If you work with a lot of documents — contracts, policy manuals, research reports — you've probably tried pasting chunks into ChatGPT and asking questions. It works until you have hundreds of files, until the key data is inside a table, or until your legal team asks where those files went.
RAGFlow addresses all three at once.
At a glance
| Fact | Detail |
|---|---|
| What it is | Open-source document question-answering engine with visual document understanding |
| Built by | InfiniFlow |
| License | Apache 2.0 |
| Platforms | Linux Intel/AMD (primary); macOS/Windows via Docker Desktop; Apple Silicon requires a custom build |
| Requires | Docker 24+, Docker Compose v2.26+, 4 CPU cores, 16 GB RAM, 50 GB disk |
| Install method | Docker Compose |
| Verdict | Install it when document accuracy matters more than setup speed |
What RAGFlow actually is
RAGFlow is an open-source engine for what the AI field calls RAG — retrieval-augmented generation. The concept: instead of an AI answering from its training data alone, you ground its answers in your own documents. You upload a file, the system indexes it, and when you ask a question the AI pulls the relevant passages and answers from those — with citations pointing to the exact source.
Most document-chat tools treat a PDF as a raw text dump. They extract characters and throw away structure: table borders, column layout, figure captions, reading order in multi-column formats. Ask about a number from a table and you risk getting the wrong figure because the tool lost the spatial context.
RAGFlow's key differentiator is what it calls deep document understanding. Before indexing a file, it parses each page's visual structure — identifying tables as tables, preserving header-to-cell relationships, reading multi-column text in the right order. The chunk of text the AI receives is faithful to what was on the page, not a jumbled extraction.
The project has 85,000+ stars on GitHub. There is a cloud-hosted version at cloud.ragflow.io if you want to try it before self-hosting.
What you get after installing it
Once running, RAGFlow is accessed through a browser. Three concepts drive everything:
Knowledge bases are document collections. You create one, upload files — PDF, Word, PowerPoint, Excel, plain text, scanned images (RAGFlow includes OCR, the technology that reads text from images), and web pages — and parsing runs automatically.
Chunking templates control how RAGFlow splits documents into searchable pieces. Chunking is the process of dividing a long document into smaller segments that can each be retrieved independently. RAGFlow ships named templates for different content types: Laws (for contracts), Paper (for research), Manual (for technical docs), Table (for spreadsheet-heavy files). You pick the template that matches your content, and the system respects that document type's structure when it splits.
Chat interfaces are the question-answering layer. You connect an assistant to one or more knowledge bases, choose an AI model (OpenAI, Claude, Gemini, DeepSeek, or a locally-hosted model), and start asking. Every answer includes click-through citations to the source passage.
Beyond the core: RAGFlow syncs from Confluence, Google Drive, Notion, and S3, so your knowledge base can stay current without manual re-uploads. An agentic workflow layer added in 2025 lets you chain steps — search a knowledge base, call an API, run code — without writing code.
The install experience
RAGFlow installs via Docker Compose. Docker is a tool that packages software and all its dependencies into a self-contained unit you can run on any machine — it's how complex software gets installed without manually managing libraries. If you don't have it, the Docker installation guide covers every major OS.
The hardware requirements are firm: 4 CPU cores, 16 GB RAM, 50 GB disk. A dedicated server or cloud instance is the practical choice for anything beyond experimentation. This is not something you casually run alongside other software on a 16 GB laptop.
The install is three steps: clone the repository, check out a stable tag, run Docker Compose. One extra Linux step — a kernel setting requires a single admin command. First-time startup takes several minutes while images download. Total time from command to first question: 20 to 40 minutes.
One limitation: the official Docker images are built for Intel/AMD processors (x86) only. Apple Silicon Macs (M1 through M4) and ARM-based servers — a different processor family — require building the Docker image from source. The RAGFlow documentation covers the process, but it adds meaningful time.
The UI earns its reputation. You can inspect exactly how each document was chunked, see the pieces the system created, and delete or adjust any chunk that was parsed incorrectly. That transparency — seeing the actual indexed content — is unusual and builds warranted trust in the answers.
Where it fits and what to compare it to
AnythingLLM — easier starting point. Lower hardware requirements, a desktop app, a visual agent builder. Document parsing is text-extraction based, so it does not preserve table structure. Faster to install; less accurate on structured documents.
Dify — broader no-code AI app builder. RAG is one feature among many. Better for building custom AI applications; RAGFlow is deeper if document question-answering is the main job.
Perplexity or ChatGPT with file uploads — cloud services that work immediately. The tradeoff: your documents leave your machine. For anything under NDA or attorney-client privilege, that is a real policy question.
Microsoft 365 Copilot — the enterprise incumbent, tightly integrated with SharePoint and Teams, expensive per seat, infrastructure not in your control. If you want data control and cost discipline, RAGFlow is worth the setup effort.
Verdict
RAGFlow is not the easiest self-hosted document intelligence tool to stand up. The hardware floor is real, the Apple Silicon gap is a friction point for Mac users, and there is no desktop app.
What it offers in return: the most accurate open-source system for asking questions against structured documents. Tables, figures, and layout survive the parsing step. Every answer is cited. You can verify. That verifiability matters in professional contexts where you need to trace an answer to its source, not just trust that the AI probably got it right.
For a legal team, a research function, or an analyst group that has been manually reading through PDFs: the setup overhead is a one-time cost. The ongoing benefit is plain-English questions against a document library that stays entirely on your own infrastructure.
If you need something running in the next hour with lighter hardware: start with AnythingLLM. If document accuracy is the priority and you have an afternoon for setup: install RAGFlow.
For your agent to install this
git clone https://github.com/infiniflow/ragflow.git
cd ragflow/docker
git checkout v0.26.4
# Linux only: sudo sysctl -w vm.max_map_count=262144
docker compose -f docker-compose.yml up -d
Open http://localhost in a browser. Create an admin account, add an AI provider API key under Settings, create a knowledge base, and upload documents.