Enable Deletion of Vectors by Metadata in PGVector

[shamspias]

This pull request adds support for deleting vectors based on metadata filters in the PGVector class. It modifies the delete and adelete methods to accept an optional filter parameter, allowing users to specify conditions on metadata fields for deletion.

Changes

• Modified Methods:

  • delete: Added filter: Optional[Dict[str, Any]] = None parameter.
  • adelete: Added filter: Optional[Dict[str, Any]] = None parameter.

• Implementation Details:

  • Utilized the existing _create_filter_clause method to generate SQLAlchemy WHERE clauses from the metadata filters.
  • Ensured backward compatibility; deletion by IDs continues to function as before.
  • Added checks to prevent accidental deletion when neither ids nor filter is provided.

Code Changes

delete Method

def delete(
    self,
    ids: Optional[List[str]] = None,
    *,
    filter: Optional[Dict[str, Any]] = None,
    collection_only: bool = False,
    **kwargs: Any,
) -> None:
    if ids is None and filter is None:
        self.logger.warning("No ids or filter provided for deletion.")
        return

    with self._make_sync_session() as session:
        stmt = delete(self.EmbeddingStore)
        if collection_only:
            collection = self.get_collection(session)
            if not collection:
                self.logger.warning("Collection not found.")
                return
            stmt = stmt.where(self.EmbeddingStore.collection_id == collection.uuid)

        if ids:
            stmt = stmt.where(self.EmbeddingStore.id.in_(ids))

        if filter:
            filter_clause = self._create_filter_clause(filter)
            stmt = stmt.where(filter_clause)

        session.execute(stmt)
        session.commit()

adelete Method

async def adelete(
    self,
    ids: Optional[List[str]] = None,
    *,
    filter: Optional[Dict[str, Any]] = None,
    collection_only: bool = False,
    **kwargs: Any,
) -> None:
    if ids is None and filter is None:
        self.logger.warning("No ids or filter provided for deletion.")
        return

    await self.__apost_init__()
    async with self._make_async_session() as session:
        stmt = delete(self.EmbeddingStore)
        if collection_only:
            collection = await self.aget_collection(session)
            if not collection:
                self.logger.warning("Collection not found.")
                return
            stmt = stmt.where(self.EmbeddingStore.collection_id == collection.uuid)

        if ids:
            stmt = stmt.where(self.EmbeddingStore.id.in_(ids))

        if filter:
            filter_clause = self._create_filter_clause(filter)
            stmt = stmt.where(filter_clause)

        await session.execute(stmt)
        await session.commit()

Usage Example

# Delete vectors where the 'category' metadata field equals 'news'
vector_store.delete(filter={"category": {"$eq": "news"}})

Testing

• Unit Tests:
  • Added tests to verify deletion by metadata filters.
  • Confirmed that vectors matching the filter criteria are deleted.
  • Ensured that deletion by IDs still works as expected.

Documentation

• Updated docstrings for delete and adelete methods to include the filter parameter.
• Provided usage examples demonstrating how to delete by metadata.

Reference Issues:

• 127