feat(vector-io): implement chunk deletion for PGVector provider

Implement delete_chunk() method with SQL DELETE operation
Add _delete_openai_chunk_from_vector_store() for OpenAI compatibility
Fix chunk_id generation to use actual chunk.chunk_id instead of generated IDs (so they can be removed by id)
Set pgvector provider to requiring Api.files as optional
Update remote_provider_spec() function to support optional_api_dependencies parameter

This allows pgvector to work in configurations without a files provider
while still supporting file operations when files API is available.

Signed-off-by: Derek Higgins <derekh@redhat.com>

llama_stack/providers/datatypes.py

@@ -226,7 +226,10 @@ API responses, specify the adapter here.
 
 
 def remote_provider_spec(
-    api: Api, adapter: AdapterSpec, api_dependencies: list[Api] | None = None
+    api: Api,
+    adapter: AdapterSpec,
+    api_dependencies: list[Api] | None = None,
+    optional_api_dependencies: list[Api] | None = None,
 ) -> RemoteProviderSpec:
     return RemoteProviderSpec(
         api=api,
@@ -234,6 +237,7 @@ def remote_provider_spec(
         config_class=adapter.config_class,
         adapter=adapter,
         api_dependencies=api_dependencies or [],
+        optional_api_dependencies=optional_api_dependencies or [],
     )
 
 

llama_stack/providers/registry/vector_io.py

@@ -410,6 +410,7 @@ See [PGVector's documentation](https://github.com/pgvector/pgvector) for more de
 """,
             ),
             api_dependencies=[Api.inference],
+            optional_api_dependencies=[Api.files],
         ),
         remote_provider_spec(
             Api.vector_io,

llama_stack/providers/remote/vector_io/pgvector/__init__.py

@@ -12,6 +12,6 @@ from .config import PGVectorVectorIOConfig
 async def get_adapter_impl(config: PGVectorVectorIOConfig, deps: dict[Api, ProviderSpec]):
     from .pgvector import PGVectorVectorIOAdapter
 
-    impl = PGVectorVectorIOAdapter(config, deps[Api.inference])
+    impl = PGVectorVectorIOAdapter(config, deps[Api.inference], deps.get(Api.files, None))
     await impl.initialize()
     return impl

llama_stack/providers/remote/vector_io/pgvector/pgvector.py

@@ -99,7 +99,7 @@ class PGVectorIndex(EmbeddingIndex):
         for i, chunk in enumerate(chunks):
             values.append(
                 (
-                    f"{chunk.metadata['document_id']}:chunk-{i}",
+                    f"{chunk.chunk_id}",
                     Json(chunk.model_dump()),
                     embeddings[i].tolist(),
                 )
@@ -159,6 +159,11 @@ class PGVectorIndex(EmbeddingIndex):
         with self.conn.cursor(cursor_factory=psycopg2.extras.DictCursor) as cur:
             cur.execute(f"DROP TABLE IF EXISTS {self.table_name}")
 
+    async def delete_chunk(self, chunk_id: str) -> None:
+        """Remove a chunk from the PostgreSQL table."""
+        with self.conn.cursor(cursor_factory=psycopg2.extras.DictCursor) as cur:
+            cur.execute(f"DELETE FROM {self.table_name} WHERE id = %s", (chunk_id,))
+
 
 class PGVectorVectorIOAdapter(OpenAIVectorStoreMixin, VectorIO, VectorDBsProtocolPrivate):
     def __init__(
@@ -265,3 +270,12 @@ class PGVectorVectorIOAdapter(OpenAIVectorStoreMixin, VectorIO, VectorDBsProtoco
         index = PGVectorIndex(vector_db, vector_db.embedding_dimension, self.conn)
         self.cache[vector_db_id] = VectorDBWithIndex(vector_db, index, self.inference_api)
         return self.cache[vector_db_id]
+
+    async def _delete_openai_chunk_from_vector_store(self, store_id: str, chunk_id: str) -> None:
+        """Delete a chunk from a PostgreSQL vector store."""
+        index = await self._get_and_cache_vector_db_index(store_id)
+        if not index:
+            raise ValueError(f"Vector DB {store_id} not found")
+
+        # Use the index's delete_chunk method
+        await index.index.delete_chunk(chunk_id)