Abstract
Architectural smells are abundant in codebases and regularly hinder the development of stable and maintainable code. Understanding and removing these elements can consume a huge amount of developers' time, who often need to prioritize implementing new features. This causes a substantial increase in Technical Debt, compromising the scalability and maintainability of the codebases, at time bringing the development to a standstill. Meanwhile, the use of Large Language Models for small error correction is constantly growing, bringing the attention of an ever-wider audience to these technologies. This study explores a first approach to use Large Language Models to suggest refactoring for architectural smells, with a focus on Cyclic Dependencies smells. We study the use of detailed prompt and Retrieval-Augmented Generation (RAG) to enhance LLMs, and we study local vs cloud LLMs. The results are promising, also validated with a series of interviews with students and developers, and highlight how additional and precise context is key to enhance the use of LLMs to propose refactoring suggestions. A multi-agent approach seems to be more suited when increasing the complexity of the smells.