pgvectorscale is an open-source PostgreSQL extension from Timescale that brings high-performance vector search to Postgres, complementing pgvector for workloads that outgrow it. Written in Rust and released under the PostgreSQL license, the project has attracted about 3,150 stars on GitHub and is actively maintained in 2026. It is designed for teams who want to keep embeddings inside their existing Postgres database instead of operating a separate vector store.
The extension implements StreamingDiskANN, a disk-efficient approximate nearest neighbour index inspired by Microsoft’s DiskANN research, together with Statistical Binary Quantization (SBQ) developed by Timescale to compress vectors with minimal accuracy loss. In practice this delivers higher queries-per-second and better recall than pgvector’s indexes at large scale, while using the same vector data types, so existing pgvector tables can adopt it without a data migration. Label-based filtered search lets you combine vector similarity with metadata conditions in one query.
In Timescale’s published benchmarks, the StreamingDiskANN index sustains considerably higher query throughput than pgvector at high recall levels, particularly on datasets with tens of millions of embeddings where in-memory indexes become expensive. The extension is installed like any other Postgres extension and works on self-managed PostgreSQL as well as Timescale’s cloud, and it coexists with pgvector rather than replacing it: pgvector supplies the vector data type and operators, while pgvectorscale supplies the scalable index. That division makes adoption incremental for teams already storing embeddings in Postgres.
pgvectorscale is free and open source under the PostgreSQL license. Timescale also offers managed cloud databases, but the extension itself costs nothing to self-host.
Developers building retrieval-augmented generation, semantic search or recommendation features on PostgreSQL who need vector performance beyond pgvector without adding another database.