I feel like I’m missing something here and I’d really like for someone to help clear things up for me. As far as I can tell, GPT is really good for any chat based applications, but if you want to perform any sort of logic beyond returning whatever GPT replies with, it seems limited at best.<p>For example, let’s say I want to conduct a survey using GPT to “chat” with a user. But let’s also assume I want to collect answers to that survey using a simple scoring system (1,2,3,4,…) extrapolated from the answer. The conversation using GPT would be something like this:<p>User: Hello<p>GPT: Hello, would you like to start a survey? Im blah blah…<p>User: Yeah sure, whatever I guess.<p>GPT: On a scale of 1-5, how cool is AI?<p>User: I guess 4<p>So now I’d need to save the answer that the user provided, but “I guess 4” is not what I should write to my database. I need to write “4,” but accurately parsing that has two issues: 1. I don’t know at what point this specific question was answered, and 2. There could be multiple integers in the answer, but due to the conversational nature, picking the right one can’t be done accurately.<p>So I guess what I’m curious about is: how do I use GPT in a scenario where my code is reacting to replies rather than a human simply chatting with it?<p>I’ve used some system prompts to ask for specific reply structures, but even those are flakey at best and inconsistent. Aside from pushing replies back through GPT to extract intent, I’m not sure where to go at this point. I’m guessing as I tinker it’ll open up more possibilities, but I’ve always found asking those with more experience to help me “skip the line” in terms of head banging.