What about the LLM calls though that are done midchain? In the Home Assistant video the multi-intent prompt gets split using what looks like a traditional llm model, which I'm assuming is vulnerable to classical hallucinations.