CodeFuse-ChatBot is an assistant aimed at the whole software development lifecycle — design, coding, testing, deployment, and operations. It combines retrieval-augmented generation, tool learning, and a sandbox so a model can act on DevOps tasks rather than only answer questions about them.
The scheduling core is a multi-agent framework, released separately as the codefuse-muagent package, which is configured rather than coded to build interactive agents. Around it sit a web crawler for named URLs, a data processor for document loading, cleaning, and text splitting, text embedding and indexing, vector and graph databases, prompt management, a sandbox for compiling and running code, and an API layer for wiring in open-source components and operations platforms.
Two retrieval modes are specific to development work: repository-level code analysis, which reads a whole code base and can write and generate code at file level, and document analysis that pairs a document knowledge base with a knowledge graph. A DevOps-specific knowledge base can be built in one step for a given domain.
Models can be open-source and self-hosted for an offline, private deployment, reached through FastChat, or accessed through the OpenAI API. The project builds on Langchain-Chatchat and codebox-api.
| Model | Size | GPU memory | Quantisation |
|---|---|---|---|
| chatgpt | – | – | – |
| codellama-34b-int4 | 34b | 20g | int4 |
Features
- Multi-agent scheduling core: a configurable dispatcher that assembles interactive agents without code changes
- Repository-level code analysis: whole-repository understanding plus file-level code writing and generation
- Document analysis: a document knowledge base combined with a knowledge graph for retrieval and inference
- Domain knowledge bases: one-step construction of a DevOps-specific knowledge base
- Sandbox: an isolated local environment for compiling and executing code and actions
- Web crawling: retrieval of content from specified URLs to fill the knowledge base
- Storage: vector and graph databases behind embedding and indexing
- Model access: open-source models served through FastChat for offline deployment, or the OpenAI API
