UC Berkeley's open-source LEANN shrinks a 201GB vector index to 6GB for laptop RAG
thisguyknowsai · x · 2026-11-04
LEANN, an open-source vector database from UC Berkeley, won Best Research Paper at MLSys 2026 and has 13.5k GitHub stars. Instead of storing embeddings for every chunk like traditional vector DBs, it keeps a tiny pruned graph and recomputes only the embeddings needed per query via graph-based selective recomputation with high-precision pruning — 97% storage savings with minimal accuracy loss.
Benchmarks: 780K email chunks 2.4GB → 79MB; 400K chat messages 1.8GB → 64MB; 38K browser history entries 130MB → 6.4MB; 60M text chunks fit in 6GB instead of 201GB.
It runs on your laptop, enabling natural-language semantic search over PDFs, Apple Mail, Chrome history, iMessages, old ChatGPT/Claude chats, and codebases. It ships with MCP integration for Claude Code: in the team's 30-task test, Claude Code with LEANN surfaced 2.1x more relevant code upfront than keyword search.
More from Infra
- oMLX 0.7.4.dev1 lets you pair a fast Mac as a prefill accelerator — Synor · 2026-11-04
- 128GB unified RAM laptop benchmarks: gemma4:e4b hits 122.8 TPS locally via Ollama — ollama · 2026-11-04
- China ordered $70B of Nvidia RTX Pro 5500 GPUs but only 1% can be delivered — rohanpaul_ai · 2026-11-04
- How to keep API bills from exploding when scaling vibe-coded apps — Ok_Class2432 · 2026-11-04
- YMTC pauses IPO on weakening market; SK Hynix halts Solidigm listing for the same reason — zephyr_z9 · 2026-11-04
- Building AI codebase search across hundreds of private repos: one POC architecture — fresh_preserve · 2026-11-04