Overview
RAGFlow is an open-source Retrieval-Augmented Generation engine built by InfiniFlow, and its central bet is that most RAG failures start with bad chunking, not bad models. Where a typical vector database just shoves text into embeddings, RAGFlow first runs what it calls ‘deep document understanding’ - parsing PDFs, scanned images, spreadsheets, slide decks, and web pages into structured, semantically meaningful chunks before anything gets embedded. That front-loaded cleaning is what lets it answer from a 200-page contract without hallucinating a clause that isn’t there.
In our evaluation the differentiator is explainability. RAGFlow visualizes each chunk and attaches traceable citations to every answer, so a legal or compliance reviewer can click through to the exact source span. Under the hood it combines vector search, BM25 full-text, custom scoring, and fused reranking rather than leaning on embeddings alone, and it has been moving toward agent orchestration - visual workflows, MCP connectors, web search, code execution, and chat channels. It is Apache-2.0 licensed and self-hostable via Docker Compose, with a managed cloud for teams that don’t want to run Elasticsearch, MySQL, MinIO, and Redis themselves.
Key Features
- Deep document understanding - parses complex, multi-format files (PDF, scans, xlsx, pptx, web) into clean semantic chunks instead of blind text splits.
- Explainable chunking and citations - visualize each chunk and trace every answer back to its source span, built for audit-heavy legal and finance work.
- Hybrid retrieval - combines vector search, BM25 full-text, custom scoring, and fused reranking for steadier relevance than embeddings alone.
- Agent orchestration - visual workflows with MCP connectors, web search, a code executor, and chat channels for multi-step tasks.
- Self-host or cloud - Apache-2.0 Docker deployment, or a managed tier with Free, Starter, Pro, and Enterprise plans.
- Configurable models - bring your own LLM and embedding provider; no vendor lock-in on the model layer.
Pricing
| Plan | Price | For |
|---|---|---|
| Self-hosted | $0 (Apache-2.0) | Engineers - run on your own infra; pay only for compute, storage, and model APIs |
| Free | $0 | Evaluators - 5 apps, 1 member, 0.1 GB storage, 500 credits/mo |
| Starter | ~$29/mo | Small teams - 50 apps, 5 members, 5 GB storage, 5,000 credits/mo (reference price may vary) |
| Pro | Custom quote | Growing orgs - unlimited apps, 20 members, 50 GB storage, 20,000 credits/mo |
| Enterprise | Custom | Production - BYOC, on-prem, dedicated support, custom SLA |
Comparison
vs. Dify: Dify is the broader visual LLM-app platform - agents, workflows, and model routing - whereas RAGFlow is narrower and deeper on document ingestion and retrieval quality. Choose Dify when you are building a general AI app; choose RAGFlow when the hard part is parsing and grounding over messy files.
vs. LangChain: LangChain is a code-first framework you assemble yourself, offering maximum flexibility at the cost of boilerplate. RAGFlow ships a working RAG pipeline with a UI and APIs out of the box. Pick LangChain for bespoke orchestration; pick RAGFlow to stand up a document Q&A system in an afternoon.
vs. Flowise: Flowise is a drag-and-drop node editor for LLM flows, friendly to non-engineers but light on document intelligence. RAGFlow’s chunking and citation tracing make it the stronger pick for knowledge-base assistants over real enterprise documents.