It’s a mix of a classification model + LLM for when none of our topics are a high probability fit.