Add pdf support to file_search for Responses API

This adds basic PDF support (using our existing `parse_pdf` function)
to the file_search tool and corresponding Vector Files API.

When a PDF file is uploaded and attached to a vector store, we parse
the pdf and then chunk its content as normal. This is not the best
solution long-term, but it does match what we've been doing so far for
PDF files in the memory tool.

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

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

@@ -9,6 +9,7 @@ import base64
 import io
 import json
 import logging
+import mimetypes
 import time
 from typing import Any
 
@@ -19,7 +20,6 @@ from numpy.typing import NDArray
 from llama_stack.apis.files import Files
 from llama_stack.apis.inference import InterleavedContent
 from llama_stack.apis.inference.inference import Inference
-from llama_stack.apis.tools.rag_tool import RAGDocument
 from llama_stack.apis.vector_dbs import VectorDB
 from llama_stack.apis.vector_io import (
     Chunk,
@@ -40,8 +40,8 @@ from llama_stack.providers.utils.memory.openai_vector_store_mixin import OpenAIV
 from llama_stack.providers.utils.memory.vector_store import (
     EmbeddingIndex,
     VectorDBWithIndex,
-    content_from_doc,
     make_overlapped_chunks,
+    parse_pdf,
 )
 
 from .config import FaissVectorIOConfig
@@ -292,20 +292,23 @@ class FaissVectorIOAdapter(OpenAIVectorStoreMixin, VectorIO, VectorDBsProtocolPr
             chunk_overlap_tokens = 400
 
         try:
+            file_response = await self.files_api.openai_retrieve_file(file_id)
+            mime_type, _ = mimetypes.guess_type(file_response.filename)
             content_response = await self.files_api.openai_retrieve_file_content(file_id)
-            content = content_response.body
-            doc = RAGDocument(
-                document_id=file_id,
-                content=content,
-                metadata=attributes,
-            )
-            content = await content_from_doc(doc)
+
+            # TODO: We can't use content_from_doc directly from vector_store
+            # but should figure out how to centralize this logic near there
+            if mime_type == "application/pdf":
+                content = parse_pdf(content_response.body)
+            else:
+                content = content_response.body.decode("utf-8")
+
             chunks = make_overlapped_chunks(
-                doc.document_id,
+                file_id,
                 content,
                 max_chunk_size_tokens,
                 chunk_overlap_tokens,
-                doc.metadata,
+                attributes,
             )
 
             if not chunks:

tests/verifications/openai_api/fixtures/test_cases/responses.yaml

@@ -39,7 +39,15 @@ test_response_file_search:
       input: "How many experts does the Llama 4 Maverick model have?"
       tools:
       - type: file_search
-        # vector_store_ids gets added by the test runner
+        # vector_store_ids param for file_search tool gets added by the test runner
+      file_content: "Llama 4 Maverick has 128 experts"
+      output: "128"
+    - case_id: "What is the "
+      input: "How many experts does the Llama 4 Maverick model have?"
+      tools:
+      - type: file_search
+        # vector_store_ids param for file_search toolgets added by the test runner
+      file_path: "pdfs/llama_stack_and_models.pdf"
       output: "128"
 
 test_response_mcp_tool:

tests/verifications/openai_api/test_responses.py

@@ -5,6 +5,7 @@
 # the root directory of this source tree.
 
 import json
+import os
 import time
 
 import httpx
@@ -38,7 +39,7 @@ def _new_vector_store(openai_client, name):
     return vector_store
 
 
-def _new_file(openai_client, name, content, tmp_path):
+def _upload_file(openai_client, name, file_path):
     # Ensure we don't reuse an existing file
     files = openai_client.files.list()
     for file in files:
@@ -46,8 +47,6 @@ def _new_file(openai_client, name, content, tmp_path):
             openai_client.files.delete(file_id=file.id)
 
     # Upload a text file with our document content
-    file_path = tmp_path / name
-    file_path.write_text(content)
     return openai_client.files.create(file=open(file_path, "rb"), purpose="assistants")
 
 
@@ -291,7 +290,7 @@ def test_response_non_streaming_web_search(request, openai_client, model, provid
     responses_test_cases["test_response_file_search"]["test_params"]["case"],
     ids=case_id_generator,
 )
-def test_response_non_streaming_file_search_simple_text(
+def test_response_non_streaming_file_search(
     request, openai_client, model, provider, verification_config, tmp_path, case
 ):
     if isinstance(openai_client, LlamaStackAsLibraryClient):
@@ -303,8 +302,17 @@ def test_response_non_streaming_file_search_simple_text(
 
     vector_store = _new_vector_store(openai_client, "test_vector_store")
 
-    file_content = "Llama 4 Maverick has 128 experts"
-    file_response = _new_file(openai_client, "test_response_non_streaming_file_search.txt", file_content, tmp_path)
+    if "file_content" in case:
+        file_name = "test_response_non_streaming_file_search.txt"
+        file_path = tmp_path / file_name
+        file_path.write_text(case["file_content"])
+    elif "file_path" in case:
+        file_path = os.path.join(os.path.dirname(__file__), "fixtures", case["file_path"])
+        file_name = os.path.basename(file_path)
+    else:
+        raise ValueError(f"No file content or path provided for case {case['case_id']}")
+
+    file_response = _upload_file(openai_client, file_name, file_path)
 
     # Attach our file to the vector store
     file_attach_response = openai_client.vector_stores.files.create(
@@ -343,7 +351,7 @@ def test_response_non_streaming_file_search_simple_text(
     assert response.output[0].status == "completed"
     assert response.output[0].queries  # ensure it's some non-empty list
     assert response.output[0].results
-    assert response.output[0].results[0].text == file_content
+    assert case["output"].lower() in response.output[0].results[0].text.lower()
     assert response.output[0].results[0].score > 0
 
     # Verify the assistant response that summarizes the results
@@ -354,13 +362,8 @@ def test_response_non_streaming_file_search_simple_text(
     assert case["output"].lower() in response.output_text.lower().strip()
 
 
-@pytest.mark.parametrize(
-    "case",
-    responses_test_cases["test_response_file_search"]["test_params"]["case"],
-    ids=case_id_generator,
-)
 def test_response_non_streaming_file_search_empty_vector_store(
-    request, openai_client, model, provider, verification_config, tmp_path, case
+    request, openai_client, model, provider, verification_config
 ):
     if isinstance(openai_client, LlamaStackAsLibraryClient):
         pytest.skip("Responses API file search is not yet supported in library client.")
@@ -371,17 +374,11 @@ def test_response_non_streaming_file_search_empty_vector_store(
 
     vector_store = _new_vector_store(openai_client, "test_vector_store")
 
-    # Update our tools with the right vector store id
-    tools = case["tools"]
-    for tool in tools:
-        if tool["type"] == "file_search":
-            tool["vector_store_ids"] = [vector_store.id]
-
     # Create the response request, which should query our vector store
     response = openai_client.responses.create(
         model=model,
-        input=case["input"],
-        tools=case["tools"],
+        input="How many experts does the Llama 4 Maverick model have?",
+        tools=[{"type": "file_search", "vector_store_ids": [vector_store.id]}],
         stream=False,
         include=["file_search_call.results"],
     )