← Back to all projects

ragflow

A self-hosted RAG engine that shows you every chunk — a document platform, not a library you import

OfficialApache-2.0
Stars
93.2k
Forks
11.2k
Open issues
1.3k
Last commit
30 Oct 2026

What is ragflow?

RAGFlow is a server you run rather than a library you import: documents go into a dataset, it splits them into chunks, and you can read every chunk on screen and hand-correct the ones that came out wrong before any of it reaches a model, with chat answers citing the chunk they came from. Around that sit an agent canvas and an ingestion pipeline you can rebuild yourself, so retrieval, tools and MCP servers can be wired together without writing code. The optical character recognition, table-structure and layout models that make sense of scans run on PDFs and images only — Word, Excel and PowerPoint files are read structurally, and the docs tell you to convert a DOCX to PDF if you want the visual parser on it. The price of all this is weight: a Docker Compose stack with Elasticsearch, MySQL, MinIO and Redis behind it, 4 cores and 16 GB of RAM as the stated minimum, and parsing that the project's own FAQ concedes is slower than LangChain's. If your documents are already clean text and you want retrieval inside your own Python application, LlamaIndex or Haystack will fit better.

What can you do with ragflow?

  • Point it at documents wherever they already sit — Files can be uploaded by hand, or pulled in from roughly thirty-five outside systems the docs group into seven kinds: team wikis and drives (Confluence, Notion, Google Drive, SharePoint, Dropbox), object storage (S3, Azure Blob, R2), databases and warehouses (MySQL, PostgreSQL, BigQuery), code and project trackers (GitHub, GitLab, Jira), mail and chat (Gmail, Outlook, Slack, Teams, Discord), business systems (Salesforce, Zendesk, Airtable) and plain REST or RSS feeds. Each connection carries its own refresh interval, an option to drop content deleted upstream, and a log of every sync it has run.
  • Choose how each document gets cut into chunks — A dataset picks a chunk template so a statute is not split the way a slide deck is: General cuts at a preset token count, and Paper, Book, Laws, Manual, Presentation, Table, Q&A, Picture and One are among the templates that follow the shape of a particular kind of document instead; an individual file can override the dataset's choice. For PDFs the parser is a second, separate choice — DeepDoc, the default, runs the character-recognition, table-structure and layout models and is slow; Naive skips them when the text is already clean; MinerU, Docling and OpenDataLoader are offered as experimental alternatives — and that dropdown is for PDFs only, so a DOCX has to be converted first.
  • Read the chunks, and correct the ones that came out wrong — Parsing results are laid out chunk by chunk; double-click one to rewrite the text, or attach keywords, questions and tags, where keywords raise that chunk's ranking for queries containing them. Individual files can be switched off without being deleted.
  • Tune retrieval on a test page before wiring up a chat — The retrieval test runs a real query against the dataset and shows what comes back. A similarity threshold, 0.2 by default, drops weak chunks; a vector similarity weight, 0.3 by default and so leaving 0.7 for keyword matching, shifts the balance between meaning and exact words; and a rerank model can take the place of vector scoring, which the docs warn significantly increases response time. Nothing here is saved for you — settings that worked have to be entered again in the chat assistant, or in the Retrieval component on an agent canvas.
  • Add a knowledge graph when an answer spans several documents — Entities and relationships are extracted into one graph across the whole dataset rather than per file, which is what makes multi-hop questions work, and the graph updates as newly uploaded files are parsed. The extraction prompts come from either GraphRAG (General) or LightRAG (Light, the default and the cheaper of the two), and RAGFlow's docs warn that building the graph takes significant memory, computational resources and tokens.
  • Build agents on a canvas, or drive the whole thing from code — Begin, Agent, Retrieval, Categorize, Switch, Iteration, Code and HTTP Request components connect into a flow, where retrieval is either a fixed step or a tool the Agent decides to call, alongside SQL, MCP servers and sub-agents. The same canvas builds ingestion pipelines out of Parser, Title chunker, Token chunker, Transformer and Indexer components, which a dataset can then use in place of a built-in template. Outside the web UI there are HTTP and Python APIs, plus a separate RAGFlow MCP server so somebody else's agent can search your datasets.

Before you choose ragflow

  • Official Docker images are built for x86 only: as of v0.26.4 the README and FAQ tell anyone on ARM64, an Apple Silicon Mac included, to build their own image, and Infinity as the document engine is unsupported there.
  • A dataset's embedding model is fixed once it holds chunks, so moving to a different one means deleting every chunk in that dataset and parsing all of its files again.
  • Not everything shown is in the open-source edition: as of v0.26.4 the docs mark the Resume chunking template Enterprise-only, and the FAQ says cloud.ragflow.io runs Enterprise with DeepDoc models trained on private data.

Star history

17 Aug to 30 Oct · +4.5k

88.7k93.2k

Frequently asked questions

Is ragflow free for commercial use?

ragflow is released under the Apache-2.0 licence — OSI-approved open source, which permits commercial use.

How can ragflow be deployed?

ragflow is available as Self-hosted / Managed cloud.