<p>Retrieving precise fire safety regulations is essential for the relevant staff in smart city development and building safety management. Requesting information via large language models (LLMs) is popular nowadays. However, the LLMs present limitations in handling fire safety regulations queries due to the complex and strict numerical requirements in fire safety regulations. We introduce a triple-phase precision control strategy as a novel approach to ensure accurate and reliable regulation queries. First, we convert fire safety regulations into a standardized JSON format to ensure data consistency and accessibility, as well as friendly to computer programming language. Second, we develop a deterministic matching algorithm to accurately align user queries with relevant regulation clauses. Finally, LLMs generate summaries according to the matching clauses and a preset prompt, preventing hallucination risks. Experimental results demonstrate that our proposed framework outperforms general approaches, including commercial LLM clients with Retrieval-Augmented Generation capabilities (which use text chunking techniques, either integrating web-based retrieval or allowing local file uploads), as well as pure LLM methods, in reducing errors and improving precision. The proposed scheme not only enhances the accuracy of fire safety regulations retrieval but also provides solutions for other domains with precision-critical requirements, advancing the application of AI in regulatory compliance.</p>