# Yuxi

> Yuxi is an open-source, self-hosted knowledge agent platform combining Agentic RAG, Neo4j knowledge graphs, LangGraph multi-agent runs, MCP/Skills and sandboxes.

- **Category:** AI Agents
- **Licence:** MIT
- **Latest version:** v0.7.2
- **Open source:** oss
- **Languages:** JavaScript, Python, Vue
- **Platforms:** Docker Compose
- **Tags:** Mcp, Ai Agents, Knowledge Base, Ocr, LLM, Rag, Vector Search, Knowledge Graph, Docker, Fastapi, Self Hosted, Vue, Neo4j, Multi Agent, Langgraph, Harness, Milvus, Llms, Agentic rag, Document Parsing, Document AI, Deepagents, Kbqa, Kgqa

## Description

Yuxi (语析) is a multi-tenant knowledge agent platform organizations deploy on their own hardware: document knowledge bases with agentic RAG, entity and relation extraction into a Neo4j graph, LangGraph-powered agent orchestration, MCP and Skills extensions, sandboxed tool execution that produces downloadable file artifacts, and user/department level access control. Agents retrieve from Milvus-backed knowledge bases using vector, BM25 and hybrid search with reranking, can dispatch several sub-agents in parallel, pause for human approval before risky operations, and write reports, tables, HTML pages or code into a persistent workdir the browser can preview. Administrators centralise model provider credentials across vendors such as OpenAI, Anthropic, DeepSeek, Qwen and OpenAI-compatible self-hosted endpoints, publish shared agents, and watch token usage and request trends on a dashboard; retrieval quality can be measured with built-in QA evaluation sets or Langfuse datasets, and login works with local passwords or an enterprise OIDC provider. Deployment is Docker Compose driven: clone the matching release tag, run the init script that writes a .env with JWT, API-key and sandbox secrets, start the stack, then open the web UI on port 5173 while the API listens on 5050. A production compose file with its own env file, TLS reverse-proxy guidance and an overlay that disables the knowledge base stack is documented for real installations. The Cube sandbox backend needs a dedicated Linux/KVM execution node, while a container sandbox profile covers simple local trials. The code is MIT licensed and versioned below 1.0 (latest stable release v0.7.2, with a v0.8.0 candidate used by the quick start), and upgrades between versions require a maintenance window with backup and migration.

## Links

- [Homepage](https://github.com/xerrors/yuxi)
- [Website](https://xerrors.github.io/Yuxi/)
- [快速开始文档](https://xerrors.github.io/Yuxi/intro/quick-start)
- [生产部署与升级](https://xerrors.github.io/Yuxi/advanced/deployment)
- [GitHub 仓库](https://github.com/xerrors/Yuxi)
- [版本记录](https://github.com/xerrors/Yuxi/releases)
- [DeepWiki 项目文档](https://deepwiki.com/xerrors/Yuxi)

---
Source: https://selfhost.directory/project/yuxi