Inspired by this article
When I read it, I was thinking 'ok sure, but this sounds like an inhibitory problem' so I set out to give ChatGPT an inner voice to remind it when it is misbehaving.
This particular implementation will only provide an opportunity for the main model to review its message, based on the feedback from the inhibitor model; this can be of course re-engineered for different outcomes.
work in progress!
env variables (e.g. in a .env file if using vscode)
- "KEY", for your OpenAI API key
- "PROMPT_SYSTEM" this is where you set your chatbot's default personality (e.g. helpful expert in area x); there is also where you might want to spell out the role of the inhibitor model, so the main model knows what to do with it
- "PROMPT_INHIBITOR" default instructions to the inhibitor model (i.e. what to look for, formatting the feedback in json etc)
my code is expecting the responses from the inhibitor in the following format:
{"decision": "pass", "explanation": ""}
{"decision": "inhibit", "explanation": ""}
so, you need to engineer your prompt accordingly.
Also, you need to educate the inhibitor about your risk appetite, otherwise it might complain about every minor thing.
- the UI will only work in Jupyter environments that handle ipywidgets properly (e.g. Databricks' Jupyter does not, vscode is marginal)
- if using vscode keep in mind:
- enter and shift-enter in the input window are captured by vscode, use the Send button instead
- in vscode the default key binding to turn a cell into markdown is "m", you need to re-bind that in the settings as vscode captures it during typing