Research on Large Model Text-to-SQL Optimization Method for Intelligent Interaction in the Field of Construction Safety

With the increasing amount of information and data in construction safety management, traditional business query and analysis methods are no longer sufficient to meet the demands for flexibility and complexity in on-site management. This study addresses the information query needs in the construction safety domain by proposing a graph-structured retrieval-enhanced large model Text-to-SQL optimization method. By standardizing database table information, addressing few-shot problems with example-based and enterprise safety domain knowledge prompt templates, and utilizing graph data-based knowledge retrieval enhancement techniques, the method enhances the large model's business understanding capability in the construction safety field and reduces computational resource consumption. Furthermore, by employing reasoning strategies such as chain-of-thought prompting, easy-to-hard progression, self-consistency, and self-correction, the accuracy of the large model in responding to complex construction safety business questions is significantly improved. Experimental results on the open-source model llama-3-sqlcoder-8b and the closed-source model GLM-4-Plus demonstrate that this method achieves substantial improvements in execution accuracy and syntactic accuracy, proving its universality and effectiveness.