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/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,