Make custom LLM less weird. Or at least give example of what response should be exactly. I'm using custom LLM, it means I only want to send response text so VAPI tts'es it for human talking with my LLM, not some weird OpenAIthingamajig object nonsense.