LEANN shrinks a 60M-chunk RAG index from 201 GB to 6 GB

UC Berkeley's Sky Computing Lab released LEANN, an open-source vector index that recomputes embeddings at query time instead of storing them. A 60M-chunk Wikipedia index drops from 201 GB to 6 GB while matching HNSW accuracy on NQ (25.5%), though retrieval takes 2.48 s versus 0.05 s. The work won the MLSys 2026 Best Research Paper award and is MIT licensed.

• 60M-chunk index cut from 201 GB to 6 GB
• NQ accuracy 25.5%, matching HNSW at 4 GB vs 188 GB
• NQ retrieval takes 2.48 s versus 0.05 s for HNSW
• MIT licensed, installs via uv pip install leann