RAGFlow is a retrieval-augmented generation engine that turns unstructured documents into a queryable context layer and pairs it with an agent runtime. Ingestion applies deep document understanding to files whose formats resist plain text extraction, chunks them with selectable templates, and keeps the chunking visible so a person can intervene; answers come back with traceable citations to the passages behind them.
Retrieval combines vector search, full-text search, and tensor scoring with re-ranking, and the surrounding platform adds an orchestrable ingestion pipeline, visual agent workflows, pre-built agent templates, MCP support, and a Python or JavaScript code-executor component. Documented workflow examples cover stock analysis with multi-agent orchestration, legal precedent analysis, and maintenance guidance sourced from internal manuals.
Input handling covers Word, slides, Excel, text, images, scanned copies, structured data, and web pages, with MinerU and Docling available as parsing methods and multi-modal models used to interpret images inside PDF or DOCX files. Data can be synchronized from Confluence, S3, Notion, Discord, and Google Drive, and chat channels include Feishu, Discord, Telegram, and Line.
Self-hosting runs through Docker Compose and needs at least 4 CPU cores, 16 GB of RAM, 50 GB of disk, and Python 3.13; the published images target x86, and ARM64 hosts build their own. Elasticsearch stores full text and vectors by default, and setting DOC_ENGINE switches that to Infinity. A hosted cloud service is also offered, with paid tiers for larger teams and dataset storage.
Features
- Deep document understanding: knowledge extraction from unstructured files with complicated formats, including scanned copies and images
- Template-based chunking: several chunking templates, with the result visualized so a person can correct it
- Grounded citations: answers link back to the key references that support them
- Hybrid retrieval: vector, full-text, and tensor scoring combined with re-ranking
- Agent platform: visual workflows, pre-built templates, memory for agents, MCP, and a Python or JavaScript code executor
- Ingestion pipeline: an orchestrable pipeline plus synchronization from Confluence, S3, Notion, Discord, and Google Drive
- Parsing choices: MinerU and Docling can be selected as document parsing methods
- Configurable models: the LLM factory and embedding models are set per deployment through the service configuration
- Doc-engine switch: Elasticsearch by default, with Infinity selectable through an environment variable
- Chat channels: Feishu, Discord, Telegram, and Line among the supported channels
