We are excited to announce the release of pgvector version 0.6.0. This update introduces parallel index builds for HNSW indexes, resulting in significant speed improvements. With this new feature, building indexes for unlogged tables can be up to 30 times faster!
HNSW, or Hierarchical Navigable Small World, is an algorithm specifically designed for approximate nearest neighbor search. It operates over multiple layers with different densities or distances between nodes, improving search accuracy and performance.
The addition of parallel processing for index builds drastically reduces the time taken to build indexes, especially for large datasets. By utilizing multiple threads or processes, the workload is distributed, allowing for faster completion times.
Additionally, the use of unlogged tables further enhances the building process. By skipping the Write-Ahead Log (WAL), which is typically used for crash recovery, building times are further optimized.
In addition to the performance enhancements, pgvector version 0.6.0 also brings improvements in memory efficiency and reduces the generation of WAL during the index building process.
Experience the power of pgvector 0.6.0 and take advantage of faster index builds with parallel processing. Upgrade now to enjoy the benefits of improved search accuracy, performance, and reduced building times.

Collections

pgvector version 0.6.0 introduces parallel index builds for HNSW indexes, resulting in significant speed improvements. The use of unlogged tables further enhances the building process. Experience the benefits of improved search accuracy, performance, and reduced building times.

pgvector 0.6.0: Faster Index Builds with Parallel Processing