Functionary models are trained to output functions. We should just provide a flag to turn off calculating BNF grammars as we are already doing and just leave to the LLM to answer appropriately.