RAGFlow is a retrieval-augmented generation (RAG) engine combined with an agent platform. It turns complex, unstructured documents into a context layer for LLMs: it parses and chunks source files, indexes them for hybrid retrieval, and answers questions with traceable citations. Agents and visual workflows built on top of it combine retrieval with tools, web search, HTTP calls, and MCP servers.
Ingestion starts with deep document understanding of complicated formats such as Word, slides, Excel, plain text, images, scanned copies, structured data, and web pages. Chunking follows selectable templates, and the chunking result is visualized so a person can inspect and adjust it. Retrieval combines vector search, BM25 full-text search, and custom scoring with fused re-ranking, and LLMs, embedding models, and rerankers are configurable per deployment.
Agentic retrieval breaks a complex question into steps, retrieves knowledge over multiple rounds, and verifies evidence before answering, with Low, Medium, High, and Ultra thinking modes that set the depth. Knowledge Compilation turns documents and datasets into structured artifacts: Wikis, Graphs, Trees, PageIndex, Mind Maps, Timelines, and Skills.
The backend is a set of Go services (Admin, Ingestor, Syncer, and API) with a React web interface, deployed with Docker Compose or run as a hosted cloud service. The document engine defaults to Elasticsearch, with Infinity as an alternative, and data sources include website sitemaps and Google BigQuery with incremental synchronization.
Features
- Deep document understanding: extracts knowledge from unstructured files with complicated layouts, including scanned copies and images
- Template-based chunking: a choice of chunking templates, with visualized chunks that a person can review and correct
- Hybrid retrieval: vector search, BM25 full-text search, and custom scoring combined with fused re-ranking
- Agentic retrieval: multi-step retrieval that decomposes questions and verifies evidence, with four thinking modes from Low to Ultra
- Knowledge Compilation: generates Wikis, Graphs, Trees, PageIndex, Mind Maps, Timelines, and Skills at the document and dataset levels
- Grounded citations: answers show key references and traceable citations
- Agent workflows: visual workflows that combine retrieval, web search, HTTP calls, code, conditions, subagents, and MCP tools
- Data sources: file uploads, website ingestion through sitemaps, and Google BigQuery with incremental sync
- Configurable models: LLMs, embedding models, and rerankers are added per deployment on a model provider page
- APIs: an API service for the web interface, SDKs, and external clients
