LEANN (Low-overhead Embedded Approximate Nearest Neighbor) is an open-source vector retrieval and local Retrieval-Augmented Generation (RAG) engine developed by the StarTrail team and awarded Best Paper at MLSys 2026. It introduces a breakthrough indexing architecture that reduces vector index storage by 97% compared to traditional HNSW indices while maintaining superior recall accuracy.
For local AI assistants running on developer laptops, mobile phones, and edge devices, standard vector databases pose a severe storage bottleneck, frequently demanding gigabytes of memory and disk space for modest document collections. LEANN circumvents this limitation through novel non-linear quantization and selective graph pruning, enabling million-vector indexing within tens of megabytes of RAM.
In addition to radical memory compression, LEANN provides sub-millisecond retrieval latencies and native integration with local embedding models and small language models (SLMs). This allows desktop software engineers to operate completely private, offline semantic search across their entire personal codebase and document library with negligible hardware overhead.
Licensed under permissive MIT terms with over 12,800 GitHub stars, LEANN bridges the critical divide between server-scale vector search performance and resource-constrained edge computing environments.
