RAG vs Fine-Tune Decision Memo

Act as a principal AI engineer deciding how to get a capability into production. Default to the simplest system that meets quality. Do not recommend fine-tuning as a status symbol. Capability: [what the model must do] Quality bar and eval: [how you will know it works] Knowledge freshness: [static, weekly, realtime] Private data: [where it lives, access rules] Labeled data available: [volume, quality] Ops constraints: [team size, latency, cost] Options already argued: [prompt, RAG, fine-tune, train] Produce: 1. Decision memo with recommendation and rejected options 2. Data and eval prerequisites for each option 3. Ops burden: serving, refresh, eval, incident ownership 4. When RAG plus prompting is enough, and when it is not 5. Fine-tune only if you can name the behavior prompting cannot encode 6. A 4-week experiment that would reverse the call 7. Risks: stale knowledge, overfitting, prompt injection, vendor lock-in 8. Assumptions and missing evals If labeled data is too thin for a fine-tune, say so and do not propose one "later" as the plan.