| Kurzfassung
|
Maintaining end-to-end traceability between natural-language requirements and sourcecode is essential for comprehension, change impact analysis, and compliance, yet remainsdifficult in practice. This thesis presents a Visual Studio Code extension that brings documentation‑to‑code traceability into the developer workflow and a retrieval‑augmented approach,TRAG, that combines deterministic vector search with lightweight large‑language‑model(LLM) verification. TRAG pre-processes a workspace by summarizing compilation units,embedding summaries, and storing vectors locally. At link time it splits documentation intosentences, optionally rewrites them for precision, retrieves top‑k candidate code summariesby cosine similarity, and asks an LLM to accept or reject each candidate using compact context;chain‑of‑thought prompting is optional.We evaluate the TRAG against the ArDoCode verification baseline on benchmarked systemswith gold‑standard links, reporting precision, recall, and F1. ArDoCode provides consistentlyhigh recall and the best F1 on JR, TM, and BBB, whereas TRAG improves precision and canexceed F1 on MS and slightly on TS (e.g., best MS 0.208 with mistral‑nemo‑cot; TS 0.316 withgemma3‑4b). Chain‑of‑thought shows mixed effects, helping when evidence is compact butreducing recall otherwise. We discuss design choices, threats to validity, and practical operating points. Overall, retrieval‑augmented verification is a viable complement to recall‑orientedbaselines: it raises precision when evidence is well‑separated.
|
