Measuring and improving LLM accuracy for structured outputs