Abstract:Graph-based Retrieval-Augmented Generation (RAG) methods have significantly enhanced the performance of large language models (LLMs) in domain-specific tasks. However, existing RAG methods do not adequately utilize the naturally inherent hierarchical knowledge in human cognition, which limits the capabilities of RAG systems. In this paper, we introduce a new RAG approach, called HiRAG, which utilizes hierarchical knowledge to enhance the semantic understanding and structure capturing capabilities of RAG systems in the indexing and retrieval processes. Our extensive experiments demonstrate that HiRAG achieves significant performance improvements over the state-of-the-art baseline methods.
Submission history
From: Haoyu Huang [view email]
[v1]
Thu, 13 Mar 2025 08:22:31 UTC (21,971 KB)
[v2]
Mon, 23 Jun 2025 11:08:00 UTC (21,952 KB)
[v3]
Fri, 26 Sep 2025 14:37:04 UTC (2,876 KB)