The experiment was based on 15 research papers, 18 evaluation questions, and a retrieval setting of k=8; the generator used gpt-oss-20b on Groq, the judge was Gemini 3.5 Flash Lite, and the four metrics were computed via RAGAS 0.2.15.