Support metadata-only update operations in PGVectorStore (design discussion)

[yukiharada1228]

Support metadata-only update operations in PGVectorStore (design discussion)

Summary

I’d like to start a discussion about whether PGVectorStore should support metadata-only update operations without re-embedding.

Currently, updating metadata requires deleting and re-adding documents, which recomputes embeddings even when the content itself has not changed.

Before working on a PR, I would like to better understand whether this fits within the intended abstraction of PGVectorStore.

---

Motivation

There are several practical cases where only metadata changes:

• Title or label updates
• Status transitions (draft → published, active → archived)
• Metadata enrichment after indexing
• Corrections of metadata errors

At the moment, the typical workaround is:

1. Query documents
2. Delete them
3. Re-add them with updated metadata

This approach:

• Recomputes embeddings unnecessarily
• Can be costly when using paid embedding APIs
• Encourages users to bypass the abstraction and run raw SQL directly

For example:

cursor.execute(
    """
    UPDATE "my_table"
    SET langchain_metadata = jsonb_set(
        COALESCE(langchain_metadata::jsonb, '{}'::jsonb),
        '{video_title}',
        to_jsonb(%s::text)
    )
    WHERE video_id = %s
    """,
    [new_title, video_id],
)

This breaks the abstraction boundary and tightly couples applications to the internal schema.

---

Proposal (for discussion)

One possible API could be:

vectorstore.update(
    filter={"video_id": 123},
    metadata={"video_title": "New Title"},
)

Or:

vectorstore.update(
    ids=["doc-id-1"],
    metadata={"status": "archived"},
)

The method would:

• Only modify metadata
• Not touch embeddings
• Return the number of updated rows

---

Design Questions

I would appreciate feedback on the following:

1. Does supporting metadata-only updates align with the intended design philosophy of PGVectorStore?
2. Should update semantics be strictly limited to metadata?
3. Is there an architectural reason this capability was intentionally omitted?

If this direction is considered appropriate, I would be happy to explore a PR.

[yukiharada1228]

After reconsidering the API design, I would like to clarify the scope.

Using a generic update() method may introduce ambiguity, as it could imply updating document content (and embeddings) and may create inconsistencies with other vector store backends that do not support partial updates.

For this reason, I am leaning toward introducing a PGVectorStore-specific method such as:

vectorstore.update_metadata(
    ids=[...],
    values={...},
)

This would clearly limit the operation to metadata-only changes, avoid affecting embeddings, and preserve the simplicity of the base VectorStore abstraction.

I would appreciate feedback on whether this narrower, backend-specific approach would be more appropriate.