← 논문 노트 목록
AI

Keyword search is all you need: Achieving RAG-Level Performance without vector databases using agentic tool use

Shreyas Subramanian, Adewale Akinfaderin, Yanyan Zhang, Ishan Singh, Mani Khanuja, Sandeep Singh, Maira Ladeira Tanke

arXiv preprint (AAAI 저작권 표기 — arXiv:2602.23368) (2025) · DOI: 10.48550/arXiv.2602.23368 · 넣은 날 2026-08-05

1. 서지정보

Subramanian, S., Akinfaderin, A., Zhang, Y., Singh, I., Khanuja, M., Singh, S., & Ladeira Tanke, M. (2025). Keyword search is all you need: Achieving RAG-Level Performance without vector databases using agentic tool use. arXiv preprint arXiv:2602.23368.

정확도 메모: 이전에 이 논문을 “AAAI 2026 논문”으로만 불렀는데(이 프로젝트 PRD·세부계획), 실제로 원문을 읽어보니 arXiv에 2025-12-19 게재된 프리프린트이고 저작권 표기만 “Association for the Advancement of Artificial Intelligence”로 돼 있다. AAAI 학회지에 정식 게재됐는지는 이 문서만으로 확인할 수 없어 “arXiv 프리프린트(AAAI 저작권 표기)“로 정정한다.

2. 연구문제

RAG(검색 증강 생성)에 벡터DB·임베딩 기반 의미 검색이 실제로 얼마나 추가 가치를 주는가 — 단순한 키워드 검색만 쓰는 에이전트형 도구 사용으로도 비슷한 성능을 낼 수 있는가?

3. 방법

두 파이프라인을 같은 데이터셋·같은 평가 기준으로 직접 비교했다.

4. 데이터

llamahub에서 가져온 6개 데이터셋(PaulGrahamEssay, Llama2Paper, HistoryOfAlexnet, BlockchainSolana, LLM Survey paper, FinanceBench) — 각각 원문 PDF + 질문 + 정답·근거 문맥이 딸려 있는 표준 RAG 벤치마크.

5. 결과

6. 한계

저자가 밝힌 한계: 큰 문서에서 성능이 떨어지고, 멀티미디어 처리가 제한적이며, 컨텍스트 윈도우 제약이 있다. 모호한 질의 처리와 장기 지식 유지에도 약하다. 자동화된 검색 시스템의 개인정보·윤리 문제(접근 제어·콘텐츠 필터링 등)도 추가 검토가 필요하다고 명시했다(Conclusion).

읽으면서 느낀 한계:

7. 원문 인용

Our empirical analysis demonstrates that tool-based keyword search implementations within an agentic framework can attain over 90% of the performance metrics compared to traditional RAG systems without using a standing vector database. — Abstract

Overall, these results suggest that while the keyword search agent generally performed slightly below the RAG baseline, it maintained competitive performance levels, consistently achieving above 88% average attainment across all three metrics without the use of a semantic search via a vector database. — Results

However, several limitations were identified, including performance degradation with large documents, restricted multimedia handling, and context window constraints. — Conclusion

관련 개념

[[rag]] · [[keyword-search]] · [[llm-agent]]

(아직 만들어진 개념 페이지는 없다. SCHEMA.md 규칙대로 “나중에 쓸 것”이라는 표시로 미리 걸어둔다.)