I would transform the most commonly occurring formats programmatically. The rest could probably be handled by GPT-3. Alternatively, divide the task between as m

I would transform the most commonly occurring formats programmatically. The rest could probably be handled by GPT-3. Alternatively, divide the task between as many cloud VMs hosting something like LLaMa as it takes to fit your time constraints.