RAGFlow
Parses mixed-format documents, retrieves with reranking, and returns cited answers; self-hosting needs substantial resources.
What it is
RAGFlow is an open-source RAG engine that connects Word, Slides, Excel, scans, images, web pages and structured data to AI question answering. A typical operation is to upload material, choose a parsing and chunking method, review the knowledge chunks the system generates, and then let the model retrieve the relevant content, rerank the results and generate answers with source citations.
Compared with solutions that only offer a chat interface or simple vector retrieval, RAGFlow puts complex document parsing, knowledge chunking, retrieval reranking, citation tracing and Agent workflows into one system. Operators can inspect chunks directly and correct parsing results, which suits production scenarios that require checking knowledge boundaries and answer provenance.
How to run it
The shortest path is to clone the repository, switch to a specified version, and then start the Docker Compose services with prebuilt images:
git clone https://github.com/infiniflow/ragflow.git
cd ragflow/docker
git checkout v0.27.1
docker compose -f docker-compose.yml up -d
docker logs -f docker-ragflow-cpu-1After the service finishes initializing, visit the server address in a browser. Then you still need to select an LLM vendor in the configuration and fill in the corresponding API Key.
Core mechanisms
- Deep document parsing: extracts knowledge from unstructured material with complex layouts, and can also use multimodal models to understand images inside PDFs or DOCX files.
- Template-based chunking: choose a chunking template according to the material type, and visualize the results so that knowledge boundaries can be adjusted by hand.
- Fusion retrieval: combines multiple recall routes with reranking and attaches key citations to the answer, making it easy to go back to the original material and verify.
- Orchestratable pipelines: ingestion pipelines, RAG workflows and Agent templates are all configurable, and business systems can also integrate through the API.
Who it is for
It suits teams that need to connect heterogeneous material such as contracts, reports, tables and scanned documents to internal question answering, especially scenarios where chunk quality and answer provenance must be checked. It can also serve as the retrieval and context layer of a self-built AI application, rather than just a chat interface.
Usage boundaries
The repository uses the Apache-2.0 license. The minimum self-hosting requirement is 4 CPU cores, 16 GB memory and 50 GB disk, and it depends on Docker 24, Docker Compose 2.26.1 and Python 3.13. The prebuilt images target x86, and ARM64 requires building them yourself. By default the system uses Elasticsearch to store full text and vectors, and it can also switch to Infinity. The chosen LLM and embedding service still need to be configured separately, and the code execution sandbox additionally depends on gVisor.