\(\text {H}^{2}\) RAG: A Hierarchical Knowledge and Hypergraph Reasoning Framework for Retrieval-Augmented Generation

Graph- and hypergraph-based retrieval-augmented generation (RAG) methods have significantly enhanced the performance of large language models (LLMs) by incorporating structured knowledge representations. However, existing approaches often suffer from semantic dilution among distantly related yet semantically similar entities. To address this issue, we propose Hierarchical Knowledge Hypergraph Retrieval-Augmented Generation ( \(\text {H}^{2}\) RAG), a novel framework that models higher-order relationships through hypergraphs, applies the Kumar community detection algorithm with bounded recursion to regulate community size, and integrates a hierarchical summary layer to improve query matching and response coherence. Unlike conventional graph-based RAG systems, \(\text {H}^{2}\) RAG adopts a dual-layer retrieval strategy encompassing hyperedge similarity, node–entity matching, community probing, and summary alignment, thereby mitigating semantic drift and capturing inter-document dependencies. Experiments on the UltraDomain dataset across Agriculture, Computer Science, Legal, and Mix domains demonstrate that \(\text {H}^{2}\) RAG surpasses strong baselines such as HyperGraphRAG and HiRAG, achieving average improvements of \(20.33\%\) in Exact Match, \(14.61\%\) in F1, and \(11.53\%\) in Generalized Score. Ablation study and sensitivity analysis further verify the complementary effects of its modules and the balance between precision and computational cost.