질의의 감정과 구체성 특성을 반영한 응답 생성을 위한 검색 증강 생성 기법

Abstract

Retrieval-augmented generation (RAG) is an approach that brings external knowledge from large collections or knowledge bases into the response generation process of large language models (LLMs). By adding this external knowledge, RAG helps reduce common problems of LLMs such as hallucination, use of outdated information, and lack of domain coverage. Recently, it has gained attention as a key method to overcome these limitations and improve the reliability of LLM applications. This paper presents a RAG method for response generation that reflects both the emotional and specificity features of queries. Unlike earlier studies that mainly optimized individual stages of the RAG pipeline, the proposed method adopts an integrated approach covering query processing, retrieval, reranking, and generation. Experimental results showed that our method outperformed conventional RAG approaches across various evaluation dimensions. Furthermore, it provides a generalizable structure that can be applied to diverse domains, rather than being limited to a specific application.