Language models are really good at free association tasks, such as semantic fuzzy search. Next token prediction is among the worst possible ways to use them (although if there's no other obvious way of getting the information out of the model, it works in a pinch).

Which project did you use?