To enhance the precision of database organization for ventilation systems in large-scale underground engineering tunnel groups and to improve the accuracy and reliability of outputs from large language models, this paper presents a knowledge service system based on large language models. Firstly, a method for constructing a knowledge graph for underground ventilation is introduced, which integrates general corpus training with specific ventilation knowledge. By segmenting structural spatial elements, analyzing fluid dynamics processes, and defining operational condition assurance requirements, the ontology layer of the knowledge graph is modeled. Secondly, a retrieval-augmented generation technology based on the ventilation knowledge graph is proposed, which includes the establishment of a vertical domain document vector database and enhances the capability of large language models in vertical domain question-answering through word vector matching. Thirdly, a knowledge service platform for underground ventilation based on a large language model has been developed. The platform encompasses functionalities such as knowledge graph visualization, user natural language interaction, and the automatic generation of report documents for underground engineering ventilation. The research not only enhances the capability of specialized ventilation knowledge question-answering but also improves the efficiency of retrieval-augmented generation of report documents.