HyGRAG: New Framework Enhances RAG with Context and Relation-Aware Graphs

This research introduces HyGRAG, a novel hierarchical graph-based framework for Retrieval-Augmented Generation (RAG) that aims to overcome limitations of existing graph-based RAG methods. Traditional approaches, whether entity-centric or chunk-centric, often retrieve information based on similarity without truly synthesizing contextual and relational knowledge, thus missing emergent understanding. HyGRAG addresses this by building hierarchical index structures using hybrid graphs that incorporate both chunk and entity nodes. It iteratively clusters these nodes and generates LLM-based summaries, creating representations that genuinely integrate both contextual and relational information. This allows for retrieval that is aware of both context and relationships, searching across various abstraction levels and expanding through community membership. Furthermore, the framework supports efficient dynamic knowledge updates through attachment-based algorithms, requiring only local re-summarization. Experimental results demonstrate that HyGRAG significantly improves the average accuracy of multi-hop reasoning tasks by 9.7%, while maintaining reasonable computational efficiency.