An agent harness that glues RAG, graphs, and multi-tenancy together

What it does

Yuxi is a multi-tenant “agent harness” that wires together LangGraph orchestration, LightRAG-powered knowledge graphs, and conventional RAG pipelines behind a Vue + FastAPI stack. It gives each tenant isolated knowledge bases, document parsing (including PDF via MinerU), and agent configuration with skills, MCP servers, and sub-agents. The frontend covers the usual suspects: chat interfaces, knowledge graph visualization, dashboard stats, and user/department permission management.

The interesting bit

The project treats knowledge graphs as first-class citizens rather than a RAG afterthought—LightRAG builds property graphs that participate directly in agent reasoning, not just retrieval. There’s also a sandboxed filesystem (/home/gem/user-data) where agents can write artifacts, preview files, and have their outputs surfaced as downloadable deliverables in the UI. The LITE mode lets you spin up without the heavy knowledge base modules if you just want the agent shell.

Key highlights

• LangGraph-based agent orchestration with sub-agents, skills, MCP tools, and middleware hooks
• LightRAG integration for graph construction and visualization via Neo4j, with property graphs feeding into agent reasoning
• Multi-tenant architecture with user/department scoping and API key authentication for external integrations
• Sandbox filesystem with artifact presentation—agents generate files that appear as collapsible cards in the chat UI
• Document parsing pipeline with embedding/rerank configurability and knowledge base evaluation metrics
• Dark mode, Docker packaging, and a “LITE” startup mode for lighter deployments

Caveats

• The README is primarily in Chinese; English documentation exists but appears secondary
• Several recent bug fixes suggest active rough edges around model switching in LightRAG and thread configuration binding
• Heavy dependency stack (LangChain ecosystem, Neo4j, Docker) means this is not a lightweight experiment

Verdict

Worth evaluating if you’re building internal agent platforms that need document-heavy RAG plus graph reasoning under multi-tenant isolation. Skip it if you want a simple API wrapper or a single-user local LLM UI—this is infrastructure, not a quick prototype.