LEANN is an open-source vector database that compresses RAG indexes by 97% through graph-based recomputation and on-demand embedding calculation, eliminating stored embeddings.
It builds on HNSW with two-level graph traversal, dynamic batching for low latency, and graph pruning for minimal metadata storage. LEANN achieves 90% recall in under 2 seconds on large datasets while outperforming baselines like Edge-RAG in storage and speed.
1. Introduction In the past few months, we’ve heard consistent feedback from users and partners: while our goal of providing a scalable, high-performa...
