Investigation of AI tools Performance in the Definition of Microservices Software Architectures

Abstract

The design of software architecture is a pivotal step in the software development lifecycle, bridging user requirements and system implementation through the definition of high-level structural designs. Despite its critical importance, architectural design remains a challenging, time-intensive, and error-prone process. This thesis investigates the performance of artificial intelligence (AI) tools, specifically large language models (LLMs), in automating the generation of software architectures focusing on Microservices-based systems. Building on prior research, this study explores how different input formats, ranging from plain-text requirements to detailed specification documents, model selection and Retrieval-Augmented Generation (RAG) techniques, affect the quality and compliance of AI-generated architectural designs with the software requirements. We apply an evaluation framework based on assessments from domain experts and we introduce a set of objective metrics to pave the road towards an automatic evaluation process. Additionally, this study explores whether smaller, locally hosted LLMs can serve as practical alternatives to commercially available AI tools. The results provide insights into the potential for AI to transform the architectural design phase of software development, enhancing design efficiency and quality while reshaping the role of software architects in collaborative human-AI workflows. This work contributes to the growing field of AI-assisted software engineering and outlines future research avenues to further integrate intelligent automation into complex system design.