// the find
codefuse-ai/codefuse-chatbot
An intelligent assistant serving the entire software development lifecycle, powered by a Multi-Agent Framework, working with DevOps Toolkits, Code&Doc Repo RAG, etc.
A DevOps-focused AI chatbot from Ant Group's CodeFuse team that wires together RAG over code repos, multi-agent orchestration, and a sandboxed code executor. It's aimed at teams who want a self-hosted ChatGPT-style assistant that understands their internal codebases and can interact with DevOps tooling. The primary README is in Chinese with an English translation that lags behind.
The multi-agent scheduling layer (extracted as the separate `codefuse-muagent` pip package) is a real architectural separation that lets you compose agents without forking the whole chatbot. The sandbox execution environment for running generated code is a meaningful safety boundary that many similar projects skip. Supporting both local open-source models via FastChat and cloud APIs (OpenAI-compatible) gives you an actual path to air-gapped deployment. The knowledge graph integration alongside vector search is a less common combination that helps with structured DevOps domain knowledge rather than pure similarity retrieval.
The last commit was July 2024 — a year stale, and there are open issues with no responses, which is a bad sign for a project this operationally complex to self-host. The setup requirements are steep: Python 3.9, CUDA 11.7, Docker, config files you copy from .example files, and a multi-process start.sh — getting this running is going to take a full afternoon. The English documentation is clearly a translation afterthought; anything non-trivial sends you to Chinese docs or WeChat. The LLM model table in the README lists only ChatGPT and one quantized CodeLlama variant as 'coming soon', meaning the model compatibility story is essentially undocumented.