The trick is training a little, then augmenting with documents using RAG. The idea that a model alone can handle complex use cases is common, but usually wrong.