pg.ddx.io  pgvector/pgvector GitHub issues and pull requests  
help / color / mirror / Atom feed
From: IgnazioDS (@IgnazioDS) <noreply+IgnazioDS@github.com>
To: pgvector/pgvector <noreply+pgvector-pgvector@github.com>
Subject: Re: [pgvector/pgvector] issue #761: Enhancement: multicolumn vector index design
Date: Tue, 26 May 2026 10:08:48 +0000
Message-ID: <gh-pgvector-pgvector-761-c4543210637@github.com> (raw)
In-Reply-To: <gh-pgvector-pgvector-761@github.com>
References: <gh-pgvector-pgvector-761@github.com>

Following this thread with interest from the application side. Building hybrid retrieval (pgvector + BM25) on top of pgvector, the metadata-filter case is consistently where the gap shows up versus dedicated vector DBs.

A clarifying question for @VladlenPopolitov on the selectivity charts: at the 0.1% selectivity case where mc_hnsw lands in the top-right, how does it behave when the filter column has high cardinality but low correlation with the vector clusters (e.g., a `tenant_id` in a multi-tenant SaaS where tenants are distributed roughly uniformly across the embedding space)? That is the workload that breaks both naive post-filtering and ACORN-style approaches, and it would help calibrate where mc_hnsw sits.

Also +1 to re-opening the discussion from #259 alongside the relaxed option. They solve different problems and users would benefit from both being on the menu.

view thread (18+ messages)  latest in thread

Message-ID: <gh-pgvector-pgvector-761-c4543210637@github.com>
Permalink:  ../gh-pgvector-pgvector-761-c4543210637@github.com/

 · 


This inbox is served by DDX for PostgreSQL; see mirroring instructions
for how to clone and mirror all data and code used for this inbox