xerrors/Yuxi

Yuxi – Private‑deployed, multi‑tenant Knowledge Agent Platform

What it is – Yuxi (pronounced “云西”) bundles a RAG engine, knowledge‑graph builder, LangGraph‑based multi‑agent orchestration, sandboxed file tools and fine‑grained team‑level permission management into a single Docker‑compose stack. It is aimed at enterprises that want to run their own AI agents on‑premise, keep data and model credentials in‑house, and let many users create and share custom agents.

Core capabilities

Module	What it does
Unified Agent Workbench	Web UI (Vue) for chat, tool calls, task status, and file download. Supports @‑mentions to pull in knowledge‑base docs, personal files or Skills.
Knowledge Base & RAG	Upload PDFs, Word, PPT, Excel, Markdown, etc.; automatic parsing (MinerU, PaddleOCR), chunking, embedding and vector indexing (Milvus). Built‑in rerank and evaluation dashboards (Langfuse‑style).
Knowledge Graph	Entity/relationship extraction from chunks, stored in Neo4j, visualised as interactive graphs and hierarchical knowledge maps.
Multi‑Agent & Extensibility	Configure a base LLM, attach Knowledge, Tools, Sub‑Agents, Skills and MCP (Model Context Protocol) services. Supports parallel SubAgent execution and plug‑in Skills.
Sandbox Workspace	Isolated file system for agents; generated Markdown, HTML, CSV, code, etc. can be previewed in‑browser and downloaded.
Team Governance	Multi‑tenant model with users, departments and shared scopes. Centralised API‑key vault for OpenAI, Anthropic, DeepSeek, Qwen, vLLM, etc.; dashboards for usage, token consumption and task queues.

Tech stack

• Frontend: Vue (http://localhost:5173) – workbench, admin console, dashboards.
• Backend: FastAPI (http://localhost:5050) – REST API, OpenAPI docs.
• Orchestration: LangGraph for agent flow control.
• Vector store: Milvus.
• Graph DB: Neo4j.
• Database: PostgreSQL.
• Document parsing: MinerU, PaddleOCR, RapidOCR.
• Containerisation: Docker Compose (all services start together).

Quick start

1. Install Docker Engine + Docker Compose.
2. Clone the repo at tag v0.7.3 and run ./scripts/init.sh (or the PowerShell equivalent) to generate .env with your large‑model API key (e.g., SiliconFlow).
3. docker compose up --build -d.
4. When http://localhost:5050/api/system/ready returns ready, open http://localhost:5173, create a super‑admin, and start creating knowledge bases, agents and Skills.

Typical use cases

• Internal help‑desk bots that answer from company manuals, policies and code bases.
• Automated report generation where an agent pulls data, runs analysis in a sandbox, and returns a formatted PDF/HTML.
• Knowledge‑graph‑driven discovery for R&D teams, linking entities extracted from technical documents.
• Multi‑step workflow automation (e.g., legal contract drafting) with human‑in‑the‑loop approvals.

Extensibility

• Add new Skills via the built‑in plugin manager (install scripts, configure permissions).
• Plug in additional MCP services for custom model back‑ends.
• Connect external knowledge sources like Dify or Notion through the unified retriever.

License – MIT for the Yuxi codebase; third‑party components retain their original licenses (see LICENSE and deployment docs).

Resources

• Project site: https://xerrors.github.io/Yuxi/
• Docs: quick‑start, model config, knowledge‑base, agent config, deployment.
• Demo video: Bilibili link in README.
• Sponsor: Fluxion AI (discounted model API access).

---

Yuxi is a full‑stack, self‑hosted platform that lets enterprises build, run and govern AI agents that reason over private documents and graph‑structured knowledge.