I’ve been playing around with ideas on how to use openai for my project, to fill in the gaps for the more generic chatbot responses my existing bot is not coded to handle. My bot is built off RASA and a completely custom response engine. I have hundreds of lines of response dialogue already written that all address specific topics. I thought openai could fill in the gaps for more general Q&A, as well as chitchat.
I was using some examples and messing around in the playground, pasting in blocks of text from some documentation, and asking it questions. Works great for pulling out the answers, but the answers are always paraphrased from the text, without following any of the other instructions in the prompt. Things like “Answer the question to the best of your ability, phrasing the answer as if you are extremely irritated” - that’s not exact, but you get the point. It always seems to ignore the way I’m asking it to present the answer.
Is this a matter of wording the prompt poorly? Or do I need to do fine tuning to teach it how to respond?
Currently, our internal system, assuming the AI is annoyed with the player at the time, when asked something like “So where can I find the blahblah?” might answer with something like. “Oh? Now you want my help? Fine. The damn blahblah is over at the whereplace. Now go away.”
Its not always that intense, but it wouldn’t make sense to get that answer, then the player asks “What’s the rule that says how many blahblah I can get per day?” having it spit out bland text from the documentation wouldn’t fit right.
Doing it using the chatbot examples I can get it to be a little bitchy, but answering questions… nothing. Any tips?