This PostgreSQL extension provides Hierarchical Navigable Small World (HNSW) indexing for efficient vector similarity search. It's designed for developers and data scientists working with high-dimensional vector data within PostgreSQL, enabling fast nearest neighbor searches for applications like recommendation systems and semantic search.

How It Works

The extension implements the HNSW algorithm, a graph-based approach for approximate nearest neighbor search. It constructs a multi-layered graph where each layer is a subset of the previous one. Searches start at the top layer and navigate downwards, efficiently converging on the most similar vectors. This method offers a strong balance between search speed and accuracy, outperforming brute-force methods.

Quick Start & Requirements

• Install via CREATE EXTENSION embedding; within a PostgreSQL database.
• Requires PostgreSQL.
• Supports Euclidean (L2), Cosine, and Manhattan distance metrics.
• Index creation requires specifying dims, m, efconstruction, and efsearch parameters.
• See official Neon documentation for detailed usage and migration: Neon Documentation

Highlighted Details

• Implements HNSW based on the ivf-hnsw library.
• Supports Euclidean, Cosine, and Manhattan distance metrics.
• Allows tuning of HNSW parameters (m, efconstruction, efsearch) for performance trade-offs.
• Enables creating HNSW indexes on vector columns for accelerated similarity searches.

Maintenance & Community

Neon is no longer committing to pg_embedding as of September 29, 2023. Support is maintained for existing users, with a strong recommendation to migrate to pgvector. Migration instructions are available in the Neon documentation.

Licensing & Compatibility

The extension's license is not explicitly stated in the README. Compatibility with commercial or closed-source applications should be verified.

Limitations & Caveats

Neon has ceased active development and recommends migration to pgvector. While existing users will be supported, new development and feature enhancements are unlikely.