HyDE 기반 멀티 홉 검색 기법을 활용한 검색 성능 향상 방안

The development of generative AI and large language models (LLMs) is revolutionizing business processes and fostering the adoption of customized LLMs for domain-specific tasks. However, Retrieval-Augmented Generation (RAG) systems designed to address hallucination and domain relevance issues face performance degradation due to error accumulation in handling complex multi-hop queries. This study proposes a framework that integrates the Hypothetical Document Embedding (HyDE) technique into multi-hop retrieval to enhance RAG performance. HyDE generates virtual documents reflecting query intent, improving retrieval accuracy by decomposing complex queries into single-hop queries and applying HyDE iteratively at each step. Experiments were conducted using precision@k, recall@k, F1 score, MAP, MRR, and hit rate as evaluation metrics. The results demonstrate that HyDE-based multi-hop retrieval improves performance across all metrics, with recall increasing by approximately 19.53% and hit rate by 21.21%. These findings confirm the effectiveness of HyDE in enhancing retrieval accuracy for multi-hop search. Future research directions include the development of optimized datasets for multi-hop retrieval and further refinement of virtual document generation strategies.