// the find
labring/FastGPT
FastGPT is a knowledge-based platform built on the LLMs, offers a comprehensive suite of out-of-the-box capabilities such as data processing, RAG retrieval, and visual AI workflow orchestration, letting you easily develop and deploy complex question-answering systems without the need for extensive setup or configuration.
FastGPT is a self-hostable platform for building RAG-based Q&A systems and visual AI agent workflows on top of any OpenAI-compatible LLM. It ships a full UI for ingesting documents into a vector store, a node-based workflow builder, and a chat interface you can embed or expose via API. The target is teams that want a private, configurable alternative to hosted products like Dify or Coze.
The workflow engine is more serious than most visual LLM builders — it handles parallel nodes, interactive mid-flow user inputs, loop runs with streaming, and bidirectional MCP (the app can act as both an MCP client and server). The vector store layer is genuinely pluggable: PostgreSQL/pgvector, Milvus, Zilliz, openGauss, and others are all supported through a config swap, not a code rewrite. Hybrid retrieval with reranking is on by default, not bolted on as an afterthought. The `.agents/` directory reveals they dogfood AI agents on their own codebase for design docs, PR review, and bug analysis — the workflow tooling is not just a demo.
The license is a custom 'FastGPT Open Source License' that explicitly prohibits offering it as a SaaS — read it before building a business on this. The project is Chinese-first: community is on Feishu, most issues are in Chinese, and the English README lags behind. Self-hosting requires MongoDB, PostgreSQL, Redis, and the app container at minimum — the Docker Compose works but this is not a lightweight deployment. The workflow engine's design docs show a history of stream resumption bugs, stale-state resets, and interactive node edge cases that suggest the engine has rough edges at scale, particularly around concurrent or long-running workflows.