feat: File search tool for Responses API

This is an initial working prototype of wiring up the `file_search` builtin
tool for the Responses API to our existing rag knowledge search tool.

I stubbed in a new test (that uses a hardcoded url hybrid of the OpenAI
and Llama Stack clients for now, only until we finish landing the vector
store APIs and insertion support).

Note that this is currently under tests/verification only because it
sometimes flakes with tool calling of the small Llama-3.2-3B model we run
in CI (and that I use as an example below). We'd want to make the test a
bit more robust in some way if we moved this over to tests/integration and
ran it in CI.

```
ollama run llama3.2:3b

INFERENCE_MODEL="meta-llama/Llama-3.2-3B-Instruct" \
llama stack run ./llama_stack/templates/ollama/run.yaml \
  --image-type venv \
  --env OLLAMA_URL="http://0.0.0.0:11434"

pytest -sv 'tests/verifications/openai_api/test_responses.py::test_response_non_streaming_file_search' \
  --base-url=http://localhost:8321/v1/openai/v1 \
  --model meta-llama/Llama-3.2-3B-Instruct
```

Signed-off-by: Ben Browning <bbrownin@redhat.com>

llama_stack/apis/agents/openai_responses.py

@@ -81,6 +81,15 @@ class OpenAIResponseOutputMessageWebSearchToolCall(BaseModel):
     type: Literal["web_search_call"] = "web_search_call"
 
 
+@json_schema_type
+class OpenAIResponseOutputMessageFileSearchToolCall(BaseModel):
+    id: str
+    queries: list[str]
+    status: str
+    type: Literal["file_search_call"] = "file_search_call"
+    results: list[dict[str, Any]] | None = None
+
+
 @json_schema_type
 class OpenAIResponseOutputMessageFunctionToolCall(BaseModel):
     call_id: str
@@ -119,6 +128,7 @@ class OpenAIResponseOutputMessageMCPListTools(BaseModel):
 OpenAIResponseOutput = Annotated[
     OpenAIResponseMessage
     | OpenAIResponseOutputMessageWebSearchToolCall
+    | OpenAIResponseOutputMessageFileSearchToolCall
     | OpenAIResponseOutputMessageFunctionToolCall
     | OpenAIResponseOutputMessageMCPCall
     | OpenAIResponseOutputMessageMCPListTools,
@@ -362,6 +372,7 @@ class OpenAIResponseInputFunctionToolCallOutput(BaseModel):
 OpenAIResponseInput = Annotated[
     # Responses API allows output messages to be passed in as input
     OpenAIResponseOutputMessageWebSearchToolCall
+    | OpenAIResponseOutputMessageFileSearchToolCall
     | OpenAIResponseOutputMessageFunctionToolCall
     | OpenAIResponseInputFunctionToolCallOutput
     |
@@ -397,9 +408,9 @@ class FileSearchRankingOptions(BaseModel):
 @json_schema_type
 class OpenAIResponseInputToolFileSearch(BaseModel):
     type: Literal["file_search"] = "file_search"
-    vector_store_id: list[str]
+    vector_store_ids: list[str]
     ranking_options: FileSearchRankingOptions | None = None
-    # TODO: add filters
+    # TODO: add filters, max_num_results
 
 
 class ApprovalFilter(BaseModel):

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

@@ -24,6 +24,7 @@ from llama_stack.apis.agents.openai_responses import (
     OpenAIResponseInputMessageContentImage,
     OpenAIResponseInputMessageContentText,
     OpenAIResponseInputTool,
+    OpenAIResponseInputToolFileSearch,
     OpenAIResponseInputToolMCP,
     OpenAIResponseMessage,
     OpenAIResponseObject,
@@ -34,6 +35,7 @@ from llama_stack.apis.agents.openai_responses import (
     OpenAIResponseOutput,
     OpenAIResponseOutputMessageContent,
     OpenAIResponseOutputMessageContentOutputText,
+    OpenAIResponseOutputMessageFileSearchToolCall,
     OpenAIResponseOutputMessageFunctionToolCall,
     OpenAIResponseOutputMessageMCPListTools,
     OpenAIResponseOutputMessageWebSearchToolCall,
@@ -198,7 +200,8 @@ class OpenAIResponsePreviousResponseWithInputItems(BaseModel):
 class ChatCompletionContext(BaseModel):
     model: str
     messages: list[OpenAIMessageParam]
-    tools: list[ChatCompletionToolParam] | None = None
+    response_tools: list[OpenAIResponseInputTool] | None = None
+    chat_tools: list[ChatCompletionToolParam] | None = None
     mcp_tool_to_server: dict[str, OpenAIResponseInputToolMCP]
     temperature: float | None
     response_format: OpenAIResponseFormatParam
@@ -388,7 +391,8 @@ class OpenAIResponsesImpl:
         ctx = ChatCompletionContext(
             model=model,
             messages=messages,
-            tools=chat_tools,
+            response_tools=tools,
+            chat_tools=chat_tools,
             mcp_tool_to_server=mcp_tool_to_server,
             temperature=temperature,
             response_format=response_format,
@@ -417,7 +421,7 @@ class OpenAIResponsesImpl:
             completion_result = await self.inference_api.openai_chat_completion(
                 model=ctx.model,
                 messages=messages,
-                tools=ctx.tools,
+                tools=ctx.chat_tools,
                 stream=True,
                 temperature=ctx.temperature,
                 response_format=ctx.response_format,
@@ -606,6 +610,12 @@ class OpenAIResponsesImpl:
                 if not tool:
                     raise ValueError(f"Tool {tool_name} not found")
                 chat_tools.append(make_openai_tool(tool_name, tool))
+            elif input_tool.type == "file_search":
+                tool_name = "knowledge_search"
+                tool = await self.tool_groups_api.get_tool(tool_name)
+                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
@@ -667,6 +677,7 @@ class OpenAIResponsesImpl:
 
         tool_call_id = tool_call.id
         function = tool_call.function
+        tool_kwargs = json.loads(function.arguments) if function.arguments else {}
 
         if not function or not tool_call_id or not function.name:
             return None, None
@@ -680,12 +691,18 @@ class OpenAIResponsesImpl:
                     endpoint=mcp_tool.server_url,
                     headers=mcp_tool.headers or {},
                     tool_name=function.name,
-                    kwargs=json.loads(function.arguments) if function.arguments else {},
+                    kwargs=tool_kwargs,
                 )
             else:
+                if function.name == "knowledge_search":
+                    response_file_search_tool = next(
+                        t for t in ctx.response_tools if isinstance(t, OpenAIResponseInputToolFileSearch)
+                    )
+                    if response_file_search_tool:
+                        tool_kwargs["vector_db_ids"] = response_file_search_tool.vector_store_ids
                 result = await self.tool_runtime_api.invoke_tool(
                     tool_name=function.name,
-                    kwargs=json.loads(function.arguments) if function.arguments else {},
+                    kwargs=tool_kwargs,
                 )
         except Exception as e:
             error_exc = e
@@ -713,6 +730,27 @@ class OpenAIResponsesImpl:
                 )
                 if error_exc or (result.error_code and result.error_code > 0) or result.error_message:
                     message.status = "failed"
+            elif function.name == "knowledge_search":
+                message = OpenAIResponseOutputMessageFileSearchToolCall(
+                    id=tool_call_id,
+                    queries=[tool_kwargs.get("query", "")],
+                    status="completed",
+                )
+                if "document_ids" in result.metadata:
+                    message.results = []
+                    for i, doc_id in enumerate(result.metadata["document_ids"]):
+                        text = result.metadata["chunks"][i] if "chunks" in result.metadata else None
+                        score = result.metadata["scores"][i] if "scores" in result.metadata else None
+                        message.results.append(
+                            {
+                                "file_id": doc_id,
+                                "filename": doc_id,
+                                "text": text,
+                                "score": score,
+                            }
+                        )
+                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")
 

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

@@ -170,6 +170,8 @@ class MemoryToolRuntimeImpl(ToolGroupsProtocolPrivate, ToolRuntime, RAGToolRunti
             content=picked,
             metadata={
                 "document_ids": [c.metadata["document_id"] for c in chunks[: len(picked)]],
+                "chunks": [c.content for c in chunks[: len(picked)]],
+                "scores": scores[: len(picked)],
             },
         )