FlowRAG Enhances Graph-Based RAG with Multi-Granularity Reasoning.

Graph-based Retrieval-Augmented Generation (GraphRAG) systems are effective for complex, knowledge-intensive queries, but often struggle with abstract queries and brittle multi-hop reasoning due to reliance on implicit semantic relevance. Existing methods frequently under-retrieve information when user queries are semantically sparse at the entity level and can be derailed by noisy activations during entity-to-entity transitions, leading to unreliable conclusions. To overcome these challenges, FlowRAG introduces a semantic-aware retrieval framework designed to enhance both semantic recall and explicit reasoning capabilities. This framework builds a quad-level heterogeneous graph that integrates passages, summaries, sentences, and entities, with summary nodes acting as crucial semantic hubs. A dual-granularity activation module then combines summary-query alignment with sentence-level matching to robustly activate relevant entities, even under paraphrasing or abstraction. A core innovation is the frequency-aware weighted flow module, which routes relevance through entity-passage links. These links are weighted by within-passage term frequency, effectively pruning noisy connections and extracting high-confidence reasoning paths. These paths serve as an explicit logical skeleton for the generation process. Extensive experiments confirm that FlowRAG achieves state-of-the-art performance on complex reasoning benchmarks.