3 ms·

Cool idea. Any early data on how much this reduces token usage / cost in practice on real workflows?