pgvectorscale

Postgres shops that outgrow pgvector's HNSW index can add pgvectorscale instead of migrating to a dedicated vector database. The extension, written in Rust with the PGRX framework, layers a StreamingDiskANN index on top of the standard pgvector data type, keeping the graph on disk rather than in RAM so indexes over tens of millions of embeddings stay affordable on ordinary instances. Statistical Binary Quantization, a Timescale-developed refinement of binary quantization, compresses vectors further, and label-filtered search applies the Filtered DiskANN approach so WHERE clauses on tags do not wreck recall; arbitrary predicates fall back to post-filtering. Timescale's benchmark on 50 million 768-dimensional Cohere embeddings measured 28x lower p95 latency and 16x higher query throughput than Pinecone's storage-optimized index at 99 percent recall, at roughly a quarter of the cost self-hosted. Installation is a Docker image, prebuilt packages, or a source build against PostgreSQL versions up to 18, the code ships under the PostgreSQL license, and Timescale Cloud runs it managed.