Hey,
I’m trying to build an MVP for my platform and I’m facing couple of issues I can’t find a way to overcome.
Idea:
- The chatbot has instructions on how to respond, what tone and voice to use, how to engage with person to provide required support.
- It has attached couple of files with information I want it to use to have better understanding on how to respond to user, how to evaluate the user and it’s behavior to respond properly
Issues:
- After couple of messages it feels like it forgest it’s instructions and loses the track and either starts to respond some unrelated garbage. However I have a point in instructions which states something like “read instructions before each response”
- Sometimes it starts to just paste the content from the files without any logic
What I’m thinking of as a solution:
- Create a two-level system of models where first level acts as a middleman between user and AI
- Where first AI just gets the prompts from user, passes them to main AI (with knowledge and instructions)
- Main AI responds, first AI checks the response and if complies with instructions forwards it to user, if not asks the main AI to reread the instructions and regenerate response
What are your thoughts on that? Did anyone had similar issues? How did you solve them?
I’d like to keep my costs to just ChatGPT Plus, but it feels I won’t get away just with that and for multi-level chat I’ll need to use API anyway.