|
Title |
Development of a RAG-Based LLM Decision Support System for Facility Safety Grade Estimation
|
|
Authors |
장주영(Juyoung Jang) ; 조민건(Mingeon Cho) ; 강건모(Gunmo Gang) ; 차기춘(Gichun Cha) ; 박승희(Seunghee Park) |
|
DOI |
https://doi.org/10.11112/jksmi.2026.30.3.1 |
|
Keywords |
생성형 AI; 대형 언어 모델(LLM); 검색 증강 생성(RAG); 의사결정 지원 시스템(DSS); 시설물 안전등급 Generative AI; Large Language Model (LLM); Retrieval-Augmented Generation (RAG); Decision Support System (DSS); Facility safety grade estimation |
|
Abstract |
Conventional safety grade estimation of infrastructure facilities is conducted based on inspection and testing results, where structural conditions are assessed according to predefined evaluation criteria. However, such assessment processes largely depend on inspectors’ experience and subjective judgment, often leading to inconsistent and less reliable outcomes. In addition, maintenance-related information is commonly stored in unstructured document formats, increasing the burden of retrieving and interpreting relevant criteria during field inspections. As a result, inspectors must manually interpret scattered information across multiple documents, which can further reduce the efficiency and objectivity of the evaluation process. To address these challenges, this study proposes a decision support system for facility safety grade estimation based on a Retrieval-Augmented Generation (RAG) architecture combined with a Large Language Model (LLM). The proposed system converts unstructured maintenance documents into a vector-based knowledge index at the paragraph level and enables evidence-grounded response generation through semantic retrieval and contextual augmentation. By retrieving relevant document segments and generating responses grounded in referenced content, the system mitigates hallucination issues in conventional LLM-based systems. Moreover, it presents explicit document evidence with responses to enhance reliability and explainability in the decision-making process. Although this study focuses on functional validation rather than quantitative performance evaluation, the results demonstrate the practical applicability of the RAG-based approach as a decision support tool for infrastructure maintenance. The proposed system is expected to improve the consistency and reliability of safety assessments and to support the digital transformation of infrastructure management practices.
|