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

UC Berkeley's open-source LEANN stores a pruned graph and recomputes embeddings at query time. It matches HNSW (a standard graph index) accuracy at 90% recall using 4 GB instead of 188 GB. The cost is search latency measured in seconds.

Why it matters: Storage is a major cost of local and personal-scale RAG, so trading compute for ~30x less disk makes on-device retrieval over large corpora practical. The seconds-level latency limits it to non-interactive or small-scale uses, but it reflects a broader shift of recomputing instead of storing as GPUs get cheaper than memory.