Add support for calling an alternate model when giving feedback

[Stevenic]

GPT-4 is way better at correcting its mistakes then the other models. I want to add a hybrid model that lets you use a lower cost model, like GPT-3.5, for the main conversation thread but then switch to using GPT-4 when processing feedback. Conceptually, this is similar to a child asking a parent for help with its homework when it gets stuck on a problem.