An Automated Risk Assessment Based on GPT and RAG for Small and Medium-Sized Construction Sites

The construction industry faces frequent safety accidents due to its complex production structure and variable site conditions, with small and medium-sized construction sites being particularly vulnerable. To address these challenges, risk assessment plays a critical role in identifying hazardous factors in advance and establishing countermeasures. However, current risk assessments rely heavily on the subjective judgment and experience of safety managers, which poses limitations, especially in sites lacking specialized personnel. This study aims to propose a more efficient risk assessment method by overcoming the limitations of large-scale generative language models. By applying Retrieval Augmented Generation (RAG; retrieval augmented generation), which provides relevant documents to the model to generate more accurate answers, we enhanced the performance of the GPT model in deriving specific and expert-level safety measures. The experimental results showed that the RAG approach allowed for the effective identification of hazardous factors tailored to the task characteristics and generated more concrete safety countermeasures compared to using GPT alone. Based on these findings, the development of an automated risk assessment system for small and medium-sized construction sites could reduce reliance on professional safety managers and enable more efficient and systematic safety management.