Long-distance and deep-buried tunnel underground engineering presents significant challenges, including complex geological conditions and the need for human-machine collaborative operations. Knowledge graphs offer an efficient means to organize large volumes of information, which is crucial for enhancing project quality and operational efficiency. However, knowledge in the field of underground engineering is often fragmented across extensive textual sources, and traditional “top-down” knowledge graph construction methods often fail to meet the requirements of this domain. To address this, a “bottom-up” approach to knowledge graph construction is proposed. This approach begins with the collection of underground engineering textual materials, followed by data cleaning. Subsequently, a text embedding model is fine-tuned using this dataset. Simultaneously, knowledge triples are extracted from the dataset through prompt engineering, and text clustering is performed to define the schema layer of the knowledge graph. The results demonstrate that the fine-tuned text embedding model significantly enhances the effectiveness of text clustering. As a case study, a knowledge graph for underground engineering was constructed using data from a tunnel project in Lesotho, and a knowledge platform was developed to support construction decision-making.