LEANN

[MLsys2026 Best Paper]: https://arxiv.org/abs/2506.08276. RAG on Everything with LEANN. Enjoy 97% storage savings while running a fast, accurate, and 100% priva

What is LEANN?

[MLsys2026 Best Paper]: https://arxiv.org/abs/2506.08276. RAG on Everything with LEANN. Enjoy 97% storage savings while running a fast, accurate, and 100% private RAG application on your personal device.

Overview of LEANN

What it is

LEANN is an open-source vector database that enables retrieval-augmented generation (RAG) on personal devices. It indexes and searches through millions of documents while using 97% less storage than traditional solutions, according to the project's claims. LEANN achieves this through graph-based selective recomputation with high-degree preserving pruning, computing embeddings on-demand instead of storing them all. It received the MLsys2026 Best Paper award.

Key features

97% storage savings compared to traditional vector databases
Graph-based selective recomputation and high-degree preserving pruning
Supports semantic search over file system, emails, browser history, chat history, agent memory (ChatGPT, Claude), live data (Slack, Twitter), codebases, and external knowledge bases
Drop-in semantic search MCP service compatible with Claude Code
No accuracy loss while using less storage
Portable: transfer knowledge bases between devices

Use cases

Semantic search over personal files, emails, browser history, and chat history (WeChat, iMessage)
Agent memory retrieval for ChatGPT and Claude
Live data search from Slack and Twitter
Codebase semantic search for coding agents (e.g., Claude Code)
Indexing large external knowledge bases (e.g., 60 million documents)

Getting started

Clone the repository and install with uv:bashgit clone https://github.com/yichuan-w/LEANN.git leanncd leannuv venvsource .venv/bin/activateuv pip install leann

License and pricing

LEANN is licensed under MIT and is open-source. The license is stated in the project metadata.

這是什麼

LEANN 是一個開源的向量資料庫，旨在讓個人裝置上的檢索增強生成 (RAG) 普及化。根據該專案的說法，它能夠索引和搜尋數百萬份文件，同時比傳統解決方案節省 97% 的儲存空間。LEANN 透過基於圖形的選擇性重算與高節點度保留剪枝來實現此效果，按需計算嵌入向量，而非全部儲存。該專案獲得了 MLsys2026 最佳論文獎。

主要功能

與傳統向量資料庫相比，可節省 97% 的儲存空間
基於圖形的選擇性重算與高節點度保留剪枝技術
支援對檔案系統、電子郵件、瀏覽器歷史記錄、聊天記錄、代理記憶 (ChatGPT, Claude)、即時資料 (Slack, Twitter)、程式碼庫及外部知識庫進行語意搜尋
可與 Claude Code 相容的即插即用語意搜尋 MCP 服務
在不犧牲準確度的情況下減少儲存空間使用
可攜帶：可在裝置之間轉移知識庫

使用情境

針對個人檔案、電子郵件、瀏覽器歷史記錄及聊天記錄 (微信、iMessage) 進行語意搜尋
為 ChatGPT 和 Claude 檢索代理記憶
搜尋 Slack 和 Twitter 的即時資料
為程式碼代理 (例如 Claude Code) 進行程式碼庫語意搜尋
索引大規模外部知識庫 (例如 6000 萬份文件)

快速開始

複製儲存庫並使用 uv 安裝：bashgit clone https://github.com/yichuan-w/LEANN.git leanncd leannuv venvsource .venv/bin/activateuv pip install leann

授權與價格

LEANN 採用 MIT 授權條款，為開源軟體。授權條款已在專案中標示。

Based on the project's public README and repository metadata.