pgvectorscale

pgvectorscale is an open-source PostgreSQL extension from Timescale that complements pgvector with higher-performance embedding search and more cost-efficient storage for AI applications. It adds a StreamingDiskANN index, Statistical Binary Quantization compression, and label-based filtered vector search. It is written in Rust using the PGRX framework and released under the PostgreSQL License.

What It Is

pgvectorscale is a vector search extension for PostgreSQL that works alongside pgvector. Application developers and DBAs create a table with a vector column, then build a diskann index on it and query with the same distance operators pgvector uses: cosine (<=>), L2 (<->), and inner product (<#>), each matched to its operator class.

The README states that on a benchmark of 50 million Cohere embeddings with 768 dimensions, PostgreSQL with pgvector and pgvectorscale achieved 28x lower p95 latency and 16x higher query throughput than Pinecone's storage-optimized (s1) index at 99% recall, at 75% less cost when self-hosted on AWS EC2.

Index and Filtering Model

The StreamingDiskANN index is inspired by Microsoft's DiskANN research. By default it uses the memory_optimized storage layout, which compresses vectors with Statistical Binary Quantization; a plain layout stores them uncompressed. Label-based filtering takes a smallint[] labels column in the index definition and uses the && overlap operator for filtered queries. Arbitrary WHERE clauses are also supported, but they are applied as post-filtering after the vector search.

Tuning and Behavior Notes

Build-time parameters include num_neighbors, search_list_size, max_alpha, num_dimensions (useful for Matryoshka embeddings), and num_bits_per_dimension. Parallel index builds are supported under documented conditions, for example when the table has no labels and SBQ storage is used. At query time, diskann.query_search_list_size and diskann.query_rescore trade accuracy against speed.

Documented limitations:

• Results can be slightly out of distance order (relaxed ordering), and a materialized CTE restores strict ordering.
• Null vectors are not indexed.
• Indexes on UNLOGGED tables are not supported.
• Building from source on Intel (x86) macOS is not supported.

Setup Path

You can run it from a pre-built TimescaleDB Docker image, build from source with Rust and cargo-pgrx, or enable it in a Timescale Cloud service. In each case, CREATE EXTENSION IF NOT EXISTS vectorscale CASCADE; also installs pgvector.