refactor: decouple file_search from legacy knowledge_search tool_groups (#5175)

## Summary

Decouples the Responses API's `file_search` tool from the legacy
`knowledge_search` tool_groups infrastructure.

**Before**: When a user passed `{"type": "file_search"}`, the code
looked up `knowledge_search` from the `builtin::rag` tool group registry
and presented it to the LLM under that name. This required
`builtin::rag` to be registered at startup and showed `knowledge_search`
in tracing spans.

**After**: The Responses API creates an inline tool definition with the
name `file_search` — matching the OpenAI convention. No dependency on
`builtin::rag` for the Responses API path. Backward compatibility
preserved: `knowledge_search` still works if used directly.

### Changes

**`streaming.py`**:
- `file_search` tool type now creates an inline
`ChatCompletionToolParam` with name `"file_search"` instead of looking
up `knowledge_search` from tool_groups
- Added `"file_search"` to `_SERVER_SIDE_BUILTIN_TOOL_NAMES`
- Output item events match both `"knowledge_search"` and `"file_search"`

**`tool_executor.py`**:
- All 6 places that checked `== "knowledge_search"` now use `in
("knowledge_search", "file_search")`
- Tracer span uses `function_name` dynamically instead of hardcoded
`"knowledge_search"`

**Recordings**: 44 new recordings for file_search tests with the updated
tool name.

### Why

- Removes unnecessary coupling to `builtin::rag` tool group registration
- LLM now sees `file_search` (matching OpenAI convention) instead of the
non-standard `knowledge_search`
- Tracing spans correctly show `file_search`
- Related to #3901 (closed without merge — this is a more surgical
approach)

## Test Plan

- All 144 unit tests in `tests/unit/providers/agents/builtin/` pass
- Integration test recordings re-generated against OpenAI

---------

Signed-off-by: Sébastien Han <seb@redhat.com>
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Co-authored-by: mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>

Author: 3 people

src/llama_stack/providers/inline/agents/builtin/responses/streaming.py

@@ -114,7 +114,7 @@
 
 # Built-in tool names that the server knows how to execute itself.
 # Anything else is either a registered function tool (client-side) or a hallucinated name.
-_SERVER_SIDE_BUILTIN_TOOL_NAMES = frozenset({"web_search", "knowledge_search"})
+_SERVER_SIDE_BUILTIN_TOOL_NAMES = frozenset({"web_search", "knowledge_search", "file_search"})
 
 # Maps OpenAI Chat Completions error codes to Responses API error codes
 _RESPONSES_API_ERROR_CODES = {
@@ -1299,7 +1299,7 @@ async def _coordinate_tool_execution(
                     id=matching_item_id,
                     status="in_progress",
                 )
-            elif tool_call.function.name == "knowledge_search":
+            elif tool_call.function.name in ("knowledge_search", "file_search"):
                 item = OpenAIResponseOutputMessageFileSearchToolCall(
                     id=matching_item_id,
                     status="in_progress",
@@ -1417,11 +1417,22 @@ def make_openai_tool(tool_name: str, tool: ToolDef) -> ChatCompletionToolParam:
                     raise ValueError(f"Tool {tool_name} not found")
                 self.ctx.chat_tools.append(make_openai_tool(tool_name, tool))
             elif input_tool.type == "file_search":
-                tool_name = "knowledge_search"
-                tool = await self.tool_executor.tool_groups_api.get_tool(tool_name)
-                if not tool:
-                    raise ValueError(f"Tool {tool_name} not found")
-                self.ctx.chat_tools.append(make_openai_tool(tool_name, tool))
+                tool_name = "file_search"
+                file_search_tool_def = ToolDef(
+                    name=tool_name,
+                    description="Search files for relevant information",
+                    input_schema={
+                        "type": "object",
+                        "properties": {
+                            "query": {
+                                "type": "string",
+                                "description": "The search query",
+                            },
+                        },
+                        "required": ["query"],
+                    },
+                )
+                self.ctx.chat_tools.append(make_openai_tool(tool_name, file_search_tool_def))
             elif input_tool.type == "mcp":
                 async for stream_event in self._process_mcp_tool(input_tool, output_messages):
                     yield stream_event

src/llama_stack/providers/inline/agents/builtin/responses/tool_executor.py

@@ -275,7 +275,7 @@ async def _emit_progress_events(
                 ),
                 sequence_number=sequence_number,
             )
-        elif function_name == "knowledge_search":
+        elif function_name in ("knowledge_search", "file_search"):
             sequence_number += 1
             yield ToolExecutionResult(
                 stream_event=OpenAIResponseObjectStreamResponseFileSearchCallInProgress(
@@ -299,7 +299,7 @@ async def _emit_progress_events(
             )
 
         # For file search, emit searching event
-        if function_name == "knowledge_search":
+        if function_name in ("knowledge_search", "file_search"):
             sequence_number += 1
             yield ToolExecutionResult(
                 stream_event=OpenAIResponseObjectStreamResponseFileSearchCallSearching(
@@ -343,7 +343,7 @@ async def _execute_tool(
                         authorization=mcp_tool.authorization,
                         session_manager=self.mcp_session_manager,
                     )
-            elif function_name == "knowledge_search":
+            elif function_name in ("knowledge_search", "file_search"):
                 response_file_search_tool = (
                     next(
                         (t for t in ctx.response_tools if isinstance(t, OpenAIResponseInputToolFileSearch)),
@@ -356,7 +356,7 @@ async def _execute_tool(
                     # Use vector_stores.search API instead of knowledge_search tool
                     # to support filters and ranking_options
                     query = tool_kwargs.get("query", "")
-                    with tracer.start_as_current_span("knowledge_search"):
+                    with tracer.start_as_current_span(function_name):
                         result = await self._execute_knowledge_search_via_vector_store(
                             query=query,
                             response_file_search_tool=response_file_search_tool,
@@ -408,7 +408,7 @@ async def _emit_completion_events(
                 sequence_number=sequence_number,
             )
             yield ToolExecutionResult(stream_event=web_completion_event, sequence_number=sequence_number)
-        elif function_name == "knowledge_search":
+        elif function_name in ("knowledge_search", "file_search"):
             sequence_number += 1
             file_completion_event = OpenAIResponseObjectStreamResponseFileSearchCallCompleted(
                 item_id=item_id,
@@ -465,7 +465,7 @@ async def _build_result_messages(
                 )
                 if has_error:
                     message.status = "failed"
-            elif function.name == "knowledge_search":
+            elif function.name in ("knowledge_search", "file_search"):
                 message = OpenAIResponseOutputMessageFileSearchToolCall(
                     id=item_id,
                     queries=[tool_kwargs.get("query", "")],