I wonder how would they calculate the metrics if the result is generated instead of retrieved? Is it likely that the LLM can generate exactly the same output as the desired result?