3,096 stars · PostgreSQL License (read from /blob/main/LICENSE, 17 lines, plain apart from the holder, which reads 'Tiger Data' — Timescale, Inc. d/b/a Tiger Data; a separate NOTICE file sits beside it pointing at the same licence. NOT Apache-2.0 and NOT Timescale's source-available TSL) · 0.9.1 (4 Sep; the year 2026 is INFERRED, not registry-confirmed — crates.io returns 404 for the name and no package registry publishes this project, so the year rests on the build targeting PostgreSQL 18, which did not exist before September 2025) · Track this in Scout
A PostgreSQL extension that makes similarity search work on more data than fits in memory.
▶Repo detailsthe review · specs · pros & cons · install
What it is
Pgvectorscale is an extension for PostgreSQL, which means a piece of code you load into the database itself. It adds an index based on an approach called DiskANN that is designed to live on disk. It works beside pgvector, the base extension that gives PostgreSQL the ability to store those lists of numbers at all.
What it is good for. Anyone whose search-by-meaning feature has grown past what a reasonably priced server can hold in memory. The alternative is a second, separate database just for this, which is another thing to run, back up and keep in step. This keeps it in the database that already holds the rest of the data.
- ⚠ Its licence is the PostgreSQL License, read from the file: short, permissive and with no conditions worth worrying about. That matters because the company behind it also sells a source-available product, and this is not it.
- It keeps one database rather than two. Your text and your numbers stay in the same place, in the same backup, inside the same transaction.
- A small waiting list — 14 problems and 9 changes — which for a database extension is a good sign rather than a quiet one.
- It is not something you run. You must already have a working PostgreSQL server with pgvector installed, and building from source wants Rust, a build helper called
cargo-pgrxpinned to one exact version, and thejqtool. - ⚠ Index building needs the database's memory setting raised by hand. The project's own instruction is
maintenance_work_mem = '2GB', and without it a real dataset either crawls or fails. - Intel-based Mac computers are explicitly not supported, so some laptops can only use it through a container. The copyright holder in the licence file reads "Tiger Data", which is the company's new name for itself and does not match the repository owner.
pgvector/pgvectorThe same job of searching by meaning inside PostgreSQL, and in fact the thing pgvectorscale depends on; the difference is that its indexes are built to live in memory, which is the limit pgvectorscale exists to push past.
Track this in Scout- supervc-stack/VectorChord
Also a disk-friendly similarity index for PostgreSQL by a different method, and the successor to an earlier project by the same group, but dual-licensed AGPL-3.0 and Elastic License v2, so not permissive; renamed from tensorchord/VectorChord.
Track this in Scout - qdrant/qdrant
The same search job done quickly, but as a separate database you run alongside PostgreSQL rather than inside it.
Track this in Scout
CREATE EXTENSION IF NOT EXISTS vectorscale CASCADE;
