Cog-RAG: Cognitive-Inspired Dual-Hypergraph with Theme Alignment Retrieval-Augmented Generation

ABSTRACT

Cog-RAG introduces a cognitive-inspired dual-hypergraph architecture designed to capture both thematic and entity-level relationships within large document collections. The framework constructs two interconnected hypergraphs: a theme graph representing global conceptual structures and an entity graph capturing fine-grained semantic relationships. Retrieval occurs through a two-stage process that first activates thematic context and then performs entity-level diffusion. This structure ensures better semantic alignment between retrieved evidence and generated responses. Experiments show strong improvements over graph-based and traditional RAG methods.