AutoFlow is a knowledge-base tool that answers questions over your documentation using graph RAG (retrieval-augmented generation backed by a knowledge graph). It is built on TiDB Vector for storage, LlamaIndex as the RAG framework, and DSPy for programming the language model, with a Next.js front end. A public instance answering questions about TiDB runs at tidb.ai.
You point a built-in crawler at official or documentation sites, and it scrapes them through sitemap URLs to build the knowledge base. Visitors then use a Perplexity-style conversational search page, or you embed a JavaScript snippet that adds a chat widget, typically in the bottom-right corner of your own site, to answer product questions in place.
The project is in an early stage of development and is deployed with Docker Compose, with a stated minimum of 4 CPU cores and 8 GB of RAM. The maintainers plan to repackage it as a Python package so it can be used as a RAG library.
Features
- Graph RAG retrieval: answers grounded in a knowledge graph plus vector search rather than vector search alone
- Conversational search page: a Perplexity-style chat interface over the crawled knowledge base
- Website crawler: sitemap-driven scraping of documentation and official sites
- Embeddable widget: a JavaScript snippet that adds the chat window to any website
- TiDB storage: chat history, vectors, JSON, and analytics stored in one database
- Framework stack: LlamaIndex for RAG, DSPy for model programming, Next.js, Tailwind CSS, and shadcn/ui for the UI
- Docker Compose deployment: backend and frontend images with a 4-core, 8 GB baseline
