title	LEANN RAG • Qwen3 0.6B
emoji	🧠
colorFrom	blue
colorTo	indigo
sdk	gradio
sdk_version	5.49.1
app_file	app.py
pinned	false
license	mit
short_description	97% smaller vector search running on CPU

🧠 LEANN RAG: The "Impossible" Vector Search

This Space demonstrates LEANN (Lightweight Embedding & Neural Network), a revolutionary vector search engine that runs highly efficient RAG (Retrieval Augmented Generation) on consumer hardware.

Unlike traditional vector databases that store heavy embedding vectors for every document (bloating storage), LEANN stores only the graph structure and re-computes embeddings on-the-fly during search. This reduces index size by 97%, allowing us to run a powerful semantic search engine entirely on the Hugging Face Free Tier (2 vCPU).

---

🏗️ Architecture: How it Works

This application is a complete RAG pipeline optimized for low-resource environments.

graph LR
    A[User PDF/Text] -->|Chunking| B(Text Chunks)
    B -->|LEANN Indexer| C{HNSW Graph Construction}
    C -->|Compress| D[Sparse Graph Index]
    
    D -->|User Query| E[Graph Traversal]
    E -->|On-Demand Compute| F[Embedding Re-calculation]
    F -->|Top-K Results| G[Qwen3-0.6B LLM]
    G -->|Response| H[User Chat]
    
    style D fill:#f96,stroke:#333,stroke-width:2px
    style F fill:#f9f,stroke:#333,stroke-width:2px

Loading