RAGFlow: Chat With Your Hardest Documents (GitHub, Scanned)

RAGFlow is a retrieval-augmented generation engine that concentrates on the part most RAG tools skip: getting clean text out of messy documents. Its DeepDoc layer runs layout analysis and OCR over PDFs, scans, slides, spreadsheets and Word files, recognizes tables and figures, and then chunks the result with templates suited to the document type (papers, manuals, contracts, Q&A sheets). You can see each chunk and fix it by hand, and every answer comes back with citations to the passages it used.

Around that it has grown into a full platform: knowledge bases, an agent canvas with MCP and a code-execution sandbox, a memory feature, sync from Confluence, S3, Notion and Google Drive, and chat channels such as Telegram and Discord. It is Apache 2.0, built by InfiniFlow, and mid-migration: Go is now the largest language in the repository, ahead of the original Python. It is also heavy, a stack of containers that wants a real server.

• Repository: github.com/infiniflow/ragflow
• Licence: Apache-2.0 (Apache License 2.0)
• Language: Go. Stars: 91.3K. Forks: 10.8K. Last push: Sep 26, 2026.
• Scan: safe, Sep 24, 2026, commit 313ca90

Who it is for

Teams building question answering over contracts, manuals, research papers or scanned archives where parsing quality decides everything, and self-hosters with a spare x86 server who want citations they can check.

Getting started

1. Raise the Linux memory-map limit that the search engine needs

sudo sysctl -w vm.max_map_count=262144

2. Clone the repository and check out the release matching the Docker image

git clone https://github.com/infiniflow/ragflow.git && cd ragflow/docker && git checkout v0.27.2

3. Start the stack, then open http://localhost and add a model provider

docker compose -f docker-compose.yml up -d

Plan for at least 4 CPU cores, 16 GB of RAM and 50 GB of disk. The prebuilt images are x86 only; on ARM, including Apple Silicon, the docs explain how to build your own. The code-execution sandbox additionally needs gVisor.

Safety scan

We cloned infiniflow/ragflow at commit 313ca90 on Sep 24, 2026 and ran the checks described on the GitHub Tools page: credential patterns, decode-and-execute code, install-time scripts, committed binaries, risky CI workflows, every host the code talks to, known vulnerabilities in pinned dependencies, and project hygiene. A person read every hit. This is what we found.

• Four secret hits, all sample SSH private keys in internal/agent/sandbox/ssh_test.go. No self-decoding code. The bare-IP URLs are SSRF tests for the agent tools and data connectors, checking which addresses are blocked.
• Three pattern hits: the Dockerfile installs Node.js with NodeSource's curl | bash setup script at build time, a Telegram URL sits in a test, and the reverse-shell regex belongs to the skill hub's security scanner, which looks for exactly that.
• tools/scripts/install.sh installs the optional ragflow-cli binary from GitHub releases into /usr/local/bin, downloads a SHA256SUMS file alongside it, and asks for sudo only if that folder is not writable. The server itself installs through Docker. No npm hooks, no committed binaries.
• Known advisories: 210 across 2,905 packages in nine lockfiles. The one critical (shell-quote) and 72 highs are in the web front end's npm build. The main uv.lock has 8 highs among 697 packages, and the Python SDK's lockfile 13 highs among 26.
• Only three workflows. The scanner's pull_request_target hit in sep-tests.yml is the name appearing in comments and conditions: the workflow is triggered by plain pull_request and runs only when a maintainer adds the ci label. Neither third-party action is pinned. Security policy and licence present; no Dependabot, CodeQL or contributing guide.
• docker/.env ships default passwords (infini_rag_flow and variants) for Elasticsearch, MySQL, MinIO and the rest, under a warning at the top of the file to change them before any non-local deployment.

What the scanner counted

The raw findings

Every hit the scanner wrote out, with a link to the exact line at the scanned commit. Secrets candidates are redacted.

By the numbers

Lines by language: Go 856K, Python 422.8K, TypeScript 257.9K, JSON 125.9K, Markdown 45.4K, C++ 44.5K.

Questions

Is RAGFlow free?

Yes. RAGFlow is Apache 2.0, so self-hosting, modifying and commercial use are all allowed without a fee. InfiniFlow also runs a hosted version at cloud.ragflow.io for people who do not want to operate the stack. Either way you supply the language and embedding models, hosted or local.

What hardware does RAGFlow need?

The README asks for at least 4 CPU cores, 16 GB of RAM, 50 GB of disk and Docker 24 or later, because the stack runs a search engine, a database, object storage and the parsing service side by side. The prebuilt images are x86 only. A GPU is optional and speeds up document parsing.

How is RAGFlow different from Dify or AnythingLLM?

All three do retrieval over your documents. RAGFlow puts most of its effort into parsing: layout analysis, OCR, table recognition and chunk templates, with the chunks visible and editable. Dify is a broader app builder and AnythingLLM a lighter desktop chat. If answers depend on complex PDFs or scans, RAGFlow is usually the stronger choice.

This post is part of GitHub Tools, where every repository is cloned and scanned before it is written up. The scan is a snapshot of one commit on one day; the repository has moved on since, so check it before you install.