Meet LEANN: An Open-Source Vector Index that Shrinks a 60M-Chunk RAG Index from 201 GB to 6 GB

UC Berkeley’s LEANN is an open-source vector index that takes a fundamentally different approach to storage: it stores a pruned graph and recomputes embeddings at query time.

How It Performs

According to the report, LEANN matches HNSW accuracy at 90% recall while using 4 GB instead of 188 GB. The trade-off is seconds-level search latency.

Key Numbers at a Glance

• 60M-chunk RAG index reduced from 201 GB to 6 GB
• 90% recall with 4 GB used versus 188 GB for the comparison
• Accuracy that matches HNSW
• Search cost: seconds-level latency

The work comes from UC Berkeley and is described as an open-source vector index. The reported approach trades fast search for dramatic storage savings by storing a pruned graph and recomputing embeddings at query time rather than storing them.

Image: LEANN banner

Source: MarkTechPost