Abstract
The goal of construction site risk and hazard identification is to enhancesafety management through automation. Existing research based on large languagemodels falls into two categories: image-text matching for collaborativereasoning, which struggles with complex hazard features, and instructionfine-tuning or dialogue guidance using professional datasets, which suffersfrom high training costs and poor generalization.To address this, we propose ahazard identification method using similar case retrieval enhancement. Byintegrating external knowledge and retrieved case contexts via promptfine-tuning, we mitigate misjudgments caused by limited domain knowledge andweak feature associations. Our method includes three modules: retrievallibrary, image similarity retrieval, and large model retrieval enhancement,enabling efficient recognition without training. Experiments on realconstruction data show significant improvements. For instance, GLM-4V'srecognition accuracy increased to 50\%, a 35.49\% boost. The method enhancesaccuracy, context understanding, and stability, offering new theoretical andtechnical support for hazard detection.