LLMs are surprisingly good at extracting data from other data, so at the end of the day there is no right or wrong, it's what works best for you use case.