PRODUCTION SCENARIO
A claims team needs 30 few-shot examples to make Gemini write summaries in its house layout: a claim-type code first, then a three-line action list. The prompt is expensive, the layout is still wrong in 8% of cases, and the team holds 1,500 approved summaries.
Which approach should the team evaluate next?
Answering here is anonymous. Nothing is saved unless you sign in.
Show answer and explanation
Answer: Supervised fine-tuning on the approved summaries, dropping the examples
Supervised fine-tuning suits a well-defined task with labeled data and is Google's route for aligning a model with custom syntax and rules a prompt cannot hold; tuning also saves cost at inference by removing few-shot examples from every request. A reformatting second call adds another request, a larger model keeps all 30 examples and their cost, and RAG makes the prompt longer rather than teaching the layout.