PGVector

PGVector is an open-source PostgreSQL extension that enables efficient storage, indexing, and similarity search for vector embeddings directly within a Postgres database. It supports both exact and approximate nearest neighbor search, and a variety of vector types and distance metrics.

Features

• Vector Storage: Store vectors alongside other data in Postgres tables.
  • Supports single-precision, half-precision (halfvec), binary (bit), and sparse (sparsevec) vectors.
  • Can store vectors with up to 16,000 dimensions (vector), 4,000 (halfvec), 64,000 (bit), or 1,000 non-zero elements (sparsevec).
• Similarity Search:
  • Exact and approximate nearest neighbor search.
  • Multiple distance metrics: L2 (Euclidean), inner product, cosine, L1 (taxicab), Hamming (for binary), and Jaccard (for binary).
• Indexing:
  • Exact search by default.
  • Approximate search via HNSW and IVFFlat indexes for faster queries at the cost of some recall.
  • Index tuning options for memory, parallelism, and recall/speed tradeoffs.
  • Support for filtering and iterative index scans to improve recall when filtering.
• Querying:
  • Retrieve nearest neighbors to a vector or another row.
  • Query by distance threshold.
  • Aggregate functions: average, sum of vectors.
  • Subvector indexing and querying.
• Hybrid Search:
  • Combine with Postgres full-text search for hybrid queries.
• Language Support:
  • Usable from any language with a Postgres client.
  • Libraries/examples available for C, C++, C#, Go, Java, JavaScript/TypeScript, Python, Ruby, Rust, and more.
• Postgres Integration:
  • ACID compliance, point-in-time recovery, JOINs, and all standard Postgres features.
  • Supports replication through WAL.
  • Scalable with standard Postgres scaling techniques (vertical and horizontal).
• Performance and Tuning:
  • Bulk loading with COPY.
  • Index build and query performance tuning.
  • Monitoring with pg_stat_statements.
  • Options for concurrent index creation, parallelism, and memory usage.
• Installation:
  • Available via compiling from source, Docker, Homebrew, PGXN, APT, Yum, pkg, conda-forge, and preinstalled on some hosted Postgres providers.
  • Works on Linux, macOS, and Windows.
• Data Flexibility:
  • Store vectors with different dimensions using untyped columns (with some limitations on indexing).
  • Support for higher-precision vectors using Postgres array types.
• Functions and Operators:
  • Rich set of operators for vector arithmetic and similarity.
  • Functions for distance calculation, normalization, quantization, and more.
• Open Source:
  • Licensed under an open-source license, available for community contributions.

Pricing

PGVector is open-source and free to use.

Links

• Source code and documentation