When would you fine-tune the generator instead of putti...

The decision to fine-tune versus prompt should be split based on the objective: fine-tune when you need to instill specific behaviors or styles, and use retrieval or prompting when you need to provide the model with external knowledge. Economically, fine-tuning requires a one-time cost of 6ND_SFT, while prompting incurs a per-query cost of 2Nk_td. The break-even point occurs after 3*D_SFT / k_td queries, which typically represents several days of production traffic.