“The evaluation of large language model-based systems has expanded beyond traditional ranking accuracy to include multiple new dimensions such as hallucination and latency.”