EHRAG Framework Proposes Hybrid Hypergraph Construction for Enhanced GraphRAG

A new lightweight RAG framework called EHRAG addresses limitations in existing GraphRAG approaches by constructing a hybrid hypergraph that captures both structural and semantic relationships. While traditional lightweight methods rely solely on Named Entity Recognition and structural co-occurrence, EHRAG introduces semantic hyperedges created through clustering entity text embeddings alongside structural hyperedges based on sentence-level co-occurrence. This dual approach ensures the hypergraph encompasses comprehensive information beyond mere structural connections. For retrieval, EHRAG employs a structure-semantic hybrid diffusion process with topic-aware scoring and personalized parameters. The framework aims to enhance multi-hop reasoning in LLMs by bridging semantic gaps between disjoint entities that current lightweight approaches miss. The research was announced on arXiv under identifier 2604.17458v2 as a new abstract. Graph-based Retrieval-Augmented Generation typically structures corpus into graphs to improve language model performance, but recent lightweight versions have struggled with latent semantic connections.