chore(api)!: remove tool_runtime.rag_tool from the API surface (#4067)

RAG aka file search is implemented via the Responses API by specifying
the file-search tool. The backend implementation remains unchanged. This
PR merely removes the directly exposed API surface which allowed users
to directly perform searches from the client.

This facility is now available via the `client.vector_store.search()`
OpenAI compatible API.

src/llama_stack/apis/tools/rag_tool.py

@@ -5,18 +5,13 @@
 # the root directory of this source tree.
 
 from enum import Enum, StrEnum
-from typing import Annotated, Any, Literal, Protocol
+from typing import Annotated, Any, Literal
 
 from pydantic import BaseModel, Field, field_validator
-from typing_extensions import runtime_checkable
 
 from llama_stack.apis.common.content_types import URL, InterleavedContent
-from llama_stack.apis.version import LLAMA_STACK_API_V1
-from llama_stack.core.telemetry.trace_protocol import trace_protocol
-from llama_stack.schema_utils import json_schema_type, register_schema, webmethod
 
 
-@json_schema_type
 class RRFRanker(BaseModel):
     """
     Reciprocal Rank Fusion (RRF) ranker configuration.
@@ -30,7 +25,6 @@ class RRFRanker(BaseModel):
     impact_factor: float = Field(default=60.0, gt=0.0)  # default of 60 for optimal performance
 
 
-@json_schema_type
 class WeightedRanker(BaseModel):
     """
     Weighted ranker configuration that combines vector and keyword scores.
@@ -55,10 +49,8 @@ Ranker = Annotated[
     RRFRanker | WeightedRanker,
     Field(discriminator="type"),
 ]
-register_schema(Ranker, name="Ranker")
 
 
-@json_schema_type
 class RAGDocument(BaseModel):
     """
     A document to be used for document ingestion in the RAG Tool.
@@ -75,7 +67,6 @@ class RAGDocument(BaseModel):
     metadata: dict[str, Any] = Field(default_factory=dict)
 
 
-@json_schema_type
 class RAGQueryResult(BaseModel):
     """Result of a RAG query containing retrieved content and metadata.
 
@@ -87,7 +78,6 @@ class RAGQueryResult(BaseModel):
     metadata: dict[str, Any] = Field(default_factory=dict)
 
 
-@json_schema_type
 class RAGQueryGenerator(Enum):
     """Types of query generators for RAG systems.
 
@@ -101,7 +91,6 @@ class RAGQueryGenerator(Enum):
     custom = "custom"
 
 
-@json_schema_type
 class RAGSearchMode(StrEnum):
     """
     Search modes for RAG query retrieval:
@@ -115,7 +104,6 @@ class RAGSearchMode(StrEnum):
     HYBRID = "hybrid"
 
 
-@json_schema_type
 class DefaultRAGQueryGeneratorConfig(BaseModel):
     """Configuration for the default RAG query generator.
 
@@ -127,7 +115,6 @@ class DefaultRAGQueryGeneratorConfig(BaseModel):
     separator: str = " "
 
 
-@json_schema_type
 class LLMRAGQueryGeneratorConfig(BaseModel):
     """Configuration for the LLM-based RAG query generator.
 
@@ -145,10 +132,8 @@ RAGQueryGeneratorConfig = Annotated[
     DefaultRAGQueryGeneratorConfig | LLMRAGQueryGeneratorConfig,
     Field(discriminator="type"),
 ]
-register_schema(RAGQueryGeneratorConfig, name="RAGQueryGeneratorConfig")
 
 
-@json_schema_type
 class RAGQueryConfig(BaseModel):
     """
     Configuration for the RAG query generation.
@@ -181,38 +166,3 @@ class RAGQueryConfig(BaseModel):
         if len(v) == 0:
             raise ValueError("chunk_template must not be empty")
         return v
-
-
-@runtime_checkable
-@trace_protocol
-class RAGToolRuntime(Protocol):
-    @webmethod(route="/tool-runtime/rag-tool/insert", method="POST", level=LLAMA_STACK_API_V1)
-    async def insert(
-        self,
-        documents: list[RAGDocument],
-        vector_store_id: str,
-        chunk_size_in_tokens: int = 512,
-    ) -> None:
-        """Index documents so they can be used by the RAG system.
-
-        :param documents: List of documents to index in the RAG system
-        :param vector_store_id: ID of the vector database to store the document embeddings
-        :param chunk_size_in_tokens: (Optional) Size in tokens for document chunking during indexing
-        """
-        ...
-
-    @webmethod(route="/tool-runtime/rag-tool/query", method="POST", level=LLAMA_STACK_API_V1)
-    async def query(
-        self,
-        content: InterleavedContent,
-        vector_store_ids: list[str],
-        query_config: RAGQueryConfig | None = None,
-    ) -> RAGQueryResult:
-        """Query the RAG system for context; typically invoked by the agent.
-
-        :param content: The query content to search for in the indexed documents
-        :param vector_store_ids: List of vector database IDs to search within
-        :param query_config: (Optional) Configuration parameters for the query operation
-        :returns: RAGQueryResult containing the retrieved content and metadata
-        """
-        ...

src/llama_stack/apis/tools/tools.py

@@ -16,8 +16,6 @@ from llama_stack.apis.version import LLAMA_STACK_API_V1
 from llama_stack.core.telemetry.trace_protocol import trace_protocol
 from llama_stack.schema_utils import json_schema_type, webmethod
 
-from .rag_tool import RAGToolRuntime
-
 
 @json_schema_type
 class ToolDef(BaseModel):
@@ -195,8 +193,6 @@ class SpecialToolGroup(Enum):
 class ToolRuntime(Protocol):
     tool_store: ToolStore | None = None
 
-    rag_tool: RAGToolRuntime | None = None
-
     # TODO: This needs to be renamed once OPEN API generator name conflict issue is fixed.
     @webmethod(route="/tool-runtime/list-tools", method="GET", level=LLAMA_STACK_API_V1)
     async def list_runtime_tools(

src/llama_stack/core/routers/tool_runtime.py

@@ -8,14 +8,9 @@ from typing import Any
 
 from llama_stack.apis.common.content_types import (
     URL,
-    InterleavedContent,
 )
 from llama_stack.apis.tools import (
     ListToolDefsResponse,
-    RAGDocument,
-    RAGQueryConfig,
-    RAGQueryResult,
-    RAGToolRuntime,
     ToolRuntime,
 )
 from llama_stack.log import get_logger
@@ -26,36 +21,6 @@ logger = get_logger(name=__name__, category="core::routers")
 
 
 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_store_ids: list[str],
-            query_config: RAGQueryConfig | None = None,
-        ) -> RAGQueryResult:
-            logger.debug(f"ToolRuntimeRouter.RagToolImpl.query: {vector_store_ids}")
-            provider = await self.routing_table.get_provider_impl("knowledge_search")
-            return await provider.query(content, vector_store_ids, query_config)
-
-        async def insert(
-            self,
-            documents: list[RAGDocument],
-            vector_store_id: str,
-            chunk_size_in_tokens: int = 512,
-        ) -> None:
-            logger.debug(
-                f"ToolRuntimeRouter.RagToolImpl.insert: {vector_store_id}, {len(documents)} documents, chunk_size={chunk_size_in_tokens}"
-            )
-            provider = await self.routing_table.get_provider_impl("insert_into_memory")
-            return await provider.insert(documents, vector_store_id, chunk_size_in_tokens)
-
     def __init__(
         self,
         routing_table: ToolGroupsRoutingTable,
@@ -63,11 +28,6 @@ class ToolRuntimeRouter(ToolRuntime):
         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

src/llama_stack/core/server/routes.py

@@ -13,7 +13,6 @@ from aiohttp import hdrs
 from starlette.routing import Route
 
 from llama_stack.apis.datatypes import Api, ExternalApiSpec
-from llama_stack.apis.tools import RAGToolRuntime, SpecialToolGroup
 from llama_stack.core.resolver import api_protocol_map
 from llama_stack.schema_utils import WebMethod
 
@@ -25,33 +24,16 @@ RouteImpls = dict[str, PathImpl]
 RouteMatch = tuple[EndpointFunc, PathParams, str, WebMethod]
 
 
-def toolgroup_protocol_map():
-    return {
-        SpecialToolGroup.rag_tool: RAGToolRuntime,
-    }
-
-
 def get_all_api_routes(
     external_apis: dict[Api, ExternalApiSpec] | None = None,
 ) -> dict[Api, list[tuple[Route, WebMethod]]]:
     apis = {}
 
     protocols = api_protocol_map(external_apis)
-    toolgroup_protocols = toolgroup_protocol_map()
     for api, protocol in protocols.items():
         routes = []
         protocol_methods = inspect.getmembers(protocol, predicate=inspect.isfunction)
 
-        # HACK ALERT
-        if api == Api.tool_runtime:
-            for tool_group in SpecialToolGroup:
-                sub_protocol = toolgroup_protocols[tool_group]
-                sub_protocol_methods = inspect.getmembers(sub_protocol, predicate=inspect.isfunction)
-                for name, method in sub_protocol_methods:
-                    if not hasattr(method, "__webmethod__"):
-                        continue
-                    protocol_methods.append((f"{tool_group.value}.{name}", method))
-
         for name, method in protocol_methods:
             # Get all webmethods for this method (supports multiple decorators)
             webmethods = getattr(method, "__webmethods__", [])

src/llama_stack/core/stack.py

@@ -31,7 +31,7 @@ from llama_stack.apis.safety import Safety
 from llama_stack.apis.scoring import Scoring
 from llama_stack.apis.scoring_functions import ScoringFunctions
 from llama_stack.apis.shields import Shields
-from llama_stack.apis.tools import RAGToolRuntime, ToolGroups, ToolRuntime
+from llama_stack.apis.tools import ToolGroups, ToolRuntime
 from llama_stack.apis.vector_io import VectorIO
 from llama_stack.core.conversations.conversations import ConversationServiceConfig, ConversationServiceImpl
 from llama_stack.core.datatypes import Provider, SafetyConfig, StackRunConfig, VectorStoresConfig
@@ -78,7 +78,6 @@ class LlamaStack(
     Inspect,
     ToolGroups,
     ToolRuntime,
-    RAGToolRuntime,
     Files,
     Prompts,
     Conversations,

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

@@ -27,7 +27,6 @@ from llama_stack.apis.tools import (
     RAGDocument,
     RAGQueryConfig,
     RAGQueryResult,
-    RAGToolRuntime,
     ToolDef,
     ToolGroup,
     ToolInvocationResult,
@@ -91,7 +90,7 @@ async def raw_data_from_doc(doc: RAGDocument) -> tuple[bytes, str]:
             return content_str.encode("utf-8"), "text/plain"
 
 
-class MemoryToolRuntimeImpl(ToolGroupsProtocolPrivate, ToolRuntime, RAGToolRuntime):
+class MemoryToolRuntimeImpl(ToolGroupsProtocolPrivate, ToolRuntime):
     def __init__(
         self,
         config: RagToolRuntimeConfig,