// the find
xerrors/Yuxi
结合知识库、知识图谱管理的 多租户 Agent Harness 平台。 An agent harness that integrates a LightRAG knowledge base and knowledge graphs. Build with LangChain + Vue + FastAPI, support DeepAgents、MinerU PDF、Neo4j 、MCP.
Yuxi is a multi-tenant enterprise knowledge base and agent orchestration platform built on FastAPI, LangGraph, and Vue 3. It combines RAG retrieval with Milvus-backed knowledge graphs, document parsing via MinerU/PaddleX, and a ChatGPT-style interface where agents can be equipped with skills, MCP tools, sub-agents, and sandboxed code execution. The target audience is teams building internal AI assistants over proprietary document corpora.
The storage architecture is thoughtful — Milvus for vector search, Neo4j for graph traversal, and PostgreSQL for relational data are all properly separated rather than crammed into one system. The document parsing pipeline is unusually complete: MinerU, PaddleX, RapidOCR, and a RAGflow-inspired chunking strategy with parsers for books, laws, QA formats, and semantic splits. The multi-tenant model is a real feature, not a label — departments, workspaces, and per-user agent configs are first-class objects. The test suite has genuine coverage: unit, integration, and e2e tests with a real database, not mocks.
The dependency footprint is brutal — Postgres, Redis, MinIO, Milvus, Neo4j, and optionally MinerU and PaddleX all need to be running before you get a working system. The 'lite mode' helps but the graph and knowledge features, which are the whole point, require the full stack. Documentation is primarily in Chinese with thin English coverage, which will stop most non-Chinese-speaking teams at the README. The project is early-stage enough that the graph implementation recently replaced LightRAG with a custom Milvus-backed graph, meaning anyone who deployed before v0.7 is looking at a non-trivial migration. Observability is thin — Langfuse integration exists but token usage tracking across nested sub-agent calls is a known rough edge.