labring/FastGPT

A knowledge-based AI agent platform that pairs out-of-the-box data processing and RAG retrieval with visual workflow orchestration.

What it is

FastGPT is a platform for building AI agents and question-answering systems on top of LLMs, offering data processing, model invocation, and RAG retrieval without extensive setup. Applications are assembled through visual Flow-based workflow orchestration, so complex conversational and plugin pipelines can be built and debugged in one place. It targets teams that want to stand up knowledge-base-backed assistants and agentic workflows, either self-hosted or via the hosted service.

Key features

Application orchestration through visual workflows, including conversational and plugin workflows with basic RPA nodes, agent skill orchestration, user interaction steps, and bidirectional MCP.
Knowledge base capabilities: reuse and mixing of multiple libraries, chunk editing and deletion, manual/direct-segmentation/QA-split import, and ingestion of txt, md, html, pdf, docx, pptx, csv, and xlsx plus URL reading and CSV batch import.
Hybrid retrieval with reranking, plus an API-backed knowledge base option.
Debugging tools: single-point search testing against a knowledge base, citation feedback that can be edited or deleted during chat, full call-chain logs, and application evaluation.
Operations features: login-free share windows, one-click iframe embedding, unified conversation records with data annotation, and application operation logs.

Tech stack

TypeScript on Next.js, organized as a pnpm/Turborepo monorepo (@fastgpt/app, @fastgpt/admin, @fastgpt/global, @fastgpt/service).
Package manager pinned to pnpm@10.33.4 (pnpm 10.x); Node.js >=20.19.0.
Chakra UI (theme typings generated via @chakra-ui/cli) and i18next / next-i18next for internationalization.
MongoDB in the test and tooling path (mongodb-memory-server, a withMongo test runner); Vitest for testing; ESLint and Prettier for linting/formatting.
Deployed with Docker Compose (guided install script), or on Sealos; root manifest version 4.0.

When to reach for it

Building knowledge-base-driven chatbots and question-answering systems over your own documents.
Assembling agent and RPA-style workflows visually rather than coding an orchestration layer.
Self-hosting a RAG platform via Docker or Sealos, or using the hosted cloud version.
Embedding an assistant into other sites through iframe or login-free share links.

When not to reach for it

You want to offer the software itself as a multi-tenant SaaS — the license permits direct commercial backend use but prohibits providing SaaS without authorization.
You need a lightweight library or a single API call; FastGPT is a full platform with MongoDB, a vector store, and multiple services to operate.
Your workflow is code-first and you would rather express pipelines in a general-purpose framework than a visual Flow editor.

Maturity signal

FastGPT has been developed since early 2023 and remains actively maintained, with commits pushed the same day as this writing, a very large following, and an unusually high fork count that signals broad self-hosting and derivative use. It is well past early experimentation (root version 4.0) and backed by a substantial contributor community and a commercial offering. The main caveat is licensing rather than upkeep: it ships under a custom FastGPT Open Source License (reported as NOASSERTION), which constrains SaaS resale and should be reviewed before commercial deployment.

Alternatives

Dify — use instead when you want a broadly comparable LLMOps platform for apps, RAG, and agents under a more conventional open-source license.
RAGFlow — use instead when deep document parsing and retrieval quality are the priority over visual workflow orchestration.
AnythingLLM — use instead when you want a simpler, more self-contained document chatbot to self-host or run on the desktop.
Langflow — use instead when you prefer a code-adjacent, component-graph builder in the LangChain ecosystem.

Notes

The default README is Simplified Chinese, with English, Bahasa Indonesia, Thai, Vietnamese, and Japanese translations available.
Licensing is the notable constraint: direct commercial use as a backend service is allowed, but offering it as a SaaS requires authorization, and copyright notices must be retained.
The repository carries extensive internal agent design and skill documentation (.agents/design, .agents/skills), reflecting an agent-assisted development process, and lists sibling projects FastGPT-plugin, AI Proxy, and Sealos.

Wiki: labring/FastGPT