RAGFlow Is the Open-Source RAG Engine That Quietly Built the Best Context Layer for AI Agents, and Nobody in the West Is Paying Attention

The RAG conversation in 2026 has two camps. One says RAG is dead, killed by long context. The other says RAG is back, because agents need grounding that does not fit in a single window. Both miss the more interesting story: while the West argues about whether retrieval matters, a Chinese team shipped a 60,000-star Apache 2.0 RAG engine that quietly became the most complete open-source context layer on GitHub. RAGFlow is not new. It just works, and it does the parts everyone else makes you do yourself.

Hi guys, Mr. Technology here.

RAGFlow is built by infiniflow, a Beijing team shipping since April 2024. Apache 2.0, ~60,000+ GitHub stars, and it bills itself as a "converged context engine" — the orchestration layer between your documents and your agents. The piece the Western press keeps missing is that RAGFlow is one of the few open-source projects in this space that has shipped every primitive a production RAG system needs in a single deployable: ingestion, parsing, hybrid retrieval, reranking, agent orchestration, MCP, memory, and multi-channel delivery.

What It Actually Is

A self-hostable Docker stack (Elasticsearch + Infinity + Redis + MySQL + MinIO), a web UI, and an agent orchestration layer with pre-built templates. The killer feature is the ingestion pipeline: RAGFlow ships MinerU and Docling as first-class parsers, OCR via PaddleOCR, layout-aware table extraction, and an orchestrable pipeline you can extend. Point it at a 200-page PDF with embedded charts, scanned pages, and Chinese mixed with English, and it returns clean structured chunks with tables intact. That pipeline alone is why most RAG projects fail. Most open-source RAG stacks hand you a chunker and tell you to figure out parsing. RAGFlow ships the parser.

The retrieval layer is hybrid by default: dense vectors + BM25 + a configurable reranker (BGE, Cohere, Jina). Hybrid wins because dense embeddings miss exact terms, BM25 misses paraphrase, the reranker arbitrates. RAGFlow ships the right combination as the default and lets you swap components.

The Agent Layer Is Where RAGFlow Pulls Ahead

Most RAG frameworks stop at "return top-k chunks to the LLM." RAGFlow ships agentic workflows as a first-class primitive: a visual graph editor, pre-built templates for extraction, classification, multi-hop QA, memory, code execution, plus MCP server and client support. The pieces that surprised me:

• Memory. Real memory for agents — episodic, semantic, short-term — with explicit APIs instead of hand-wavy "we store chat history."
• Code execution. A Python and JavaScript sandbox inside the agent runtime.
• MCP. Both client and server. Wire RAGFlow into Claude Desktop as an MCP server, or have its agents call out to MCP tools.
• Multi-channel delivery. Built-in connectors for Feishu, Discord, Telegram, Line (June 2026). Build the agent once, deploy to chat ops.

What It Beats, What It Loses To

Versus Dify. Dify is the more popular Western RAG-plus-agent platform and easier to demo. RAGFlow is harder to install, uglier in the UI. But the document parsing is materially better, the hybrid retrieval defaults are more honest, and the agent memory is implemented instead of promised. For messy PDFs, RAGFlow wins. For a five-minute demo, Dify wins.

Versus LangChain + LlamaIndex + a vector DB. You can build anything RAGFlow ships in roughly six weeks with three engineers, three packages, and one AWS bill you do not want to look at. RAGFlow is the assembled version. Every team rebuilds the same parsing-pipeline-reranker-MCP stack; RAGFlow ships it already wired, and the parts hardest to get right (OCR, table extraction, layout detection) are where it is best in class.

Versus closed SaaS RAG. Glean, Hebbia, and the rest charge $30k-$300k/year for what is essentially a hosted RAGFlow with nicer CSS. If your data is sensitive or your CFO hates $100k line items, RAGFlow is the answer.

What I Don't Love

The UI is dated. Looks like a 2019 admin panel. It works. For a 60k-star project, the front-end investment is the most visible gap.

English docs trail Chinese. English README is complete but lags the Chinese originals by weeks.

The Docker stack is heavy. Elasticsearch + Infinity + Redis + MySQL + MinIO is a lot of moving parts for a single-node dev setup. Slim-mode deployment helps.

Community is concentrated in Asia. Discord is bilingual; English GitHub discussions are slower. If you are a Western team, write internal docs because the public ones will not answer your deployment question in 24 hours.

The Take

RAGFlow is not the flashiest RAG engine on GitHub. It is the most complete — the only Apache 2.0 project I have used in 2026 that ships parsing, hybrid retrieval, reranking, agents, MCP, memory, code execution, and multi-channel delivery in a single deployable that actually works on real PDFs.

The reason nobody in the U.S. press has noticed is the same reason nobody noticed LLaMA-Factory for two years: it ships out of Beijing, the docs lead in Chinese, and the English coverage trail followed the Chinese adopters instead of the other way around. That is a Western press problem, not a project problem.

If you are evaluating RAG stacks in 2026, the answer is RAGFlow for the messy-document-plus-agent use case, Dify for the polished demo, LangChain/LlamaIndex for primitives, and a hosted SaaS like Glean when your CFO has signed off and your data is not sensitive. RAGFlow is the boring, complete option that ships, and in 2026, that is still the highest compliment in infrastructure.

— Mr. Technology

*RAGFlow: github.com/infiniflow/ragflow — Apache 2.0, ~60,000+ GitHub stars. Built by infiniflow (Beijing). Stack: Elasticsearch + Infinity + Redis + MySQL + MinIO. Parsers: MinerU, Docling, PaddleOCR. Models: GPT-5, Claude, Gemini 3 Pro, DeepSeek V4, Qwen 3, GLM-5, Llama 4, Mistral, Ollama. Channels: Feishu, Discord, Telegram, Line. MCP: client + server. Memory: episodic/semantic/short-term. Code execution: Python + JS. Deployment: docker compose -f docker/docker-compose.yml up -d; web UI on :9380.*