RAGFlow
InfiniFlow's Apache 2.0 RAG engine built on deep document parsing. DeepDoc handles layout, tables and OCR before chunking, and every citation points back to a highlighted region in the source PDF.
Overview
InfiniFlow built RAGFlow around a claim most retrieval frameworks avoid making out loud: retrieval quality is bounded by parsing quality. The DeepDoc component runs layout analysis, table structure recognition and OCR before anything reaches a chunker, and the chunking itself is template-driven, so a resume, a research paper, a regulatory filing and a product manual each get handled differently instead of by one fixed token window. Retrieval defaults to Elasticsearch for hybrid keyword and vector search, with InfiniFlow's own Infinity engine available as the alternative backend.
The licence is Apache 2.0, unmodified, with no feature gating and no per-seat cost. Worth stating plainly, because several products in this category reserve the parts you actually need for a paid tier. Every answer returns citations that point back to a highlighted region in the source document, which is what ends arguments in a compliance review rather than starting them. Recent work has pushed into agentic territory: v0.27.0 on 24 August 2026 introduced knowledge compilation at document and dataset level, and the 2026 releases added orchestrable ingestion pipelines alongside connectors for Google BigQuery and Azure DevOps.
Resourcing deserves attention. The documented minimum is four CPU cores, 16 GB of RAM and 50 GB of disk with Docker 24.0.0 or later, and that is a floor rather than a target once Elasticsearch is running alongside the parser models. First ingestion of a large scanned corpus is slow on CPU, so reach for the GPU image if you have a card free. Version numbers are still 0.x, currently v0.27.1 from 28 August 2026, and the API has moved between minor releases, so pin it. English deployment documentation is solid. The deeper knowledge about tuning parsers and writing chunk templates still lives mostly in Chinese GitHub issues and the project blog.
Key Features
- ✓ DeepDoc parsing stack running layout analysis, table structure recognition and OCR before chunking, tuned heavily on Chinese-language documents including scanned finance and government forms
- ✓ Template-based chunking per document type, so a resume, a research paper and a technical manual are each split by a different strategy rather than one fixed token window
- ✓ Grounded citations that highlight the exact source region in the original PDF, visible in the UI for verification
- ✓ Elasticsearch by default for hybrid keyword and vector retrieval, swappable for InfiniFlow's own Infinity engine as the storage backend
- ✓ Agentic workflow builder plus orchestrable ingestion pipelines, with document-level and dataset-level knowledge compilation added in v0.27.0
- ✓ Docker Compose deployment with a slim image option, a growing set of data-source connectors including Google BigQuery and Azure DevOps, and a Python client published as ragflow-sdk on PyPI
Where it holds
- • Plain Apache 2.0 with nothing held back for a paid tier, which several comparable RAG products cannot say
- • 89,500 stars and roughly monthly releases, with v0.27.1 landing 28 August 2026
- • Chinese document handling is measurably better than the Western RAG stacks, which is the entire reason to pick it
- • Citation grounding is designed in rather than bolted on, and it shortens compliance conversations considerably
Where it breaks
- • The default Docker Compose stack pulls Elasticsearch, so four cores, 16 GB RAM and 50 GB disk is the documented floor rather than a comfortable target
- • Still on 0.x version numbers and the API surface has shifted between minor releases. Pin your version in production
- • Initial ingestion of a large scanned corpus is slow, because the layout and OCR models run per page. Use the GPU Docker image if you have the hardware
- • English docs cover deployment properly, but deeper parser tuning and chunking-template authoring is explained far better in the Chinese issues and blog posts
My Take
Document parsing is the hard part of RAG and it is where most frameworks quietly give up. RAGFlow does not. DeepDoc runs layout recognition, table structure detection and OCR over PDFs, slides and scanned forms before anything is chunked, then shows the source region behind every citation, which is the feature that gets a Chinese enterprise procurement team to sign off. Apache 2.0, no feature gates, v0.27.1 tagged 28 August 2026. Budget at least four CPU cores, 16 GB of RAM and 50 GB of disk, because the default stack brings Elasticsearch with it.
Quick Info
- Pricing:
- open-source
- Openness:
- Open source
- Licence:
- Apache 2.0
- Starting at:
- Apache 2.0 and free to self-host, with no feature gates, no per-seat pricing and no enterprise-only retrieval tier. Your only cost is infrastructure plus whatever you spend on model API calls to DeepSeek, Qwen or an OpenAI-compatible endpoint. InfiniFlow runs a hosted service at cloud.ragflow.io for teams that prefer not to operate the stack. Self-hosting floor is four CPU cores, 16 GB RAM and 50 GB disk with Docker 24.0.0 or newer.
- Added:
- Aug 2026
- Updated:
- Aug 2026
Use Cases
Judge it on your own work
The notes above say where RAGFlow holds and where it breaks. The fastest check is your own workload.
Visit website ↗Alternatives to RAGFlow
Dify
freemiumOpen-source LLM app platform from LangGenius for agentic workflows and RAG pipelines. Modified Apache 2.0: self-host freely, multi-tenant SaaS needs a commercial licence.
LlamaIndex
freemiumOpen-source Python and TypeScript framework for RAG and agentic workflows, paired with LlamaCloud for document parsing, extraction and managed indexes.
MinerU
open-sourceDocument-to-Markdown engine from Shanghai AI Lab's OpenDataLab. Relicensed off AGPLv3 in April 2026, reported at 95.4 on OmniDocBench v1.6 and running on ten domestic AI chip families.
Glean
paidEnterprise search and agent platform that indexes company systems with permissions intact, then exposes that context to internal agents and outside AI clients.