Automating Generation of UML Sequence Diagrams Using LLM

Manually creating UML sequence diagrams from Software Requirements Specification (SRS) papers takes a lot of effort and is prone to errors. Traditional rule-based NLP methods struggle with ambiguity and require extensive human supervision. Large language models (LLMs) and graph-based semantic retrieval using Neo4j are combined in the hybrid Graph Retrieval-Augmented Generation (GraphRAG) architecture shown in this study. Our approach improves accuracy by embedding extracted UML components in a graph and utilizing retrieval-enhanced prompting for guided diagram production. Our technique was tested using four different generating strategies: (i) no prompting or RAG, (ii) only prompting, (iii) retrieval based on FAISS, and (iv) the proposed Neo4j-based retrieval. Across five domain-specific SRS documents, the Neo4j-enhanced solution achieves the highest levels of semantic alignment, structural completeness, and overall accuracy, outperforming all baselines. We also introduce a multi-metric evaluation methodology created especially for the quality of sequence diagrams. These results show that structured retrieval is a viable method for developing AI-assisted software modeling.