Summary Youtu-RAG is an agentic retrieval-augmented generation system for local deployment, autonomous decision-making, and memory-driven Q&A. It targets users needing robust, privacy-preserving personal knowledge base management. The system enhances traditional RAG via autonomous strategy selection and continuous learning through dual-layer memory, evolving Q&A capabilities beyond passive retrieval.

How It Works The system employs a "Local Deployment · Autonomous Decision · Memory-Driven" paradigm. Agents autonomously select optimal retrieval strategies and tool calls. Its dual-layer memory includes short-term conversational context and long-term knowledge accumulation for continuous Q&A learning and self-evolution. A file-centric architecture supports diverse data formats, integrating with MinIO for local object storage, ensuring data privacy.

Quick Start & Requirements Requires Python 3.12+ and uv. Local MinIO object storage is necessary. Model deployment includes Youtu-Embedding (required) and optional Youtu-Parsing/HiChunk. Installation: clone, uv sync, activate env. Configuration via .env needs LLM/embedding service details, API keys. Start with start.sh or uvicorn; frontend at http://localhost:8000. Docs: https://youtu-rag-docs.vercel.app.

Highlighted Details

• File-Centric Architecture: Supports 12+ formats (PDF, Excel, Images, Databases).
• Adaptive Retrieval Engine: Autonomous strategy selection, diverse tool calls (web search, vector retrieval, DB queries).
• Dual-Layer Memory: Short-term conversational and long-term knowledge accumulation for evolving Q&A.
• Ready-to-Use Agents: 8+ specialized agents (Web Search, Text2SQL, Excel) for varied scenarios.
• Lightweight WebUI: Framework-free, native HTML/CSS/JS interface, zero dependencies, streaming responses.
• Security & Control: Local deployment ensures data isolation; MinIO for large-scale file management.

Maintenance & Community Contribution guidelines cover bug reports, feature suggestions, documentation, and code improvements. Specifics on active maintainers, community channels, or a roadmap are absent from the README.

Licensing & Compatibility Licensed under the MIT License, permitting broad commercial use and integration within closed-source applications.

Limitations & Caveats LLM challenges with long document context impact deep Q&A. OCR/HiChunk parsing can cause significant upload delays; single-file imports are recommended. Knowledge base operations are limited to single knowledge base management. The Memoria-Bench evaluation benchmark is pending release.