What strategies do you use to evaluate and benchmark the accuracy of an LLM-powered query generation system?