Prompt2Graph: Leveraging LLMs to Construct Knowledge Graphs from Technical Manuals

A structured knowledge graph (KG) can significantly support information retrieval and reasoning in technical contexts when extracting information from unstructured instructional documents. This paper presents a low-overhead process to map tool-part relational triples from textual instructions using a large language model (LLM) and a domain ontology. By using LoRA to fine-tune an LLM to identify domain-specific entities (Tools and Parts) and relations (for instance, "uses," "removes," and "attached") from maintenance and assembly manuals, the approach achieves respectable precision and recall against a test set with expert annotations. A thorough error analysis documents common extraction failures (missing triples, incorrect relation types, etc.) and discusses how ontology changes may resolve such issues. Extraction performance per relation and entity is then reported to highlight which relation and entity types are better represented by the model. An expert review of illustrative extracted triples indicates that the resulting KG will likely be accurately and usefully employed in practice. Finally, lessons learned and future improvements (rule-based post-processing, better alignment with procedural local steps, etc.) are explored to improve KG construction in this domain.