LEANN Open-Source Vector Index Compresses 60M-Chunk RAG Index from 201 GB to 6 GB

Researchers have released LEANN, an open-source vector index presented at MLSys 2026 that reduces a 60-million-chunk RAG index from 201 GB down to just 6 GB — a 33x compression ratio — without significant retrieval quality degradation. LEANN achieves this through a novel learned, compressed index structure that trades a small amount of additional query-time compute for dramatic memory and storage savings. For developers operating large-scale RAG pipelines, this breakthrough directly addresses one of the primary infrastructure bottlenecks: the cost and complexity of hosting large vector indexes in memory or on high-throughput storage. The open-source release means teams can immediately evaluate LEANN as a drop-in replacement or complement to FAISS, HNSW, or other standard ANN indexes. This has direct implications for reducing cloud storage costs and enabling RAG at scale on less expensive hardware.