I’m using the RealtimeAPI. I’ve built a system where, if my system detects that the RealtimeAPI has made a mistake, I delete conversation items using “conversation.item.delete” and trigger a new response from the API using “response.create”. The instructions I provide to “response.create” are very simple - usually, “Ask the user to clarify X” or “Use function Y instead”
However, often times, this doesn’t trigger a response that follows these instructions. Usually, it just tries to repeat the same output that it had created (which we deleted, because it was wrong).
Have I missed something? Has anyone experienced the same issue? What’s going on here?