AutoFlow is an open-source Graph RAG conversational knowledge base tool from PingCAP, built on TiDB Vector, LlamaIndex and DSPy, aimed at teams that want to turn their own documentation sites into a searchable, answer-generating chatbot.
What it is
AutoFlow is a knowledge base and conversational search application maintained by PingCAP. It combines graph RAG — retrieval over a knowledge graph rather than flat vector chunks alone — with a chat interface, so that questions are answered from ingested content instead of from a general-purpose model's memory. The project lives in the TypeScript ecosystem and is written in TypeScript, with a Next.js frontend and a separate backend service distributed as Docker images.
The concrete problem it solves is standing up a documentation question-answering system without wiring together a crawler, an embedding pipeline, a vector store and a chat UI from scratch. It replaces that assembly work: the built-in website crawler navigates official and documentation sites and scrapes sitemap URLs, the content lands in TiDB Vector storage together with chat history and JSON, and the Perplexity-style conversational search page serves answers back. TiDB is used not only as the vector store but also for chat history and analytics, so a single database backs the whole application.
Key capabilities
- Graph RAG retrieval combining a knowledge graph with vector search, using TiDB Vector as the underlying store.
- A Perplexity-style Conversational Search page as the primary user-facing interface.
- A built-in website crawler that navigates official and documentation sites and scrapes sitemap URLs for ingestion.
- An embeddable JavaScript snippet that adds the conversational search window to an existing site, typically in the bottom right corner.
- Docker images published as
tidbai/backend and tidbai/frontend on Docker Hub.
- A stack built on LlamaIndex for retrieval orchestration and DSPy for programming foundation models rather than prompting them.
- A frontend built with Next.js, Tailwind CSS and shadcn/ui.
Who uses it and how
- Product and developer documentation teams that want a chat window answering questions from their own docs site rather than from a general model.
- Teams already running TiDB or MySQL-compatible infrastructure, since the topics list includes
mysql and the design consolidates vectors, chat history and analytics in one database.
- Website owners who want search embedded into an existing page by pasting a JavaScript snippet, without rebuilding the front end around it.
- Deployments on serverless vector storage, reflected in the
serverless and vector-database topics and in the TiDB Serverless Vector Storage description.
- Self-hosters running the full application locally or on their own hardware using the provided Docker Compose deployment.
Getting started
Deploy with Docker Compose, which the README sizes at 4 CPU cores and 8GB RAM, pulling the tidbai/backend and tidbai/frontend images. A hosted live demo runs at https://tidb.ai, with deployment documentation at https://autoflow.tidb.ai.
How it compares
No list of paid products that AutoFlow replaces is provided in the facts, and the facts name no directly comparable alternative tools in this registry. It therefore stands alone here, with LlamaIndex and DSPy appearing as libraries it is built on rather than as competitors.
When to use it — and when not to
Use it when a self-hosted, Apache-2.0 licensed conversational knowledge base over your own content is the goal, and when operating a Docker Compose stack backed by TiDB is acceptable. Do not pick it if a stable, production-frozen API matters: the README carries an explicit warning that AutoFlow is still in the early stages of development, and states that the next move is to become a Python package installable with pip install autoflow-ai. The project is under active maintenance — the most recent push recorded is 2026-04-27, with 74 open issues — but the packaging and interface are still moving, so plan for change.
