A new standard for evaluating retrieval-augmented generation (RAG) systems using large language models (LLMs) has been proposed, emphasizing explicit measurement criteria and cluster-aware inference. This approach reveals that previous benchmarks may overstate progress, suggesting a need for the industry to adopt more rigorous evaluation methods.