Yes, I get back only JSON. I found that providing a sample object is the best way to be consistent. I also have some basic validation on fields and have follow-up calls if needed. For example, dialog must be 1 character and at max 255. If I see empty dialog or too long dialog (or any other invalid fields), I pass back the JSON to GPT as well as the basic list of 'this property is wrong due to X' and make it provide a new JSON object.
GPT4 is much better than GPT3.5 at producing valid JSON and nothing else.
What I’ve found quite useful is to extract text from the first { and last }. This solves the problem of GPT adding some “helpful” explanations outside the JSON.
There are also some more lenient JSON parsers that will handle comments and other non-standard things. They can be a bit hard to add in to some frameworks as it’s kind of the opposite of what you are normally trying to do which is reject bad input.
