fix: Pydantic validation error with list-type metadata in vector search (#3797)

src/llama_stack/providers/utils/memory/openai_vector_store_mixin.py

@@ -23,6 +23,9 @@
     content_from_data_and_mime_type,
     make_overlapped_chunks,
 )
+from llama_stack.providers.utils.vector_io.vector_utils import (
+    sanitize_metadata_for_attributes,
+)
 from llama_stack_api import (
     Chunk,
     Files,
@@ -635,7 +638,7 @@ async def openai_search_vector_store(
                     file_id=chunk.metadata.get("document_id", ""),
                     filename=chunk.metadata.get("filename", ""),
                     score=score,
-                    attributes=chunk.metadata,
+                    attributes=sanitize_metadata_for_attributes(chunk.metadata),
                     content=content,
                 )
                 data.append(response_data_item)

src/llama_stack/providers/utils/vector_io/vector_utils.py

@@ -7,6 +7,7 @@
 import hashlib
 import re
 import uuid
+from typing import Any
 
 
 def generate_chunk_id(document_id: str, chunk_text: str, chunk_window: str | None = None) -> str:
@@ -37,6 +38,28 @@ def sanitize_collection_name(name: str, weaviate_format=False) -> str:
     return s
 
 
+def sanitize_metadata_for_attributes(metadata: dict[str, Any]) -> dict[str, str | float | bool]:
+    """
+    Filter metadata to primitives for VectorStoreSearchResponse.attributes compatibility.
+
+    Converts dict[str, Any] to dict[str, str | float | bool]:
+    - Preserves: str, bool
+    - Converts: int/float -> float, list -> comma-separated string
+    - Filters: dict, None, other types
+    """
+    sanitized: dict[str, str | float | bool] = {}
+    for key, value in metadata.items():
+        if isinstance(value, bool):
+            sanitized[key] = value
+        elif isinstance(value, int | float):
+            sanitized[key] = float(value)
+        elif isinstance(value, str):
+            sanitized[key] = value
+        elif isinstance(value, list):
+            sanitized[key] = ", ".join(str(item) for item in value)
+    return sanitized
+
+
 class WeightedInMemoryAggregator:
     @staticmethod
     def _normalize_scores(scores: dict[str, float]) -> dict[str, float]:

tests/unit/providers/vector_io/test_vector_utils.py

@@ -4,7 +4,10 @@
 # This source code is licensed under the terms described in the LICENSE file in
 # the root directory of this source tree.
 
-from llama_stack.providers.utils.vector_io.vector_utils import generate_chunk_id
+from llama_stack.providers.utils.vector_io.vector_utils import (
+    generate_chunk_id,
+    sanitize_metadata_for_attributes,
+)
 from llama_stack_api import Chunk, ChunkMetadata
 
 # This test is a unit test for the chunk_utils.py helpers. This should only contain
@@ -78,3 +81,27 @@ def test_chunk_serialization():
     serialized_chunk = chunk.model_dump()
     assert serialized_chunk["chunk_id"] == "test-chunk-id"
     assert "chunk_id" in serialized_chunk
+
+
+def test_sanitize_metadata_for_attributes():
+    """Test sanitization of metadata for VectorStoreSearchResponse.attributes."""
+    # metadata with lists should be converted to strings
+    metadata = {
+        "tags": ["transformers", "h100-compatible", "region:us"],
+        "model_name": "granite-3.3-8b",
+        "score": 0.95,
+        "active": True,
+        "count": 42,
+        "nested": {"key": "value"},  # Should be filtered out
+    }
+    result = sanitize_metadata_for_attributes(metadata)
+
+    # Lists converted to comma-separated strings
+    assert result["tags"] == "transformers, h100-compatible, region:us"
+    # Primitives preserved
+    assert result["model_name"] == "granite-3.3-8b"
+    assert result["score"] == 0.95
+    assert result["active"] is True
+    assert result["count"] == 42.0  # int -> float
+    # Complex types filtered out
+    assert "nested" not in result