Deductive Software Architecture Recovery via Chain-of-thought Prompting

As software evolves, software architecture recovery techniques can help for effective maintenance. We envision a deductive software architecture recovery approach supported by Large Language Models (LLMs). Unlike existing inductive (bottom-up) recovery techniques, which reconstruct architecture by considering the properties observed at implementation level, our top-down approach starts with architectural properties and seeks their manifestations in the implementation. It employs a known Reference Architecture (RA) and involves two phases: RA definition and code units classification. A proof-of-concept with GPT-4 emulates deductive reasoning via chain-of-thought prompting. It demonstrates the deductive SAR approach, applying it to the Android application K-9 Mail and achieving a 70% accuracy in classifying 54 classes and 184 methods. The future plans focus on evaluating and refining the approach through ground-truth assessments, deeper exploration of reference architectures, and advancing toward automated human-like software architecture explanations. We highlight the potential for LLMs in achieving more comprehensive and explainable software architecture recovery.