GPT functions allow the LLM to infer whether it has captured enough information in a conversation to then create a pre-defined JSON object that can be passed to an external system. We have tested this concept with a helpdesk capture flow that has a chatgpt conversation until it infers it has enough information for a ticket and then creates a proper object and returns it.
This is a great new feature and allows us to implement plug-in type functionality for dynamic chatgpt conversations.
Maybe we can add an advanced option in the action (alongside temperature etc) to determine whether to pass a function array or not. Or may be better to have its own 'completion with function' action that returns the function name and object as separate properties if triggered (or a standard chat reply if not).
https://learn.microsoft.com/en-us/azure/ai-services/openai/how-to/function-calling
I have tested this inference with GPT-4 and it's performing well (better than 3.5) so, when coupled with additional models, we'd have a borad suite of powerful tools.
