In response to the challenges of inefficient data management and difficult process reuse caused by the dispersion of multimodal data and the lack of inter-document connections in the field of steam turbine blade machining processes, this paper proposes a multimodal cross-document graph modeling technique based on Large Language Models (LLM) and Technical Language Processing (TLP). Initially, a multimodal graph ontology framework is designed, followed by the application of TLP technology to address issues such as complex terminology, non-standardized text, and cross-document semantic inconsistencies in process documentation. To tackle the problems faced by LLMs in understanding domain-specific process knowledge, including term ambiguity, lack of contextual information, and difficulties in cross-document semantic association, a template-based hierarchical prompt engineering method is introduced to facilitate the automatic extraction and integration of process information. The results show that TLP solves the problem of poor adaptability in the field of LLM and significantly improves the accuracy of term recognition. The context learning augmentation rules of LLM solve the problem of weak TLP generalization ability and improve the ability to deal with new terms, and the two are fused to form two-way enhancement. At the same time, the constructed map can intuitively display the multimodal information and its correlation of process documents, which significantly improves the efficiency of knowledge integration and reuse.