StarTrail-org/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.

What it solves

LEANN is a lightweight vector database designed to enable high-performance Retrieval-Augmented Generation (RAG) on personal hardware. It solves the problem of excessive storage requirements in traditional vector databases, allowing users to index millions of documents on a laptop without sacrificing search accuracy.

How it works

LEANN utilizes a technique called graph-based selective recomputation with high-degree preserving pruning. Instead of storing every embedding for every piece of data, it computes embeddings on-demand. It also employs smart graph pruning and the Compressed Sparse Row (CSR) format to minimize the storage overhead of the graph itself.

Who it’s for

It is built for individuals who want to create a private, local AI assistant capable of semantic search across their personal data—such as file systems, emails, browser history, chat logs (WeChat, iMessage), and agent memories (ChatGPT, Claude)—without relying on cloud services or expensive hardware.

Highlights

  • Extreme Storage Efficiency: Reduces storage needs by up to 97% compared to traditional vector DBs (e.g., 60 million chunks in 6GB vs 201GB).
  • Privacy-First: Operates entirely on the local machine, ensuring data never leaves the device.
  • Broad Integration: Supports a wide range of data sources and is fully compatible as a semantic search MCP service for Claude Code.
  • Flexible Backend: Supports multiple embedding and LLM providers, including local engines like Ollama and LM Studio, as well as cloud APIs.

Related

  • Project
    HKUDS/LightRAG

    A lightweight, graph-based RAG framework that combines knowledge graphs and vector embeddings to provide efficient, high-quality retrieval and global context understanding.

  • Project
    mudler/LocalRecall

    A lightweight, offline RESTful API for managing vector-based knowledge bases, enabling AI agents to handle local long-term and short-term memory without cloud services.

  • Project
    infiniflow/infinity

    An AI-native database designed for LLM applications that provides high-speed hybrid search across dense and sparse embeddings, tensors, and full-text data.

  • Project
    Bessouat40/RAGLight

    A lightweight and modular Python library for implementing Retrieval-Augmented Generation (RAG), allowing users to easily integrate various LLMs and vector stores to chat with their documents.

  • Project
    tryAGI/LangChain

    A C# implementation of the LangChain framework that enables .NET developers to build composable AI applications using LLMs, vector databases, and document loaders.