TransArC-AI

TransArC-AI extends the TransArC idea by using an LLM to generate a simple architecture mode (SAM). In this approach, instead of requiring a hand-made SAM, a large language model (such as GPT-4) is prompted to extract or invent the main component names from the SAD (and optionally from code). These names serve as a minimal architecture model (i.e. a list of components). Then, as in TransArC, these LLM-derived components are matched to code. The goal is to bridge the SAD–code gap without manual modeling.

• How it works: Given the software architecture text and the codebase, the system asks the LLM to list likely component names. That list of names forms a “Simple Software Architecture Model” (SSAM). Finally, code elements with matching names or descriptions are linked to the documentation. This pipeline avoids needing an explicit UML model.
• Effectiveness: TransArC-AI achieved very competitive results. Using GPT-4o, it obtained a weighted F1 of about 0.86, nearly as good as the original TransArC with a hand-made model (F1 0.87). It also substantially outperformed the ArDoCode baseline (which scored ~0.62). This shows that LLMs can automatically infer the key architectural components.