mxngjxa/lshrs

Separates vector storage from LSH bucket membership by storing only band signatures in Redis while keeping heavy embeddings in your primary database, enabling memory-efficient scaling. Supports streaming ingestion from PostgreSQL or Parquet with configurable band/row parameters for target similarity thresholds, and exposes dual retrieval modes—fast collision-based top-k or cosine-reranked top-p filtering with optional vector reranking from your data store.

Available on PyPI.