// the find
timescale/pgvectorscale
Postgres extension for vector search (DiskANN), complements pgvector for performance and scale. Postgres OSS licensed.
pgvectorscale is a Rust-based Postgres extension that adds a StreamingDiskANN index on top of pgvector, trading some result ordering strictness for dramatically faster ANN search at scale. It's for teams already on Postgres who are hitting pgvector's HNSW performance ceiling at tens of millions of vectors and don't want to operate a separate vector database. Backed by Timescale, so it has real engineering behind it.
The DiskANN-derived index genuinely changes the cost curve — SBQ compression means the index fits in less RAM than HNSW, which matters when your dataset exceeds available memory. Label-based filtered search is built into the index structure (not post-filtered), so you're not paying full scan cost when filtering by category. The upgrade path is well-maintained: there's a SQL migration file for every version pair going back to 0.0.2, which is a sign someone actually thinks about operators running in production. Parallel index builds are automatic for large datasets, configurable for smaller ones.
The relaxed ordering behavior is a footgun: results can come back slightly out of distance order by default, and you need a materialized CTE workaround to get strict ordering — this will silently produce wrong-looking results if you don't read the docs carefully. Label values are constrained to smallint range (-32768 to 32767) and must be stored as integer arrays; any string-based or hierarchical tagging scheme requires a separate lookup table and schema redesign. Parallel builds don't work with label-filtered indexes yet, so your largest workloads with filtering get no parallelism benefit. The macOS x86 build is broken with no fix in sight, which rules out a common developer laptop.