, used for pagination.
+        :param include: Additional fields to include in the response.
+        :param limit: A limit on the number of objects to be returned.
+        :param order: The order to return the input items in.
+        """
+        # TODO: support after/before pagination
+        if after or before:
+            raise NotImplementedError("After/before pagination is not supported yet")
+        if include:
+            raise NotImplementedError("Include is not supported yet")
+
+        response_with_input = await self.get_response_object(response_id)
+        input_items = response_with_input.input
+
+        if order == Order.desc:
+            input_items = list(reversed(input_items))
+
+        if limit is not None and len(input_items) > limit:
+            input_items = input_items[:limit]
+
+        return ListOpenAIResponseInputItem(data=input_items)

llama_stack/providers/utils/sqlstore/api.py

@@ -0,0 +1,90 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the terms described in the LICENSE file in
+# the root directory of this source tree.
+
+from collections.abc import Mapping
+from enum import Enum
+from typing import Any, Literal, Protocol
+
+from pydantic import BaseModel
+
+
+class ColumnType(Enum):
+    INTEGER = "INTEGER"
+    STRING = "STRING"
+    TEXT = "TEXT"
+    FLOAT = "FLOAT"
+    BOOLEAN = "BOOLEAN"
+    JSON = "JSON"
+    DATETIME = "DATETIME"
+
+
+class ColumnDefinition(BaseModel):
+    type: ColumnType
+    primary_key: bool = False
+    nullable: bool = True
+    default: Any = None
+
+
+class SqlStore(Protocol):
+    """
+    A protocol for a SQL store.
+    """
+
+    async def create_table(self, table: str, schema: Mapping[str, ColumnType | ColumnDefinition]) -> None:
+        """
+        Create a table.
+        """
+        pass
+
+    async def insert(self, table: str, data: Mapping[str, Any]) -> None:
+        """
+        Insert a row into a table.
+        """
+        pass
+
+    async def fetch_all(
+        self,
+        table: str,
+        where: Mapping[str, Any] | None = None,
+        limit: int | None = None,
+        order_by: list[tuple[str, Literal["asc", "desc"]]] | None = None,
+    ) -> list[dict[str, Any]]:
+        """
+        Fetch all rows from a table.
+        """
+        pass
+
+    async def fetch_one(
+        self,
+        table: str,
+        where: Mapping[str, Any] | None = None,
+        order_by: list[tuple[str, Literal["asc", "desc"]]] | None = None,
+    ) -> dict[str, Any] | None:
+        """
+        Fetch one row from a table.
+        """
+        pass
+
+    async def update(
+        self,
+        table: str,
+        data: Mapping[str, Any],
+        where: Mapping[str, Any],
+    ) -> None:
+        """
+        Update a row in a table.
+        """
+        pass
+
+    async def delete(
+        self,
+        table: str,
+        where: Mapping[str, Any],
+    ) -> None:
+        """
+        Delete a row from a table.
+        """
+        pass

llama_stack/providers/utils/sqlstore/sqlite/sqlite.py

@@ -0,0 +1,161 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the terms described in the LICENSE file in
+# the root directory of this source tree.
+from collections.abc import Mapping
+from typing import Any, Literal
+
+from sqlalchemy import (
+    JSON,
+    Boolean,
+    Column,
+    DateTime,
+    Float,
+    Integer,
+    MetaData,
+    String,
+    Table,
+    Text,
+    select,
+)
+from sqlalchemy.ext.asyncio import create_async_engine
+
+from ..api import ColumnDefinition, ColumnType, SqlStore
+from ..sqlstore import SqliteSqlStoreConfig
+
+TYPE_MAPPING: dict[ColumnType, Any] = {
+    ColumnType.INTEGER: Integer,
+    ColumnType.STRING: String,
+    ColumnType.FLOAT: Float,
+    ColumnType.BOOLEAN: Boolean,
+    ColumnType.DATETIME: DateTime,
+    ColumnType.TEXT: Text,
+    ColumnType.JSON: JSON,
+}
+
+
+class SqliteSqlStoreImpl(SqlStore):
+    def __init__(self, config: SqliteSqlStoreConfig):
+        self.engine = create_async_engine(config.engine_str)
+        self.metadata = MetaData()
+
+    async def create_table(
+        self,
+        table: str,
+        schema: Mapping[str, ColumnType | ColumnDefinition],
+    ) -> None:
+        if not schema:
+            raise ValueError(f"No columns defined for table '{table}'.")
+
+        sqlalchemy_columns: list[Column] = []
+
+        for col_name, col_props in schema.items():
+            col_type = None
+            is_primary_key = False
+            is_nullable = True  # Default to nullable
+
+            if isinstance(col_props, ColumnType):
+                col_type = col_props
+            elif isinstance(col_props, ColumnDefinition):
+                col_type = col_props.type
+                is_primary_key = col_props.primary_key
+                is_nullable = col_props.nullable
+
+            sqlalchemy_type = TYPE_MAPPING.get(col_type)
+            if not sqlalchemy_type:
+                raise ValueError(f"Unsupported column type '{col_type}' for column '{col_name}'.")
+
+            sqlalchemy_columns.append(
+                Column(col_name, sqlalchemy_type, primary_key=is_primary_key, nullable=is_nullable)
+            )
+
+        # Check if table already exists in metadata, otherwise define it
+        if table not in self.metadata.tables:
+            sqlalchemy_table = Table(table, self.metadata, *sqlalchemy_columns)
+        else:
+            sqlalchemy_table = self.metadata.tables[table]
+
+        # Create the table in the database if it doesn't exist
+        # checkfirst=True ensures it doesn't try to recreate if it's already there
+        async with self.engine.begin() as conn:
+            await conn.run_sync(self.metadata.create_all, tables=[sqlalchemy_table], checkfirst=True)
+
+    async def insert(self, table: str, data: Mapping[str, Any]) -> None:
+        async with self.engine.begin() as conn:
+            await conn.execute(self.metadata.tables[table].insert(), data)
+            await conn.commit()
+
+    async def fetch_all(
+        self,
+        table: str,
+        where: Mapping[str, Any] | None = None,
+        limit: int | None = None,
+        order_by: list[tuple[str, Literal["asc", "desc"]]] | None = None,
+    ) -> list[dict[str, Any]]:
+        async with self.engine.begin() as conn:
+            query = select(self.metadata.tables[table])
+            if where:
+                for key, value in where.items():
+                    query = query.where(self.metadata.tables[table].c[key] == value)
+            if limit:
+                query = query.limit(limit)
+            if order_by:
+                if not isinstance(order_by, list):
+                    raise ValueError(
+                        f"order_by must be a list of tuples (column, order={['asc', 'desc']}), got {order_by}"
+                    )
+                for order in order_by:
+                    if not isinstance(order, tuple):
+                        raise ValueError(
+                            f"order_by must be a list of tuples (column, order={['asc', 'desc']}), got {order_by}"
+                        )
+                    name, order_type = order
+                    if order_type == "asc":
+                        query = query.order_by(self.metadata.tables[table].c[name].asc())
+                    elif order_type == "desc":
+                        query = query.order_by(self.metadata.tables[table].c[name].desc())
+                    else:
+                        raise ValueError(f"Invalid order '{order_type}' for column '{name}'")
+            result = await conn.execute(query)
+            if result.rowcount == 0:
+                return []
+            return [dict(row._mapping) for row in result]
+
+    async def fetch_one(
+        self,
+        table: str,
+        where: Mapping[str, Any] | None = None,
+        order_by: list[tuple[str, Literal["asc", "desc"]]] | None = None,
+    ) -> dict[str, Any] | None:
+        rows = await self.fetch_all(table, where, limit=1, order_by=order_by)
+        if not rows:
+            return None
+        return rows[0]
+
+    async def update(
+        self,
+        table: str,
+        data: Mapping[str, Any],
+        where: Mapping[str, Any],
+    ) -> None:
+        if not where:
+            raise ValueError("where is required for update")
+
+        async with self.engine.begin() as conn:
+            stmt = self.metadata.tables[table].update()
+            for key, value in where.items():
+                stmt = stmt.where(self.metadata.tables[table].c[key] == value)
+            await conn.execute(stmt, data)
+            await conn.commit()
+
+    async def delete(self, table: str, where: Mapping[str, Any]) -> None:
+        if not where:
+            raise ValueError("where is required for delete")
+
+        async with self.engine.begin() as conn:
+            stmt = self.metadata.tables[table].delete()
+            for key, value in where.items():
+                stmt = stmt.where(self.metadata.tables[table].c[key] == value)
+            await conn.execute(stmt)
+            await conn.commit()

llama_stack/providers/utils/sqlstore/sqlstore.py

@@ -0,0 +1,72 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the terms described in the LICENSE file in
+# the root directory of this source tree.
+
+
+from enum import Enum
+from pathlib import Path
+from typing import Annotated, Literal
+
+from pydantic import BaseModel, Field
+
+from llama_stack.distribution.utils.config_dirs import RUNTIME_BASE_DIR
+
+from .api import SqlStore
+
+
+class SqlStoreType(Enum):
+    sqlite = "sqlite"
+    postgres = "postgres"
+
+
+class SqliteSqlStoreConfig(BaseModel):
+    type: Literal["sqlite"] = SqlStoreType.sqlite.value
+    db_path: str = Field(
+        default=(RUNTIME_BASE_DIR / "sqlstore.db").as_posix(),
+        description="Database path, e.g. ~/.llama/distributions/ollama/sqlstore.db",
+    )
+
+    @property
+    def engine_str(self) -> str:
+        return "sqlite+aiosqlite:///" + Path(self.db_path).expanduser().as_posix()
+
+    @classmethod
+    def sample_run_config(cls, __distro_dir__: str, db_name: str = "sqlstore.db"):
+        return cls(
+            type="sqlite",
+            db_path="${env.SQLITE_STORE_DIR:" + __distro_dir__ + "}/" + db_name,
+        )
+
+    # TODO: move this when we have a better way to specify dependencies with internal APIs
+    @property
+    def pip_packages(self) -> list[str]:
+        return ["sqlalchemy[asyncio]"]
+
+
+class PostgresSqlStoreConfig(BaseModel):
+    type: Literal["postgres"] = SqlStoreType.postgres.value
+
+    @property
+    def pip_packages(self) -> list[str]:
+        raise NotImplementedError("Postgres is not implemented yet")
+
+
+SqlStoreConfig = Annotated[
+    SqliteSqlStoreConfig | PostgresSqlStoreConfig,
+    Field(discriminator="type", default=SqlStoreType.sqlite.value),
+]
+
+
+def sqlstore_impl(config: SqlStoreConfig) -> SqlStore:
+    if config.type == SqlStoreType.sqlite.value:
+        from .sqlite.sqlite import SqliteSqlStoreImpl
+
+        impl = SqliteSqlStoreImpl(config)
+    elif config.type == SqlStoreType.postgres.value:
+        raise NotImplementedError("Postgres is not implemented yet")
+    else:
+        raise ValueError(f"Unknown sqlstore type {config.type}")
+
+    return impl

llama_stack/providers/utils/telemetry/tracing.py

@@ -34,6 +34,8 @@ logger = get_logger(__name__, category="core")
 INVALID_SPAN_ID = 0x0000000000000000
 INVALID_TRACE_ID = 0x00000000000000000000000000000000
 
+ROOT_SPAN_MARKERS = ["__root__", "__root_span__"]
+
 
 def trace_id_to_str(trace_id: int) -> str:
     """Convenience trace ID formatting method
@@ -178,7 +180,8 @@ async def start_trace(name: str, attributes: dict[str, Any] = None) -> TraceCont
 
     trace_id = generate_trace_id()
     context = TraceContext(BACKGROUND_LOGGER, trace_id)
-    context.push_span(name, {"__root__": True, **(attributes or {})})
+    attributes = {marker: True for marker in ROOT_SPAN_MARKERS} | (attributes or {})
+    context.push_span(name, attributes)
 
     CURRENT_TRACE_CONTEXT.set(context)
     return context

llama_stack/providers/utils/tools/mcp.py

@@ -0,0 +1,100 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the terms described in the LICENSE file in
+# the root directory of this source tree.
+
+from contextlib import asynccontextmanager
+from typing import Any
+
+try:
+    # for python < 3.11
+    import exceptiongroup
+
+    BaseExceptionGroup = exceptiongroup.BaseExceptionGroup
+except ImportError:
+    pass
+
+import httpx
+from mcp import ClientSession
+from mcp import types as mcp_types
+from mcp.client.sse import sse_client
+
+from llama_stack.apis.common.content_types import ImageContentItem, InterleavedContentItem, TextContentItem
+from llama_stack.apis.tools import (
+    ListToolDefsResponse,
+    ToolDef,
+    ToolInvocationResult,
+    ToolParameter,
+)
+from llama_stack.distribution.datatypes import AuthenticationRequiredError
+from llama_stack.log import get_logger
+
+logger = get_logger(__name__, category="tools")
+
+
+@asynccontextmanager
+async def sse_client_wrapper(endpoint: str, headers: dict[str, str]):
+    try:
+        async with sse_client(endpoint, headers=headers) as streams:
+            async with ClientSession(*streams) as session:
+                await session.initialize()
+                yield session
+    except BaseException as e:
+        if isinstance(e, BaseExceptionGroup):
+            for exc in e.exceptions:
+                if isinstance(exc, httpx.HTTPStatusError) and exc.response.status_code == 401:
+                    raise AuthenticationRequiredError(exc) from exc
+        elif isinstance(e, httpx.HTTPStatusError) and e.response.status_code == 401:
+            raise AuthenticationRequiredError(e) from e
+
+        raise
+
+
+async def list_mcp_tools(endpoint: str, headers: dict[str, str]) -> ListToolDefsResponse:
+    tools = []
+    async with sse_client_wrapper(endpoint, headers) as session:
+        tools_result = await session.list_tools()
+        for tool in tools_result.tools:
+            parameters = []
+            for param_name, param_schema in tool.inputSchema.get("properties", {}).items():
+                parameters.append(
+                    ToolParameter(
+                        name=param_name,
+                        parameter_type=param_schema.get("type", "string"),
+                        description=param_schema.get("description", ""),
+                    )
+                )
+            tools.append(
+                ToolDef(
+                    name=tool.name,
+                    description=tool.description,
+                    parameters=parameters,
+                    metadata={
+                        "endpoint": endpoint,
+                    },
+                )
+            )
+    return ListToolDefsResponse(data=tools)
+
+
+async def invoke_mcp_tool(
+    endpoint: str, headers: dict[str, str], tool_name: str, kwargs: dict[str, Any]
+) -> ToolInvocationResult:
+    async with sse_client_wrapper(endpoint, headers) as session:
+        result = await session.call_tool(tool_name, kwargs)
+
+        content: list[InterleavedContentItem] = []
+        for item in result.content:
+            if isinstance(item, mcp_types.TextContent):
+                content.append(TextContentItem(text=item.text))
+            elif isinstance(item, mcp_types.ImageContent):
+                content.append(ImageContentItem(image=item.data))
+            elif isinstance(item, mcp_types.EmbeddedResource):
+                logger.warning(f"EmbeddedResource is not supported: {item}")
+            else:
+                raise ValueError(f"Unknown content type: {type(item)}")
+        return ToolInvocationResult(
+            content=content,
+            error_code=1 if result.isError else 0,
+        )

llama_stack/templates/bedrock/build.yaml

@@ -29,3 +29,5 @@ distribution_spec:
     - inline::rag-runtime
     - remote::model-context-protocol
 image_type: conda
+additional_pip_packages:
+- sqlalchemy[asyncio]

llama_stack/templates/bedrock/run.yaml

@@ -35,6 +35,9 @@ providers:
         type: sqlite
         namespace: null
         db_path: ${env.SQLITE_STORE_DIR:~/.llama/distributions/bedrock}/agents_store.db
+      responses_store:
+        type: sqlite
+        db_path: ${env.SQLITE_STORE_DIR:~/.llama/distributions/bedrock}/responses_store.db
   telemetry:
   - provider_id: meta-reference
     provider_type: inline::meta-reference
@@ -96,6 +99,9 @@ providers:
 metadata_store:
   type: sqlite
   db_path: ${env.SQLITE_STORE_DIR:~/.llama/distributions/bedrock}/registry.db
+inference_store:
+  type: sqlite
+  db_path: ${env.SQLITE_STORE_DIR:~/.llama/distributions/bedrock}/inference_store.db
 models:
 - metadata: {}
   model_id: meta.llama3-1-8b-instruct-v1:0

llama_stack/templates/cerebras/build.yaml

@@ -29,3 +29,5 @@ distribution_spec:
     - remote::tavily-search
     - inline::rag-runtime
 image_type: conda
+additional_pip_packages:
+- sqlalchemy[asyncio]

llama_stack/templates/cerebras/run.yaml

@@ -41,6 +41,9 @@ providers:
         type: sqlite
         namespace: null
         db_path: ${env.SQLITE_STORE_DIR:~/.llama/distributions/cerebras}/agents_store.db
+      responses_store:
+        type: sqlite
+        db_path: ${env.SQLITE_STORE_DIR:~/.llama/distributions/cerebras}/responses_store.db
   eval:
   - provider_id: meta-reference
     provider_type: inline::meta-reference
@@ -99,6 +102,9 @@ providers:
 metadata_store:
   type: sqlite
   db_path: ${env.SQLITE_STORE_DIR:~/.llama/distributions/cerebras}/registry.db
+inference_store:
+  type: sqlite
+  db_path: ${env.SQLITE_STORE_DIR:~/.llama/distributions/cerebras}/inference_store.db
 models:
 - metadata: {}
   model_id: llama3.1-8b

llama_stack/templates/ci-tests/build.yaml

@@ -30,3 +30,5 @@ distribution_spec:
     - inline::rag-runtime
     - remote::model-context-protocol
 image_type: conda
+additional_pip_packages:
+- sqlalchemy[asyncio]

llama_stack/templates/ci-tests/run.yaml

@@ -38,6 +38,9 @@ providers:
         type: sqlite
         namespace: null
         db_path: ${env.SQLITE_STORE_DIR:~/.llama/distributions/ci-tests}/agents_store.db
+      responses_store:
+        type: sqlite
+        db_path: ${env.SQLITE_STORE_DIR:~/.llama/distributions/ci-tests}/responses_store.db
   telemetry:
   - provider_id: meta-reference
     provider_type: inline::meta-reference
@@ -99,6 +102,9 @@ providers:
 metadata_store:
   type: sqlite
   db_path: ${env.SQLITE_STORE_DIR:~/.llama/distributions/ci-tests}/registry.db
+inference_store:
+  type: sqlite
+  db_path: ${env.SQLITE_STORE_DIR:~/.llama/distributions/ci-tests}/inference_store.db
 models:
 - metadata: {}
   model_id: accounts/fireworks/models/llama-v3p1-8b-instruct

llama_stack/templates/dell/build.yaml

@@ -30,3 +30,6 @@ distribution_spec:
     - remote::tavily-search
     - inline::rag-runtime
 image_type: conda
+additional_pip_packages:
+- sqlalchemy[asyncio]
+- sqlalchemy[asyncio]

llama_stack/templates/dell/run-with-safety.yaml

@@ -41,6 +41,9 @@ providers:
         type: sqlite
         namespace: null
         db_path: ${env.SQLITE_STORE_DIR:~/.llama/distributions/dell}/agents_store.db
+      responses_store:
+        type: sqlite
+        db_path: ${env.SQLITE_STORE_DIR:~/.llama/distributions/dell}/responses_store.db
   telemetry:
   - provider_id: meta-reference
     provider_type: inline::meta-reference
@@ -99,6 +102,9 @@ providers:
 metadata_store:
   type: sqlite
   db_path: ${env.SQLITE_STORE_DIR:~/.llama/distributions/dell}/registry.db
+inference_store:
+  type: sqlite
+  db_path: ${env.SQLITE_STORE_DIR:~/.llama/distributions/dell}/inference_store.db
 models:
 - metadata: {}
   model_id: ${env.INFERENCE_MODEL}

llama_stack/templates/dell/run.yaml

@@ -37,6 +37,9 @@ providers:
         type: sqlite
         namespace: null
         db_path: ${env.SQLITE_STORE_DIR:~/.llama/distributions/dell}/agents_store.db
+      responses_store:
+        type: sqlite
+        db_path: ${env.SQLITE_STORE_DIR:~/.llama/distributions/dell}/responses_store.db
   telemetry:
   - provider_id: meta-reference
     provider_type: inline::meta-reference
@@ -95,6 +98,9 @@ providers:
 metadata_store:
   type: sqlite
   db_path: ${env.SQLITE_STORE_DIR:~/.llama/distributions/dell}/registry.db
+inference_store:
+  type: sqlite
+  db_path: ${env.SQLITE_STORE_DIR:~/.llama/distributions/dell}/inference_store.db
 models:
 - metadata: {}
   model_id: ${env.INFERENCE_MODEL}

llama_stack/templates/dependencies.json

@@ -1,843 +0,0 @@
-{
-  "bedrock": [
-    "aiosqlite",
-    "autoevals",
-    "blobfile",
-    "boto3",
-    "chardet",
-    "chromadb-client",
-    "datasets",
-    "emoji",
-    "faiss-cpu",
-    "fastapi",
-    "fire",
-    "httpx",
-    "langdetect",
-    "matplotlib",
-    "mcp",
-    "nltk",
-    "numpy",
-    "openai",
-    "opentelemetry-exporter-otlp-proto-http",
-    "opentelemetry-sdk",
-    "pandas",
-    "pillow",
-    "psycopg2-binary",
-    "pymongo",
-    "pypdf",
-    "pythainlp",
-    "redis",
-    "requests",
-    "scikit-learn",
-    "scipy",
-    "sentencepiece",
-    "tqdm",
-    "transformers",
-    "tree_sitter",
-    "uvicorn"
-  ],
-  "cerebras": [
-    "aiosqlite",
-    "autoevals",
-    "blobfile",
-    "cerebras_cloud_sdk",
-    "chardet",
-    "chromadb-client",
-    "datasets",
-    "emoji",
-    "faiss-cpu",
-    "fastapi",
-    "fire",
-    "httpx",
-    "langdetect",
-    "matplotlib",
-    "nltk",
-    "numpy",
-    "openai",
-    "opentelemetry-exporter-otlp-proto-http",
-    "opentelemetry-sdk",
-    "pandas",
-    "pillow",
-    "psycopg2-binary",
-    "pymongo",
-    "pypdf",
-    "pythainlp",
-    "redis",
-    "requests",
-    "scikit-learn",
-    "scipy",
-    "sentencepiece",
-    "tqdm",
-    "transformers",
-    "tree_sitter",
-    "uvicorn",
-    "sentence-transformers --no-deps",
-    "torch torchvision --index-url https://download.pytorch.org/whl/cpu"
-  ],
-  "ci-tests": [
-    "aiosqlite",
-    "autoevals",
-    "blobfile",
-    "chardet",
-    "chromadb-client",
-    "datasets",
-    "emoji",
-    "fastapi",
-    "fire",
-    "fireworks-ai",
-    "httpx",
-    "langdetect",
-    "matplotlib",
-    "mcp",
-    "nltk",
-    "numpy",
-    "openai",
-    "opentelemetry-exporter-otlp-proto-http",
-    "opentelemetry-sdk",
-    "pandas",
-    "pillow",
-    "psycopg2-binary",
-    "pymongo",
-    "pypdf",
-    "pythainlp",
-    "redis",
-    "requests",
-    "scikit-learn",
-    "scipy",
-    "sentencepiece",
-    "sqlite-vec",
-    "tqdm",
-    "transformers",
-    "tree_sitter",
-    "uvicorn",
-    "sentence-transformers --no-deps",
-    "torch torchvision --index-url https://download.pytorch.org/whl/cpu"
-  ],
-  "dell": [
-    "aiohttp",
-    "aiosqlite",
-    "autoevals",
-    "blobfile",
-    "chardet",
-    "chromadb-client",
-    "datasets",
-    "emoji",
-    "faiss-cpu",
-    "fastapi",
-    "fire",
-    "httpx",
-    "huggingface_hub",
-    "langdetect",
-    "matplotlib",
-    "nltk",
-    "numpy",
-    "openai",
-    "opentelemetry-exporter-otlp-proto-http",
-    "opentelemetry-sdk",
-    "pandas",
-    "pillow",
-    "psycopg2-binary",
-    "pymongo",
-    "pypdf",
-    "pythainlp",
-    "redis",
-    "requests",
-    "scikit-learn",
-    "scipy",
-    "sentencepiece",
-    "tqdm",
-    "transformers",
-    "tree_sitter",
-    "uvicorn",
-    "sentence-transformers --no-deps",
-    "torch torchvision --index-url https://download.pytorch.org/whl/cpu"
-  ],
-  "fireworks": [
-    "aiosqlite",
-    "autoevals",
-    "blobfile",
-    "chardet",
-    "chromadb-client",
-    "datasets",
-    "emoji",
-    "faiss-cpu",
-    "fastapi",
-    "fire",
-    "fireworks-ai",
-    "httpx",
-    "langdetect",
-    "matplotlib",
-    "mcp",
-    "nltk",
-    "numpy",
-    "openai",
-    "opentelemetry-exporter-otlp-proto-http",
-    "opentelemetry-sdk",
-    "pandas",
-    "pillow",
-    "psycopg2-binary",
-    "pymongo",
-    "pypdf",
-    "pythainlp",
-    "redis",
-    "requests",
-    "scikit-learn",
-    "scipy",
-    "sentencepiece",
-    "tqdm",
-    "transformers",
-    "tree_sitter",
-    "uvicorn",
-    "sentence-transformers --no-deps",
-    "torch torchvision --index-url https://download.pytorch.org/whl/cpu"
-  ],
-  "groq": [
-    "aiosqlite",
-    "autoevals",
-    "blobfile",
-    "chardet",
-    "datasets",
-    "emoji",
-    "faiss-cpu",
-    "fastapi",
-    "fire",
-    "httpx",
-    "langdetect",
-    "litellm",
-    "matplotlib",
-    "nltk",
-    "numpy",
-    "openai",
-    "opentelemetry-exporter-otlp-proto-http",
-    "opentelemetry-sdk",
-    "pandas",
-    "pillow",
-    "psycopg2-binary",
-    "pymongo",
-    "pypdf",
-    "pythainlp",
-    "redis",
-    "requests",
-    "scikit-learn",
-    "scipy",
-    "sentencepiece",
-    "tqdm",
-    "transformers",
-    "tree_sitter",
-    "uvicorn"
-  ],
-  "hf-endpoint": [
-    "aiohttp",
-    "aiosqlite",
-    "autoevals",
-    "blobfile",
-    "chardet",
-    "chromadb-client",
-    "datasets",
-    "emoji",
-    "faiss-cpu",
-    "fastapi",
-    "fire",
-    "httpx",
-    "huggingface_hub",
-    "langdetect",
-    "matplotlib",
-    "mcp",
-    "nltk",
-    "numpy",
-    "openai",
-    "opentelemetry-exporter-otlp-proto-http",
-    "opentelemetry-sdk",
-    "pandas",
-    "pillow",
-    "psycopg2-binary",
-    "pymongo",
-    "pypdf",
-    "pythainlp",
-    "redis",
-    "requests",
-    "scikit-learn",
-    "scipy",
-    "sentencepiece",
-    "tqdm",
-    "transformers",
-    "tree_sitter",
-    "uvicorn"
-  ],
-  "hf-serverless": [
-    "aiohttp",
-    "aiosqlite",
-    "autoevals",
-    "blobfile",
-    "chardet",
-    "chromadb-client",
-    "datasets",
-    "emoji",
-    "faiss-cpu",
-    "fastapi",
-    "fire",
-    "httpx",
-    "huggingface_hub",
-    "langdetect",
-    "matplotlib",
-    "mcp",
-    "nltk",
-    "numpy",
-    "openai",
-    "opentelemetry-exporter-otlp-proto-http",
-    "opentelemetry-sdk",
-    "pandas",
-    "pillow",
-    "psycopg2-binary",
-    "pymongo",
-    "pypdf",
-    "pythainlp",
-    "redis",
-    "requests",
-    "scikit-learn",
-    "scipy",
-    "sentencepiece",
-    "tqdm",
-    "transformers",
-    "tree_sitter",
-    "uvicorn",
-    "sentence-transformers --no-deps",
-    "torch torchvision --index-url https://download.pytorch.org/whl/cpu"
-  ],
-  "llama_api": [
-    "aiosqlite",
-    "autoevals",
-    "blobfile",
-    "chardet",
-    "chromadb-client",
-    "datasets",
-    "emoji",
-    "fastapi",
-    "fire",
-    "httpx",
-    "langdetect",
-    "litellm",
-    "matplotlib",
-    "mcp",
-    "nltk",
-    "numpy",
-    "openai",
-    "opentelemetry-exporter-otlp-proto-http",
-    "opentelemetry-sdk",
-    "pandas",
-    "pillow",
-    "psycopg2-binary",
-    "pymongo",
-    "pypdf",
-    "pythainlp",
-    "redis",
-    "requests",
-    "scikit-learn",
-    "scipy",
-    "sentencepiece",
-    "sqlite-vec",
-    "tqdm",
-    "transformers",
-    "tree_sitter",
-    "uvicorn",
-    "sentence-transformers --no-deps",
-    "torch torchvision --index-url https://download.pytorch.org/whl/cpu"
-  ],
-  "meta-reference-gpu": [
-    "accelerate",
-    "aiosqlite",
-    "autoevals",
-    "blobfile",
-    "chardet",
-    "chromadb-client",
-    "datasets",
-    "emoji",
-    "fairscale",
-    "faiss-cpu",
-    "fastapi",
-    "fbgemm-gpu-genai==1.1.2",
-    "fire",
-    "httpx",
-    "langdetect",
-    "lm-format-enforcer",
-    "matplotlib",
-    "mcp",
-    "nltk",
-    "numpy",
-    "openai",
-    "opentelemetry-exporter-otlp-proto-http",
-    "opentelemetry-sdk",
-    "pandas",
-    "pillow",
-    "psycopg2-binary",
-    "pymongo",
-    "pypdf",
-    "pythainlp",
-    "redis",
-    "requests",
-    "scikit-learn",
-    "scipy",
-    "sentence-transformers",
-    "sentencepiece",
-    "torch",
-    "torchao==0.8.0",
-    "torchvision",
-    "tqdm",
-    "transformers",
-    "tree_sitter",
-    "uvicorn",
-    "zmq"
-  ],
-  "nvidia": [
-    "aiohttp",
-    "aiosqlite",
-    "blobfile",
-    "chardet",
-    "datasets",
-    "faiss-cpu",
-    "fastapi",
-    "fire",
-    "httpx",
-    "matplotlib",
-    "nltk",
-    "numpy",
-    "openai",
-    "opentelemetry-exporter-otlp-proto-http",
-    "opentelemetry-sdk",
-    "pandas",
-    "pillow",
-    "psycopg2-binary",
-    "pymongo",
-    "pypdf",
-    "redis",
-    "requests",
-    "scikit-learn",
-    "scipy",
-    "sentencepiece",
-    "tqdm",
-    "transformers",
-    "uvicorn"
-  ],
-  "ollama": [
-    "aiohttp",
-    "aiosqlite",
-    "autoevals",
-    "blobfile",
-    "chardet",
-    "chromadb-client",
-    "datasets",
-    "emoji",
-    "faiss-cpu",
-    "fastapi",
-    "fire",
-    "httpx",
-    "langdetect",
-    "matplotlib",
-    "mcp",
-    "nltk",
-    "numpy",
-    "ollama",
-    "openai",
-    "opentelemetry-exporter-otlp-proto-http",
-    "opentelemetry-sdk",
-    "pandas",
-    "peft",
-    "pillow",
-    "psycopg2-binary",
-    "pymongo",
-    "pypdf",
-    "pythainlp",
-    "redis",
-    "requests",
-    "scikit-learn",
-    "scipy",
-    "sentencepiece",
-    "torch",
-    "tqdm",
-    "transformers",
-    "tree_sitter",
-    "trl",
-    "uvicorn"
-  ],
-  "open-benchmark": [
-    "aiosqlite",
-    "autoevals",
-    "blobfile",
-    "chardet",
-    "chromadb-client",
-    "datasets",
-    "emoji",
-    "fastapi",
-    "fire",
-    "httpx",
-    "langdetect",
-    "litellm",
-    "matplotlib",
-    "mcp",
-    "nltk",
-    "numpy",
-    "openai",
-    "opentelemetry-exporter-otlp-proto-http",
-    "opentelemetry-sdk",
-    "pandas",
-    "pillow",
-    "psycopg2-binary",
-    "pymongo",
-    "pypdf",
-    "pythainlp",
-    "redis",
-    "requests",
-    "scikit-learn",
-    "scipy",
-    "sentencepiece",
-    "sqlite-vec",
-    "together",
-    "tqdm",
-    "transformers",
-    "tree_sitter",
-    "uvicorn"
-  ],
-  "passthrough": [
-    "aiosqlite",
-    "autoevals",
-    "blobfile",
-    "chardet",
-    "chromadb-client",
-    "datasets",
-    "emoji",
-    "faiss-cpu",
-    "fastapi",
-    "fire",
-    "httpx",
-    "langdetect",
-    "matplotlib",
-    "mcp",
-    "nltk",
-    "numpy",
-    "openai",
-    "opentelemetry-exporter-otlp-proto-http",
-    "opentelemetry-sdk",
-    "pandas",
-    "pillow",
-    "psycopg2-binary",
-    "pymongo",
-    "pypdf",
-    "pythainlp",
-    "redis",
-    "requests",
-    "scikit-learn",
-    "scipy",
-    "sentencepiece",
-    "tqdm",
-    "transformers",
-    "tree_sitter",
-    "uvicorn",
-    "sentence-transformers --no-deps",
-    "torch torchvision --index-url https://download.pytorch.org/whl/cpu"
-  ],
-  "remote-vllm": [
-    "aiosqlite",
-    "autoevals",
-    "blobfile",
-    "chardet",
-    "chromadb-client",
-    "datasets",
-    "emoji",
-    "faiss-cpu",
-    "fastapi",
-    "fire",
-    "httpx",
-    "langdetect",
-    "matplotlib",
-    "mcp",
-    "nltk",
-    "numpy",
-    "openai",
-    "opentelemetry-exporter-otlp-proto-http",
-    "opentelemetry-sdk",
-    "pandas",
-    "pillow",
-    "psycopg2-binary",
-    "pymongo",
-    "pypdf",
-    "pythainlp",
-    "redis",
-    "requests",
-    "scikit-learn",
-    "scipy",
-    "sentencepiece",
-    "tqdm",
-    "transformers",
-    "tree_sitter",
-    "uvicorn",
-    "sentence-transformers --no-deps",
-    "torch torchvision --index-url https://download.pytorch.org/whl/cpu"
-  ],
-  "sambanova": [
-    "aiosqlite",
-    "blobfile",
-    "chardet",
-    "chromadb-client",
-    "faiss-cpu",
-    "fastapi",
-    "fire",
-    "httpx",
-    "litellm",
-    "matplotlib",
-    "mcp",
-    "nltk",
-    "numpy",
-    "opentelemetry-exporter-otlp-proto-http",
-    "opentelemetry-sdk",
-    "pandas",
-    "pillow",
-    "psycopg2-binary",
-    "pymongo",
-    "pypdf",
-    "redis",
-    "requests",
-    "scikit-learn",
-    "scipy",
-    "sentencepiece",
-    "tqdm",
-    "transformers",
-    "uvicorn",
-    "sentence-transformers --no-deps",
-    "torch torchvision --index-url https://download.pytorch.org/whl/cpu"
-  ],
-  "starter": [
-    "aiosqlite",
-    "autoevals",
-    "blobfile",
-    "chardet",
-    "chromadb-client",
-    "datasets",
-    "emoji",
-    "fastapi",
-    "fire",
-    "fireworks-ai",
-    "httpx",
-    "langdetect",
-    "litellm",
-    "matplotlib",
-    "mcp",
-    "nltk",
-    "numpy",
-    "openai",
-    "opentelemetry-exporter-otlp-proto-http",
-    "opentelemetry-sdk",
-    "pandas",
-    "pillow",
-    "psycopg2-binary",
-    "pymongo",
-    "pypdf",
-    "pythainlp",
-    "redis",
-    "requests",
-    "scikit-learn",
-    "scipy",
-    "sentencepiece",
-    "sqlite-vec",
-    "tqdm",
-    "transformers",
-    "tree_sitter",
-    "uvicorn",
-    "sentence-transformers --no-deps",
-    "torch torchvision --index-url https://download.pytorch.org/whl/cpu"
-  ],
-  "tgi": [
-    "aiohttp",
-    "aiosqlite",
-    "autoevals",
-    "blobfile",
-    "chardet",
-    "chromadb-client",
-    "datasets",
-    "emoji",
-    "faiss-cpu",
-    "fastapi",
-    "fire",
-    "httpx",
-    "huggingface_hub",
-    "langdetect",
-    "matplotlib",
-    "mcp",
-    "nltk",
-    "numpy",
-    "openai",
-    "opentelemetry-exporter-otlp-proto-http",
-    "opentelemetry-sdk",
-    "pandas",
-    "pillow",
-    "psycopg2-binary",
-    "pymongo",
-    "pypdf",
-    "pythainlp",
-    "redis",
-    "requests",
-    "scikit-learn",
-    "scipy",
-    "sentencepiece",
-    "tqdm",
-    "transformers",
-    "tree_sitter",
-    "uvicorn",
-    "sentence-transformers --no-deps",
-    "torch torchvision --index-url https://download.pytorch.org/whl/cpu"
-  ],
-  "together": [
-    "aiosqlite",
-    "autoevals",
-    "blobfile",
-    "chardet",
-    "chromadb-client",
-    "datasets",
-    "emoji",
-    "faiss-cpu",
-    "fastapi",
-    "fire",
-    "httpx",
-    "langdetect",
-    "matplotlib",
-    "mcp",
-    "nltk",
-    "numpy",
-    "openai",
-    "opentelemetry-exporter-otlp-proto-http",
-    "opentelemetry-sdk",
-    "pandas",
-    "pillow",
-    "psycopg2-binary",
-    "pymongo",
-    "pypdf",
-    "pythainlp",
-    "redis",
-    "requests",
-    "scikit-learn",
-    "scipy",
-    "sentencepiece",
-    "together",
-    "tqdm",
-    "transformers",
-    "tree_sitter",
-    "uvicorn",
-    "sentence-transformers --no-deps",
-    "torch torchvision --index-url https://download.pytorch.org/whl/cpu"
-  ],
-  "verification": [
-    "aiosqlite",
-    "autoevals",
-    "blobfile",
-    "chardet",
-    "chromadb-client",
-    "datasets",
-    "emoji",
-    "fastapi",
-    "fire",
-    "httpx",
-    "langdetect",
-    "litellm",
-    "matplotlib",
-    "mcp",
-    "nltk",
-    "numpy",
-    "openai",
-    "opentelemetry-exporter-otlp-proto-http",
-    "opentelemetry-sdk",
-    "pandas",
-    "pillow",
-    "psycopg2-binary",
-    "pymongo",
-    "pypdf",
-    "pythainlp",
-    "redis",
-    "requests",
-    "scikit-learn",
-    "scipy",
-    "sentencepiece",
-    "sqlite-vec",
-    "tqdm",
-    "transformers",
-    "tree_sitter",
-    "uvicorn",
-    "sentence-transformers --no-deps",
-    "torch torchvision --index-url https://download.pytorch.org/whl/cpu"
-  ],
-  "vllm-gpu": [
-    "aiosqlite",
-    "autoevals",
-    "blobfile",
-    "chardet",
-    "chromadb-client",
-    "datasets",
-    "emoji",
-    "faiss-cpu",
-    "fastapi",
-    "fire",
-    "httpx",
-    "langdetect",
-    "matplotlib",
-    "mcp",
-    "nltk",
-    "numpy",
-    "openai",
-    "opentelemetry-exporter-otlp-proto-http",
-    "opentelemetry-sdk",
-    "pandas",
-    "pillow",
-    "psycopg2-binary",
-    "pymongo",
-    "pypdf",
-    "pythainlp",
-    "redis",
-    "requests",
-    "scikit-learn",
-    "scipy",
-    "sentencepiece",
-    "tqdm",
-    "transformers",
-    "tree_sitter",
-    "uvicorn",
-    "vllm",
-    "sentence-transformers --no-deps",
-    "torch torchvision --index-url https://download.pytorch.org/whl/cpu"
-  ],
-  "watsonx": [
-    "aiosqlite",
-    "autoevals",
-    "blobfile",
-    "chardet",
-    "datasets",
-    "emoji",
-    "faiss-cpu",
-    "fastapi",
-    "fire",
-    "httpx",
-    "ibm_watson_machine_learning",
-    "langdetect",
-    "matplotlib",
-    "mcp",
-    "nltk",
-    "numpy",
-    "openai",
-    "opentelemetry-exporter-otlp-proto-http",
-    "opentelemetry-sdk",
-    "pandas",
-    "pillow",
-    "psycopg2-binary",
-    "pymongo",
-    "pypdf",
-    "pythainlp",
-    "redis",
-    "requests",
-    "scikit-learn",
-    "scipy",
-    "sentencepiece",
-    "tqdm",
-    "transformers",
-    "tree_sitter",
-    "uvicorn",
-    "sentence-transformers --no-deps",
-    "torch torchvision --index-url https://download.pytorch.org/whl/cpu"
-  ]
-}

llama_stack/templates/fireworks/build.yaml

@@ -31,3 +31,6 @@ distribution_spec:
     - inline::rag-runtime
     - remote::model-context-protocol
 image_type: conda
+additional_pip_packages:
+- sqlalchemy[asyncio]
+- sqlalchemy[asyncio]

llama_stack/templates/fireworks/run-with-safety.yaml

@@ -46,6 +46,9 @@ providers:
         type: sqlite
         namespace: null
         db_path: ${env.SQLITE_STORE_DIR:~/.llama/distributions/fireworks}/agents_store.db
+      responses_store:
+        type: sqlite
+        db_path: ${env.SQLITE_STORE_DIR:~/.llama/distributions/fireworks}/responses_store.db
   telemetry:
   - provider_id: meta-reference
     provider_type: inline::meta-reference
@@ -111,6 +114,9 @@ providers:
 metadata_store:
   type: sqlite
   db_path: ${env.SQLITE_STORE_DIR:~/.llama/distributions/fireworks}/registry.db
+inference_store:
+  type: sqlite
+  db_path: ${env.SQLITE_STORE_DIR:~/.llama/distributions/fireworks}/inference_store.db
 models:
 - metadata: {}
   model_id: accounts/fireworks/models/llama-v3p1-8b-instruct

llama_stack/templates/fireworks/run.yaml

@@ -41,6 +41,9 @@ providers:
         type: sqlite
         namespace: null
         db_path: ${env.SQLITE_STORE_DIR:~/.llama/distributions/fireworks}/agents_store.db
+      responses_store:
+        type: sqlite
+        db_path: ${env.SQLITE_STORE_DIR:~/.llama/distributions/fireworks}/responses_store.db
   telemetry:
   - provider_id: meta-reference
     provider_type: inline::meta-reference
@@ -106,6 +109,9 @@ providers:
 metadata_store:
   type: sqlite
   db_path: ${env.SQLITE_STORE_DIR:~/.llama/distributions/fireworks}/registry.db
+inference_store:
+  type: sqlite
+  db_path: ${env.SQLITE_STORE_DIR:~/.llama/distributions/fireworks}/inference_store.db
 models:
 - metadata: {}
   model_id: accounts/fireworks/models/llama-v3p1-8b-instruct

llama_stack/templates/groq/build.yaml

@@ -26,3 +26,5 @@ distribution_spec:
     - remote::tavily-search
     - inline::rag-runtime
 image_type: conda
+additional_pip_packages:
+- sqlalchemy[asyncio]

llama_stack/templates/groq/run.yaml

@@ -41,6 +41,9 @@ providers:
         type: sqlite
         namespace: null
         db_path: ${env.SQLITE_STORE_DIR:~/.llama/distributions/groq}/agents_store.db
+      responses_store:
+        type: sqlite
+        db_path: ${env.SQLITE_STORE_DIR:~/.llama/distributions/groq}/responses_store.db
   telemetry:
   - provider_id: meta-reference
     provider_type: inline::meta-reference
@@ -99,6 +102,9 @@ providers:
 metadata_store:
   type: sqlite
   db_path: ${env.SQLITE_STORE_DIR:~/.llama/distributions/groq}/registry.db
+inference_store:
+  type: sqlite
+  db_path: ${env.SQLITE_STORE_DIR:~/.llama/distributions/groq}/inference_store.db
 models:
 - metadata: {}
   model_id: groq/llama3-8b-8192