Skip to content
AI News HubLIVE
More
Summary only

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 built LEANN, an open-source vector index that stores almost no embeddings and recomputes them at query time. Benchmarks show a 60M-chunk Wikipedia index cut from 201 GB to 6 GB, matching HNSW downstream accuracy at 90% recall at the cost of seconds-level search.

SourceMarkTechPostAuthor: Asif Razzaq
Meet LEANN: An Open-Source Vector Index that Shrinks a 60M-Chunk RAG Index from 201 GB to 6 GB
Report an error

The correction channel is not available yet. You can copy the article reference below for later.

Correction instructions

Summary only

Under the source content permission, this page provides an editorial summary, key points, and a link to the original article.

Key points and analysis

Article intelligence

EngineersAdvanced

Key points

  • LEANN keeps a pruned graph and recomputes embeddings on demand, shrinking a 60M-chunk index from 201 GB to 6 GB (4 GB vs 188 GB in the paper’s 76 GB corpus).
  • It matches HNSW’s 25.5% downstream accuracy on NQ at 90% recall, but retrieval takes 2.48 s versus 0.05 s for HNSW.
  • Two-level search, dynamic batching, and high-degree preserving pruning make recomputation practical, though it still accounts for about 76% of query latency.
  • The code is MIT licensed and ships ready apps for documents, mail, browser history, and more, plus an MCP semantic search server.

Highlights and analysis are generated automatically and may contain errors. Check the original source.

More recent stories on this topic · Models