pgvectorscale enhances PostgreSQL's pgvector extension with Rust-based implementations for high-performance, cost-efficient vector search. It targets AI applications and developers needing to scale vector similarity search, offering significant improvements over existing solutions like Pinecone in benchmarks.

How It Works

pgvectorscale introduces a new StreamingDiskANN index type, inspired by Microsoft's DiskANN research. It also incorporates Statistical Binary Quantization (SBQ) for improved vector compression and label-based filtering for efficient, precise searches. This Rust implementation, built with the PGRX framework, allows for deeper integration and optimization within PostgreSQL.

Quick Start & Requirements

• Installation: Pre-built Docker containers, installation from source, or via Timescale Cloud.
• Prerequisites: PostgreSQL, Rust (for source builds), cargo-pgrx. macOS X86 (Intel) builds are not supported.
• Setup: Docker installation is straightforward. Source builds require Rust and cargo-pgrx setup.
• Docs: pgvectorscale Documentation

Highlighted Details

• Achieves 28x lower p95 latency and 16x higher query throughput than Pinecone's storage-optimized index on a 50M embedding dataset.
• Offers label-based filtering for optimized metadata searches.
• Supports cosine, L2, and inner product distance queries.
• Allows tuning of index build-time and query-time parameters for accuracy/performance trade-offs.

Maintenance & Community

• Active development by Timescale.
• Community contributions are welcomed.
• Discord for community engagement.

Licensing & Compatibility

• Apache 2.0 License.
• Compatible with PostgreSQL and pgvector.

Limitations & Caveats

• Building on macOS X86 (Intel) is not supported.
• Indices on UNLOGGED tables are not yet implemented.
• diskann index uses relaxed ordering; strict ordering requires a materialized CTE.