Five-Step Semantic Analysis Middleware for Atlassian Jira Software Requirements Transformation Using Generative Large Language Model

The article presents a framework for transforming requirement artifacts in Agile development environments into a structured knowledge base using large language models (LLMs) and graph-based methods. The study focuses on addressing key limitations of contemporary requirements engineering, including tool fragmentation, weak traceability, and insufficient adaptability to semi-structured data. The proposed system operates as an intermediary layer between Atlassian Jira and a graph-based knowledge repository, implementing a five-step methodology: requirement clustering, test linkage, dependency identification, deduplication, and historical reconstruction. The core component is the LLM, which the system interacts with via a JSON-oriented protocol that includes instructions for data interpretation, the expected structure of the response, and permitted actions (e.g., creation, updating, traceability). The modular system architecture includes a user interaction layer, an orchestration agent, a semantic core, a graph database (Neo4j), and an access control subsystem. Both the theoretical and experimental phases of the study were carried out with the support of SoftServe Inc. As part of the experimental deployment, the system processed over 18,000 Jira issues. Pilot testing demonstrated high response accuracy (90.4%) and the potential for significant time savings for analysts and testers. The system also proved capable of reconstructing historical requirement states, detecting duplicates, and identifying logical inconsistencies in dependencies – features particularly valuable for complex products with multi-layered structures. However, the system remains at the prototype stage and has a number of limitations, including dependency on the quality of prompt formulation, challenges in interpreting LLM decisions, and the need for carefully configured secure data access. The results should be considered preliminary; further testing across different projects, domains, and larger query volumes is needed to assess the scalability, robustness, and practical applicability of the proposed approach.