Graph RAG for your docs, with a TiDB backend and a JS widget

What it does AutoFlow ingests websites and documentation via a built-in crawler, then answers questions through a Perplexity-style chat interface. It also provides an embeddable JavaScript widget that drops a search window onto third-party sites. Under the hood, it uses TiDB Serverless Vector Storage to persist chat history, vectors, JSON, and analytics.

The interesting bit Rather than flat vector retrieval, it uses Graph RAG—layering knowledge graphs on top of LlamaIndex and DSPy—to reason over relationships between content chunks. That architecture is more ambitious than a standard vector-store chatbot, though the README leaves the exact graph construction details unsaid.

Key highlights

• Graph RAG pipeline built on LlamaIndex and DSPy for structured retrieval
• Built-in crawler with sitemap scraping for automated doc ingestion
• Embeddable JavaScript widget for drop-in conversational search on external sites
• Unified storage in TiDB Serverless for vectors, chat history, and JSON data
• Docker-based deployment with published frontend and backend images

Caveats

• Explicitly early-stage; the maintainers warn the project is still evolving
• The roadmap points toward a future Python package (autoflow-ai), so the current TypeScript stack may be transitional
• Deployment requires 4 CPU cores and 8 GB RAM, which is hefty for a docs bot

Verdict Best suited for teams already using TiDB who need a self-hosted knowledge base with graph reasoning. If you are looking for a stable, language-agnostic API today, the upcoming Python rewrite is not here yet.