VectorChord as a drop in replacement for pgvector

[vaclcer]

Hi, first of all, I am merely a user, not a db guru.. So, I would like to use VectorChord in an external project that supports pgvector natively (its Open WebUI). My question is: is VectorChord a drop in replacement? If so, will I benefit right away without any special configuration (on either or both sides)?

Or do I need to pre-configure it and/or use it specifically (in comparison to pgvector) to really see any benefits along the line?

Thanks!

[usamoi]

We use PostgreSQL data types provided by pgvector, so all SDKs for pgvector works.

The following two changes are required:

• When building the index, you need to specify the index access method vchordrq provided by VectorChord, rather than hnsw or ivfflat provided by pgvector.
• When performing searches, you need to specify GUC parameters such as vchordrq.probes, instead of pgvector's hnsw.ef_search and ivfflat.probes.

Since specifying the index access method and search parameters is already required when using pgvector, this does not introduce much difference.

[vaclcer]

https://www.perplexity.ai/search/0238421a-709a-4b87-92eb-d94e18191e3d

does this make sense? :)

By default, indeces are created as follows:

self.session.execute(
    text(
        "CREATE INDEX IF NOT EXISTS idx_document_chunk_vector "
        "ON document_chunk USING ivfflat (vector vector_cosine_ops) WITH (lists = 100);"
    )
)
self.session.execute(
    text(
        "CREATE INDEX IF NOT EXISTS idx_document_chunk_collection_name "
        "ON document_chunk (collection_name);"
    )
)

[usamoi]

@vaclcer You probably need to change it to:

self.session.execute(
    text(
        "CREATE INDEX IF NOT EXISTS idx_document_chunk_vector "
        "ON document_chunk USING vchordrq (vector vector_cosine_ops) WITH (options = "build.internal.lists = [100]")
    )
)
self.session.execute(
    text(
        "CREATE INDEX IF NOT EXISTS idx_document_chunk_collection_name "
        "ON document_chunk (collection_name);"
    )
)

and then change

SET ivfflat.probes = 1

somewhere else to

SET vchordrq.probes = 1

following https://docs.vectorchord.ai/vectorchord/usage/indexing.html.