I think an interesting use case could be to extend on the "kor" package. I think it solves quite a common problem - generating clear structured output with natural language. But I think the kor package is missing embedding support, so its hard to work with existing data
