본문 바로가기
  • Home

Comparative Analysis of BM25 and Embedding Retrieval Performance by Question Type in a Records Schedule and Regulations-Based RAG(Retrieval-Augmented Generation) Question Answering System

  • Journal of the Korean Society for Library and Information Science
  • 2026, 60(3), pp.125~143
  • DOI : 10.4275/KSLIS.2026.60.3.125
  • Publisher : 한국문헌정보학회
  • Research Area : Interdisciplinary Studies > Library and Information Science
  • Received : July 16, 2026
  • Accepted : August 5, 2026
  • Published : August 31, 2026

Yuna Choi 1 PARK JONG DO 2 Haein Lee 1 CHO JANE 2

1인천대학교 문헌정보학과
2인천대학교

Excellent Accredited

ABSTRACT

The Retrieval-Augmented Generation structure is a technology that generates responses by referencing a knowledge base, thereby supporting reliable responses. However, the quality of answers from a RAG structure can vary depending on the search performance of the knowledge base. In particular, when the knowledge base consists of norms in various specific domains composed of structured attributes or domain specific vocabulary, it is difficult to assume that embedding-based methods will unconditionally guarantee superior performance. Thereby, this study created a knowledge base based on the rules and regulations of a university's archive. Then, two different systems using the same LLM were built, with two different retrieval methods, BM25 and embedding. And then, the study compared and evaluated how search methods can vary depending on the user's query type and the structure of the knowledge base. Experimental results showed that BM25 demonstrated superior performance direct-answer and error-premising queries, where exact keyword matching is crucial, while the embedding method showed superior performance in inference and term-variation queries.

Journal Copyright Policy

BY-NC-ND 아이콘

This license enables reusers to copy and distribute the material in any medium or format in unadapted form only, for noncommercial purposes only, and only if attribution is given to the creator.

Citation status

* References for papers published after 2025 are currently being built.