yichuan-w/LEANN at 11.4K Stars With +186 Today: Berkeley RAG System Achieves 97% Storage Savings by Replacing Stored Embeddings With Graph Recomputation

GitHub·medium signal

LEANN from Berkeley Sky Computing Lab uses graph-based selective recomputation instead of permanently storing embeddings — maintaining only a pruned graph structure and recalculating vectors on-the-fly during search. High-degree preserving pruning retains hub nodes while eliminating redundant connections, achieving identical search quality to traditional vector databases at 1/30th the storage cost. The system supports local-first RAG across emails, browser history, chat, and code repos with zero cloud dependency, using HNSW or DiskANN backends.