Persona-driven automated extraction of non-functional requirements using LLM agents

The extraction of non-functional requirements (NFRs) is a crucial yet complex task in information system development, often requiring extensive domain expertise. Traditional persona-driven approaches rely on expert-driven manual processes, which can be inconsistent and inefficient. To address this challenge, we propose an LLM-based autonomous agent framework for persona-driven NFR extraction, integrating Retrieval-Augmented Generation (RAG) to enhance contextual adaptation. The framework systematically automates persona generation, scenario-based experience simulation, interview question formulation, response evaluation, and final NFR specification, incorporating iterative refinement mechanisms. To evaluate its effectiveness, we conducted a case study using a real-world procurement specification, comparing RAG-enabled and non-RAG conditions. Experimental results demonstrate that RAG improves the contextual relevance of interview questions, with an average Context Relevancy Score of 0.5135 compared to 0.4795 without RAG. Additionally, RAG-enabled interviews exhibited broader information coverage in the initial round, as indicated by Convex Hull Volume analysis. However, dynamic flow control mechanisms ensured that non-RAG conditions, through iterative refinement, achieved comparable final NFR completeness. Human evaluation confirmed that the proposed framework generates consistent, structured, and comprehensive NFR specifications while reducing dependency on manual expertise. These findings highlight the potential of integrating LLM agents and RAG to enhance automation and coverage in NFR extraction. Future work will focus on refining adaptive RAG application mechanisms and optimizing persona-driven question generation strategies to further improve the efficiency and accuracy of NFR specification processes.