xerrors/Yuxi

What it solves

Yuxi is a self-deployable, multi-tenant knowledge agent platform designed for teams that need full control over their data, models, and permissions. It streamlines the process of turning unstructured documents into actionable intelligence by combining RAG, knowledge graphs, and multi-agent orchestration in a single workspace.

How it works

The platform integrates several core components to handle the end-to-end AI workflow:

• Knowledge Processing: Uses engines like MinerU, PaddleX, and RapidOCR to parse documents (PDF, Word, PPT, etc.) into chunks and vector indices via Milvus.
• Hybrid Retrieval: Combines vector search with knowledge graphs (Neo4j) to extract entities and relationships for more accurate, traceable answers.
• Agent Orchestration: Powered by LangGraph, it allows the creation of agents that can delegate complex tasks to asynchronous SubAgents, utilizing MCP (Model Context Protocol) and custom Skills for extended capabilities.
• Secure Execution: Tasks run in isolated sandbox file systems, allowing agents to generate and save files (Markdown, HTML, CSV) that users can preview and download.
• Governance: A multi-tenant system manages user/department permissions, API keys for various LLM providers, and system monitoring via a dashboard.

Who it’s for

Teams and enterprises requiring private deployment of AI agents to maintain data sovereignty while enabling collaborative knowledge management and complex task automation.

Highlights

• Traceable RAG: Provides precise source citations and built-in tools to evaluate retrieval and answer quality.
• Knowledge Graph Integration: Automatically extracts entities and relations to visualize knowledge as interactive topology maps or mind maps.
• Multi-Agent Parallelism: Supports a main agent planning tasks and dispatching them to multiple SubAgents for parallel execution.
• Sandbox Workspace: Isolated environments for generating and interacting with file-based deliverables directly in the browser.
• MCP & Skills Ecosystem: Native compatibility with the Model Context Protocol and a plugin system for installing and editing skills.