RAGFlow
infiniflow · AI search
GitHub stars
93,091Awaiting first sync
No measured star history yet
What is RAGFlow?
RAGFlow is built for teams whose knowledge sits in messy files: scanned PDFs, slide decks, spreadsheets and manuals. Its parser reads layout, tables and images, and every chunk it creates can be viewed and corrected before a chat assistant uses it. Answers come with references back to the source passages.
RAGFlow is an open-source retrieval-augmented generation engine. You upload documents such as PDFs, Word files, slides and scans, it splits them into chunks you can inspect, and chat assistants answer from them with traceable citations. A visual agent canvas and an API connect it to other tools.
Core capabilities
- Parse PDFs, Word, slides, spreadsheets, images and scans into chunks you can view and edit.
- Pick a chunking template that fits the document, such as manuals, papers, laws, slides or Q&A.
- Connect hosted models or local Ollama and vLLM models, and build agents on a visual canvas.
See and adjust how documents are split
Each dataset uses a built-in template, including General, Q&A, Manual, Table, Paper, Book, Laws, Presentation, One, Audio and Email. For PDFs, the default DeepDOC parser handles OCR, table structure and page layout, and you can switch to a plainer parser, choose page ranges or turn on a vision model to describe images and tables. Parsed chunks can be read, edited and given keywords, then tried out in Retrieval testing.
- Document parsing
- Chunking
- OCR
- Retrieval testing
Chat with citations, agents and your choice of models
A Chat links one or more datasets to a language model and a system prompt, and shows citations with its answers. For multi-step jobs, the Agent canvas connects components such as Retrieval, Categorize, HTTP requests, MCP tools and code, and a workflow agent can be embedded in a web page. Model providers are set up under User settings: use hosted APIs, OpenAI-compatible endpoints, or local models through Ollama or vLLM. An HTTP API lets other applications call it.
- Chat
- Agents
- Ollama
- API
Run it with Docker Compose
RAGFlow is licensed under Apache-2.0. The docs recommend a starting point of 4 CPU cores, 16 GB RAM and 50 GB of free disk, with more needed for local models, and Docker 24 or newer with Compose 2.26.1 or newer. Images cover Linux x86-64 and ARM64 from v1.0.0-rc2, and the README says macOS is not currently supported. The default stack uses Elasticsearch, MySQL, MinIO, Kvrocks, NATS and ClickHouse. Change the default passwords before exposing it.
- Self-hosted
- Apache-2.0
- Docker Compose
- Hardware
Where it fits
Use cases
- 01
Build a searchable policy and manual library
Create a dataset, choose a parsing template such as Manual or Laws, and upload your PDFs and Word files. Open the Chunk page to check the results, then create a Chat that answers staff questions with citations.
- 02
Check why an answer came out wrong
Open the parsed file to see each chunk, add keywords or questions to the ones that matter, and edit text by hand. Run Retrieval testing with a sample question to confirm the right chunks now come back.
- 03
Run a support agent over product material
On the agent canvas, add a Categorize step to sort incoming questions, a Retrieval step to look up product documents, and an HTTP request to call your ticket system. Embed the finished agent in your website as a chat widget.