I’m building a prompt to role-play a sales call. It works really well, however there is one problem I haven’t been able to fix for months. GPT-3.5 seems to love asking more than 1 question in a message. For example:
Hey there! Is this John? I’m Alice from Bank Name, and I’m calling about your application. How are you doing today?
I’ve tried everything, but the only sureway approach was to use stop-sequences, which is not great, because I can’t know if I need to add a question mark to the end of the message or not.
Has anyone managed to overcome this?
So is your prompt something like “you’re a bank services salesman…”, kinda open just telling the role? Did you try with a few-shot approach to try to form the answer in a certain way, like presenting a couple of short interactions with one question each?
Maybe you can also incorporate your prompt? I tried to replicate this, but seems to work ok with this prompt:
“Act as a bank sales person, while being nice working hard to find out about his situation and which services he could be interested with. Once you found a suiting service, try to close the deal. You’re playing the bank sales person, I’ll give you the customer answers. Ask one question at a time and then stop.”
I also tried adding one/few shot with examples, but with the few tries I did it seemed to mess up the format of question-answer. I tried to mitigate the several-questions-problem by providing examples where the question is summarised at first, then the whole conversational style answer is given, to force the answer generation to focus on one question only.
Use a question mark token as a stop sequence.
You’ll know if you need to add a question mark at the end by the absence of a punctuation mark at the end of the response.