LEANN

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

This database stores 60M docs in 6 GB #aiagents #ai #opensource #vectordatabase #leann #claude LEANN is a vector database that skips storing every embedding. It keeps a pruned graph and recomputes embeddings only for the chunks on your search path. Their paper claims 97% less storage with the same search quality, all running locally on your laptop. In this Short: - What a vector database is (in plain words) - Why normal ones get huge: 201 GB for 60M chunks - LEANN's trick: store less, compute on demand - The numbers: 6 GB, 79 MB for emails - Why it's popular and what it changes Repo: StarTrail-org/LEANN Paper: arxiv.org Install: uv pip install leann Numbers are from the project's README (storage table + a 30-task Claude Code benchmark vs BM25). Results are the authors' own, so test it on your data. LEANN, LEANN vector database, vector database, RAG, retrieval augmented generation, local RAG, private AI, offline AI, vector search, semantic search, embeddings, FAISS alternative, Claude Code MCP, MCP server, open source AI, GitHub trending, AI tools, MLSys, Berkeley Sky Lab, AI on laptop, uv pip install, python AI, shorts #LEANN #RAG #VectorDatabase #LocalAI #OpenSource #AITools #Shorts

LEANN省下磁盘，查询要等多久？ #Shorts UC Berkeley 的 LEANN Explaine 用约 4 GB 索引达到与 HNSW 相同的 25.5% 问答准确率，但语义检索需 2.48 秒。 LEANN 是 UC Berkeley Sky Computing Lab 开发的开源向量索引，不是大语言模型；它这次展示的关键变化，是不再把完整的文本嵌入向量长期存进索引，而是在查询时按需重算。对想在本地给文档、邮件或代码库做语义检索的开发者来说，这可能大幅降低存储门槛，但省下的磁盘空间会转化成查询计算和等待时间。 论文评测的一个设置中，LEANN 用约 4 GB 索引达到与 HNSW 相同的 NQ 下游问答准确率 25.5%，但检索耗时是 2.48 秒，对比 HNSW 的 0.05 秒。注意，90% Recall@3 是检索目标，不是问答准确率；论文结果来自团队自己的基准评测，现有材料没有第三方复现。建库仍需先为所有文本块生成嵌入，CPU 延迟也没有公开数据。想试用时，可以先用 `uv pip install leann` 安装，再拿自己的语料对比索引大小、召回率和端到端延迟，判断它适不适合低频、本地检索场景。#LEANN #向量检索 #RAG #人工智能 ——— 🔗 论文：arxiv.org 🔗 GitHub：StarTrail-org/LEANN 这类底层拆解如果对你有用，订阅 + 打开小铃铛，新一期更新第一时间到你这儿。 ⏱ 章节 00:00 压缩幅度惊人，但LEANN的主张不是“又小又快”：它把存储优势和查询变慢同时 00:34 LEANN的关键变化不是取消向量计算，而是不再长期保存全部完整向量。 00:44 查询不是重算全库，而是先筛候选，再为搜索实际触及的节点重算嵌入。 01:14 在指定NQ评测中，LEANN达到与HNSW相同的下游准确率，但检索阶段明显更 01:35 LEANN更像是存储受限场景的选择，而不是低延迟检索的通用替代品；实际是否划 🔗 资源 · 代码：StarTrail-org/LEANN #Shorts 🔎 从本视频继续选型： zhiyong.dev --- 🔎 相关 AI 工具、模型与应用： zhiyong.dev

What if you could build a private AI assistant that searches millions of documents without a massive vector index? LEANN claims up to 97% less storage by recomputing embeddings on demand instead of storing Interesting for local RAG and AI coding agents StarTrail-org/LEANN

LEANN is an open source project that turns a big pile of files into a small local search index. It keeps a small map and rebuilds what it needs when you search, so your data stays on your machine. See the code at StarTrail-org/LEANN t.co