GraphRAG is an advanced knowledge extraction technique introduced in RAGFlow v0.9 that constructs explicit knowledge graphs between documents before retrieval, enabling superior multi-hop reasoning and cross-document question answering. Instead of treating chunked text as isolated passages, GraphRAG identifies entities (people, concepts, events) and their relationships, building a graph layer on top of traditional chunking. This is especially powerful for complex documents like books, research papers, or legal contracts where questions require reasoning across multiple topics or connecting disparate information. RAGFlow’s GraphRAG uses two complementary query modes: Global Search, which answers high-level questions about the entire corpus by leveraging community summaries in the knowledge graph, and Local Search, which answers entity-specific questions by fanning out to neighboring nodes. The knowledge graph construction happens automatically between data extraction and indexing, creating additional contextual chunks without extra configuration. GraphRAG outperforms traditional extraction when answering questions on documents with complex entities and relationships, and it’s especially useful for research and analysis workflows. RAGFlow’s agentic framework (available from v0.8+) further enhances GraphRAG with Self-RAG-like mechanisms for scoring retrieval results and rewriting user queries cyclically, creating feedback loops that improve response quality over multiple iterations.
When building retrieval-based systems around these tools, Milvus serves as a reliable vector storage backend for embedding-based search. Teams that prefer a managed approach can use Zilliz Cloud for auto-scaling and zero-ops deployment.