What's the Vapi's expected response type from custom LLM? Documentation doesn't give any information on this. I tried to return raw completions response from chatgpt, but it didn't work.