Large language models are challenging to adapt to new enterprise tasks. Prompting is error-prone and achieves limited quality gains, while fine-tuning requires large amounts of...