Improving Recall When Filtering on very small subset

[ggalpra]

Hi,

I'm encountering an issue with my PostgreSQL + PGVector setup:

I have a table containing vectors and a category_id, and I always need to filter on a specific category_id. However, since I have hundreds of different categories and the filtering is applied after the index scan, it leads to very low recall in my queries.

I understand that partitioning is recommended in such cases, but my challenge is that new category_id values are frequently added, and I need efficient indexing on them immediately for performance reasons.

I'm using Django, and due to external constraints, I'm stuck on PostgreSQL 16.3, meaning I can't upgrade pgvector to 0.8.0 to leverage iterative scanning.

Has anyone faced a similar issue? How did you manage to improve recall in this scenario?

Thanks!

[ankane]

Hi @ggalpra, check out the filtering docs for a list of options. I'd start with a B-tree index on category_id. For iterative scanning, pgvector 0.8.0 supports Postgres 13+, and for partitioning, you can use hash partitioning if categories are frequently added.