상세 보기
초록
This paper proposes DRAG, a Retrieval-Augmented Generation (RAG) method based on query-conditioned Retrieval Units. Existing RAG systems often rely on fixed-size chunking or sliding-window methods, which do not sufficiently reflect query semantics and document structure. DRAG reformulates document segmentation as a query-conditioned retrieval unit construction task and dynamically constructs Retrieval Units by considering both query semantics and document structure. Experimental results show that DRAG improves Recall@20 by 9.0% and reduces the Redundant Context token Ratio (RTR) by 14.1% compared with Fixed Chunking, while improving F1 Score by 7.0% compared with Semantic Chunking. These results demonstrate that the proposed method effectively enhances both retrieval performance and context utilization efficiency.
키워드
- 제목
- 질의 조건 기반 Retrieval Unit을 적용한 검색 증강 생성 기법
- 제목 (타언어)
- A RAG Method based on Query-Conditioned Retrieval Units
- 저자
- 김영아; 정선미; 김현주; 김창근
- 발행일
- 2026-08
- 유형
- Y
- 저널명
- 한국정보기술학회논문지
- 권
- 24
- 호
- 8
- 페이지
- 69 ~ 80
- 언어
- KOR
- 출판사
- 한국정보기술학회
- 발행국가
- 대한민국
- 분량
- 12 페이지
- ISSN
- E 2093-7571
P 1598-8619