What if you could ensure that an LLM only returned specific syntactic structure (e.g. JSON or XML), specific semantic structure (e.g. a date or a number), or even complete templates (e.g. a sentence with a blank to fill in)? What if the model never deviated from that structure?
This is great, can't wait to start playing with it