Research27 October 2026, 21:54 UTC4 min read

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

UC Berkeley’s Sky Computing Lab introduced LEANN, an open-source vector index that reduces storage needs by recomputing embeddings at query time instead of storing them. A 60 million-chunk Wikipedia index shrinks from 201 GB to 6 GB in the repo benchmark, and a 76 GB corpus setup drops from 173 GB (embeddings) + 15 GB (graph) to 4 GB in the paper’s tests. The project won the Best Research Paper…

1 source

Key points

LEANN reduces a 60M-chunk Wikipedia index from 201 GB to 6 GB by recomputing embeddings on demand
Matches HNSW accuracy at 90% recall while using 4 GB instead of 188 GB for a 60M-passage datastore
MIT-licensed open-source tool with apps for documents, Apple Mail, and Slack/Twitter data retrieval

LEANN achieves this by storing only a pruned graph and a compact product quantization table. At query time, it re-embeds only nodes on the search path, using techniques like two-level search, dynamic batching, and high-degree preserving pruning. Performance matches HNSW’s downstream accuracy at 90% recall while using under 5% of raw-data storage. Benchmarks show LEANN’s retrieval time ranges from 1.12 seconds (GPQA) to 7.12 seconds (HotpotQA) on an RTX 4090, compared to HNSW’s 0.05 seconds. The MIT-licensed tool runs on laptops and workstations, with apps for documents, Apple Mail, Chrome history, and more.

Full story from MarkTechPost · by Asif RazzaqOpen source ↗