UC Berkeley's Sky Computing Lab has developed LEANN, an open-source vector index designed to significantly reduce storage requirements for Retrieval-Augmented Generation (RAG) systems. LEANN achieves this by recomputing embeddings at query time rather than storing them, a departure from traditional methods.
Key points
- LEANN vector index reduces RAG index size from 201 GB to 6 GB.
- Embeddings are recomputed at query time, not stored on disk.
- Enables semantic search and RAG on personal data via laptops.
- Matches HNSW accuracy using significantly less storage.
- Research paper won Best Research Paper award at MLSys 2026.
The effectiveness of LEANN is demonstrated by its ability to shrink a 60 million chunk Wikipedia index from 201 GB down to just 6 GB. The research paper detailing LEANN also won the Best Research Paper award at MLSys 2026.
LEANN enables semantic search and RAG over personal data, making it feasible to run on a laptop. The index stores a pruned graph and a compact product quantization (PQ) table, rebuilding embeddings on demand for nodes within the search path.
By the numbers
- LEANN index size for 60M chunks
- 6 GB
- Original index size for 60M chunks
- 201 GB
- LEANN index size in paper's setup
- 4 GB
This approach utilizes techniques like two-level search, where cheap PQ distances rank neighbors first, and dynamic batching to group candidates for efficient recomputation. LEANN's performance matches HNSW downstream accuracy at 90% recall, using less than 5% of raw-data storage.
For instance, LEANN achieved 25.5% NQ accuracy, identical to HNSW, but used only 4 GB compared to HNSW's 188 GB. While HotpotQA retrieval on a Mac took 44.80 seconds, this represented 50.1% of the total pipeline time.
Sources and references
Original report
Meet LEANN: An Open-Source Vector Index that Shrinks a 60M-Chunk RAG Index from 201 GB to 6 GB
MarkTechPost
Read on MarkTechPost (opens in a new tab)Linked in the original report
- Linksky.cs.berkeley.edu (opens in a new tab)
- Codegithub.com/StarTrail-org/LEANN (opens in a new tab)
- Paperarxiv.org/abs/2506.08276 (opens in a new tab)
- Papermlsys.org/media/mlsys-2026/best-papers/Best-Resea… (opens in a new tab)
- Paperarxiv.org/html/2506.08276v2 (opens in a new tab)
- Codegithub.com/facebookresearch/faiss (opens in a new tab)
Ref 00NUI6SE
Cite this
APA
Interestana AI Editorial. (2026, October 27). LEANN Vector Index Reduces RAG Index Size Significantly. Interestana. https://interestana.com/articles/meet-leann-an-open-source-vector-index-that-shrinks-a-60m-chunk-rag-index-from-201-gb-to-6-gb-00nui6se
MLA
“LEANN Vector Index Reduces RAG Index Size Significantly.” Interestana, 27 Oct. 2026, interestana.com/articles/meet-leann-an-open-source-vector-index-that-shrinks-a-60m-chunk-rag-index-from-201-gb-to-6-gb-00nui6se.
Chicago
Interestana AI Editorial. “LEANN Vector Index Reduces RAG Index Size Significantly.” Interestana, October 27, 2026. https://interestana.com/articles/meet-leann-an-open-source-vector-index-that-shrinks-a-60m-chunk-rag-index-from-201-gb-to-6-gb-00nui6se.
BibTeX
@misc{interestana_00nui6se,
author = {{Interestana AI Editorial}},
title = {{LEANN Vector Index Reduces RAG Index Size Significantly}},
howpublished = {Interestana},
year = {2026},
month = oct,
url = {https://interestana.com/articles/meet-leann-an-open-source-vector-index-that-shrinks-a-60m-chunk-rag-index-from-201-gb-to-6-gb-00nui6se},
note = {Summary of reporting by MarkTechPost, https://www.marktechpost.com/2026/10/27/leann-vector-index-mlsys-2026/}
}Credit the source
Original reporting: MarkTechPost, “Meet LEANN: An Open-Source Vector Index that Shrinks a 60M-Chunk RAG Index from 201 GB to 6 GB”, https://www.marktechpost.com/2026/10/27/leann-vector-index-mlsys-2026/(opens in a new tab)
Follow this story
Get Computing & Chips in your inbox — new stories in 40 words or fewer, once a week.
Subscribe