Abstract: As critical special equipment， elevators heavily rely on periodic inspection reports for fault diagnosis and risk assessment. Existing approaches primarily focus on sensor and time-series data while overlooking the rich knowledge embedded in textual inspection reports. To address this limitation， this paper proposes a large language model （LLM）-based method for constructing an elevator inspection and diagnostic knowledge graph. Based on elevator inspection technical specifications， an ontology for elevator inspection and diagnosis is first developed. On this basis， contextual prompt learning is employed to automatically extract entities and relationships from inspection reports. This method addresses challenges such as limited domain-specific semantic representation and reasoning capability， and mitigates the limitations of general-purpose LLMs. Subsequently， entity alignment is performed through predefined rules， and a high-quality building elevator diagnostic knowledge graph is constructed. The proposed method requires neither manual annotation nor model training， and can efficiently process multi-source heterogeneous elevator inspection data. Experimental results demonstrate that LLM-based prompt learning under domain ontology constraints significantly improves knowledge graph coverage and accuracy. Compared with zero-shot learning， few-shot learning improves entity recognition precision， recall， and F₁-score by 6.1 percentage point， 27.20 percentage point， and 17.61 percentage point， respectively， thereby providing effective data support for downstream intelligent fault diagnosis.