Yuxi

Yuxi is an open-source, self-deployable multi-tenant knowledge agent platform that its README summarizes as "Cloud Agents + Knowledge RAG". It brings knowledge-base retrieval, knowledge graphs, LangGraph multi-agent orchestration, MCP/Skills, sandbox tools and permission management into one workspace. It is released under the MIT License and started with Docker Compose.

What It Is

Yuxi targets teams that want to run an agent product on their own infrastructure while keeping control of data, models and permissions. Users chat with agents in a unified workbench, reference knowledge bases, personal files or Skills with @, and watch task steps, tool-call status and context token usage. Generated files can be previewed and downloaded.

Knowledge and RAG

Documents in formats such as PDF, Word, PPT, Excel and Markdown are parsed (MinerU, PaddleX and RapidOCR engines are listed), chunked and indexed as vectors in Milvus. Embedding and rerank models are configurable, with a retrieval test workbench and RAG evaluation using question-answer sets, including generated single-hop and multi-hop QA. Entities and relations can be extracted into Neo4j to build a knowledge graph, and a file-structure knowledge map can be browsed. External knowledge sources such as Dify and Notion can be connected.

Agents, Skills and Sandbox

An agent combines a model, prompts, knowledge bases, tools, MCP servers and SubAgents. A main agent can plan and dispatch SubAgents that run asynchronously in parallel. Skills can be installed online from skills.sh and ModelScope, edited, and given permissions and dependencies. Long-running tasks run through asynchronous workers, with human approval cards for sensitive actions. Sandboxed workspaces restrict file access and let agents produce Markdown, tables, HTML pages and code, previewed in the browser.

Team Governance and Setup

Administrators manage tenants, users, departments, sharing scope for knowledge bases, agents, Skills and models, plus API keys and a dashboard of requests, token use and queue status. Model providers include OpenAI, DeepSeek, Qwen, Zhipu, Moonshot, MiniMax, SiliconFlow, OpenRouter, ModelScope and OpenAI-compatible endpoints. Deployment uses Docker Compose with services including LangGraph, Vue, FastAPI, Milvus, Neo4j, PostgreSQL, MinerU and PaddleOCR; the docs describe an init script and a production deployment and upgrade guide.