RAG 기반 LLM을 적용한 국방기술 연구개발 과제 기획 지원 생성형 AI 시스템 개발에 대한 연구

다국어 초록 (Multilingual Abstract)

Generative AI models have rapidly advanced in recent years, enabling creative and productive use of information across various domains. At the center of this progress are Large Language Models (LLMs), which have evolved into powerful tools capable of search, analysis, and task automation. However, their increasing adoption has highlighted the persistent challenge of hallucination—instances where models generate inaccurate or unverifiable information. Retrieval-Augmented Generation (RAG) has emerged as a promising solution to mitigate this issue, particularly within the defense sector, where closed-network operations and the need for domain-specific language models are critical. Despite this potential, research on the application of RAG-LLM systems in defense environments remains limited.
This study proposes a RAG-LLM–based AI system designed to support defense technology research and development (R&D) project planning. Using Request for Proposals (RFP) data from a component localization development support project, we implemented a LangChain-based RAG architecture under multiple experimental conditions and integrated it with an LLM to design a task-oriented AI planning system through prompt engineering techniques. Evaluation of the system demonstrated that, when responding to project-planning inquiries, it produced accurate answers with high promptness, matching the original document information. These findings suggest that the system can effectively support planning processes in defense technology R&D. However, due to the limited dataset and the use of only a single document for validation, future research should expand the system to diverse and extensive planning-related documents to enhance reliability and practical applicability.

목차 (Table of Contents)

• 목차
• 그림 목차
• 표 목차
• 약어
• Ⅰ. 서론
• Ⅱ. 관련 연구
• 2.1 생성형 AI
• 2.2 LLM
• 2.3 RAG
• 2.4 프롬프트 엔지니어링
• 2.5 RAG-LLM 활용 및 선행연구 동향
• 2.6 국방연구개발
• 2.7 국방기술 연구개발
• 2.8 국방기술 연구개발 단계에서의 문제점 및 연구 필요성
• Ⅲ. 본론
• 3.1 시스템 설계
• 3.2 데이터 선정
• 3.3 데이터 로드 및 분할
• 3.4 임베딩(Embeddings) 및 벡터 DB화
• 3.5 검색(Retriever)
• 3.6 프롬프트 엔지니어링
• 3.7 LLM
• 3.8 성능평가 방안
• Ⅳ. 성능평가 결과
• Ⅴ. 결론
• 5.1 연구의 결론 및 시사점
• 5.2 연구의 한계 및 향후 연구 방향
• 참고문헌
• Abstract
• 감사의 말