LEANN: A Lightweight Vector Database for Local RAG Applications

LEANN is an innovative vector database designed to democratize personal AI by turning a standard laptop into a private retrieval-augmented generation (RAG) system. Developed at the Berkeley Sky Computing Lab by core contributors Yichuan Wang and Zhifei Li, alongside active contributors Gabriel Dehan and Aakash Suresh, LEANN aims to provide complete data privacy with zero cloud telemetry while achieving significant storage savings.

Core Architecture and Storage Efficiency

Traditional vector databases often store every individual embedding, which can consume substantial disk space. LEANN utilizes a graph-based selective recomputation approach with high-degree preserving pruning. Instead of storing all embeddings permanently, it computes them on-demand during search paths and leverages smart graph pruning alongside CSR formats to reduce overhead.

According to project documentation and benchmarks:

• Storage Reduction: Reported savings reach up to 97% compared to traditional solutions without noted accuracy loss.
• Example Scale: Indexing 60 million text chunks reportedly fits into roughly 6GB instead of 201GB.
• Search Backends: Supports HNSW as a default option for maximum storage efficiency, alongside DiskANN for advanced search performance needs.

Supported Data Sources and Integrations

LEANN supports a wide range of personal data types and live data streams through Model Context Protocol (MCP) servers:

• Personal Files: Standard documents including PDF, TXT, and Markdown files.
• Browser and Chat History: Integrations for Google Chrome history, Apple Mail, WeChat, ChatGPT conversation exports, Claude conversation exports, and iMessage databases.
• Live Platforms: Real-time RAG support for platforms like Slack and Twitter via MCP servers.
• Developer Tools: Native integration with Claude Code utilizing AST-aware chunking for programming languages like Python, Java, C#, and TypeScript.

Practical Implications and Verification

While LEANN offers local privacy and reduced storage footprints, users must verify compatibility requirements before installation. Setup procedures depend heavily on the operating system and chosen backend, such as requiring specific development tools for DiskANN or granting Full Disk Access on macOS to index local databases like Apple Mail or iMessage. Developers can inspect the arXiv preprint (arXiv:2506.08276) associated with the project and review the official GitHub repository for complete installation steps, CLI commands, and reproducible benchmark scripts.