Full-Code · Orchestration Frameworksactive

RAGFlow

Type: full-code · Vendor: InfiniFlow · Language: Python, TypeScript · License: Apache-2.0 · Status: active · Status in practice: mature

Links: homepage docs repo

Open-source RAG engine that pairs deep document-understanding (DeepDoc) layout-aware parsing with an agentic, graph-orchestrated workflow runtime, MCP support, and an extensive citation/traceability surface.

Description. RAGFlow (78k+ GitHub stars, Apache-2.0) is InfiniFlow's open-source RAG engine. Its defining feature is DeepDoc — a layout-aware, template-driven document parser that handles PDF, Word, Excel, slides, scanned copies, structured data, and web pages with format-specific chunking strategies rather than naive splitting. From v0.8 onward RAGFlow became agentic: a graph-based workflow editor lets users compose retrieval + reasoning + tool stages, with MCP support, GraphRAG, and (from late 2025) memory for AI agents. Citations are first-class and traceable to source chunks with visualisations.

Agent loop shape. Document ingestion runs DeepDoc parsing → template-based chunking → embedding → indexing (with optional GraphRAG knowledge-graph construction). At query time, a no-code graph workflow defines retrieval + rerank + reasoning + tool stages. Each query enters the graph at the start node and traverses with branch/loop semantics. MCP exposes retrieval to external agents.

Primary use cases

enterprise document QA over complex format mixes (PDF, Office, scans, structured data)
agentic RAG with no-code workflow composition over retrieval + reasoning + tools
high-fidelity citation rendering with chunk-level provenance to source
MCP-server integration so external agents can call RAGFlow's retrieval as tools

flowchart TD subgraph Ingest DOC[(Documents<br/>PDF/Word/Excel/scans)] --> DEEPDOC[DeepDoc parser<br/>layout-aware, template-based] DEEPDOC --> CHUNK[Per-format chunking] CHUNK --> IDX[(Index:<br/>vector + sparse + optional KG)] end subgraph Query Q[User query] --> WORKFLOW[Graph workflow editor<br/>no-code DAG] WORKFLOW --> RET[Retrieval node] RET --> RERANK[Rerank node] RERANK --> REASON[Reasoning / generator] REASON --> CITE[Citation node] CITE --> ANS[Answer + traceable citations] end IDX --> RET MCP[MCP server] -.expose retrieval.-> EXT[External agents]

Key concepts

DeepDoc (docs) — Layout-aware, template-driven parser for complex document formats.
Graph workflow editor → orchestrator-workers — No-code DAG for composing RAG + agent stages.
Template-based chunking — Per-format chunking templates instead of naive token splits.
Citation rendering → citation-attribution — Traceable chunk-level citations attached to answers.