Comment by mpartel

Some LLM APIs let you give a schema or regex for the answer. I think it works because LLMs give a probability for every possible next token, and you can filter that list by what the schema/regex allows next.