Building Software Functional Requirements Lists Using RAG with Distinct LLMs in Multiple Interactions

This work explores the automated elicitation of functional software requirements using Large Language Models (LLMs). The proposed approach employs multiple LLMs in conjunction with Retrieval-Augmented Generation (RAG) across iterative refinement rounds. This process enables the generation of candidate requirements from high-level software descriptions, with validation through similarity analysis. The method addresses key challenges in current AI-based elicitation approaches, including lack of iterative refinement, hallucination control, and semantic convergence. Preliminary results demonstrate that multi-round feedback and cross-model integration improve alignment with expert-defined requirements. This study marks the initial phase of a broader doctoral research project titled Agent Family for Software Engineering Teams, which aims to develop a set of intelligent agents to support various stages of the software engineering lifecycle, including requirement elicitation, project management, and software quality, ultimately forming a cohesive and responsible AI ecosystem for software teams.