just set confidence threshhold low, then prompt the llm to clarify picked up phrases that dont logically make sense etc