LEANN is a radically compact vector index that makes retrieval-augmented generation practical on personal devices: it shrinks the search index by about 97% compared with conventional approaches, so you can run fast, fully private RAG over your own files on a laptop. The MLSys 2026 best-paper project has about 13,000 stars on GitHub.
Standard vector databases store a full embedding for every chunk, which means the index for a personal document collection can be larger than the documents themselves โ a real problem on a laptop or phone. LEANN takes a different route: it keeps a compact graph structure and recomputes embeddings on the fly during search instead of storing them all, trading a little compute for an enormous saving in storage while keeping accuracy and latency competitive. The project wraps this index in a ‘RAG on everything’ toolkit that can index documents, emails, chat history, browser data and code, and it integrates with the frameworks developers already use, exposing connectors for LangChain, LlamaIndex and local model runtimes such as Ollama. Because everything runs locally, your personal data never leaves the machine, which is precisely the scenario large cloud indexes cannot serve.
LEANN is free and open source under the MIT license. There are no hosted tiers; you run it on your own hardware and pay only for your own compute.
LEANN is for privacy-conscious users and developers who want semantic search and RAG over personal or sensitive data on-device, and for anyone building local-first AI applications where shipping data to a cloud vector database is not an option.