I’m currently developing an assistant AI chatbot for a game I am authoring. It currently uses RASA for the intent classification and entity tagging, and most of the work beyond that is my own custom code.
I’m nearing the point of development where the basic framework is in place, and it’s time to start adding content.
Adding in multiple responses that can handle the variety of question types that could be asked is beyond complicated. Basic stuff I have down. Who are the blahblahs? Are they dangerous? Do they have a home base? Stuff like that.
But there are other topics that are more nuanced and classifying all the different questions is becoming daunting. Hence I stumble upon OpenAI. I play with the playground. Beyond impressed at how well it does some sample tasks I gave it.
Sorry, I tend to ramble, probably could have gotten to the actual question sooner, but I feel context is important.
This is for a game that takes place in a post apocalyptic future. The current world is detailed, and there are lots of new terms, slang, technologies, etc that I’d like the AI to be able to offer information on.
Beyond attempting to excessively filter out non-in-game words/topics (I dont want it to provide answers if someone asks about COVID or WWII) is there a way I can restrict it to only the information in the training data I provide, while still utilizing all it’s learned in the GPT3 models? For example, I’d like it to be able to rephrase things, and answer with “emotion”, but not be able to answer questions about facts/events/etc in 2019. I’m guessing not, filtering out what it knows and doesn’t know would be an excessive project.
Another question would be if it’s possible to integrate new terms and slang into it’s responses. This wouldn’t be that big of a deal, but it would be nice if the AI is “upset with” the end user that it could integrate custom language/slang/game terminology into it’s replies.