Key Concepts
- 1Prompting (with good context, RAG, and few-shot examples) solves most commerce use cases without the cost and rigidity of fine-tuning — it's the right default
- 2Fine-tuning bakes behavior into the model weights; it shines for consistent style, domain-specific formats, or narrow high-volume tasks where prompting is inefficient
- 3The tradeoffs: fine-tuning costs data, training, and maintenance, and creates a version you must re-tune as base models improve — a real ongoing burden
- 4Prompting is faster to iterate: you can change behavior instantly, A/B test, and roll back — impossible with a fine-tuned model that requires retraining
- 5RAG vs fine-tuning is a common confusion: use RAG for knowledge (facts that change, like catalog), fine-tuning for behavior/format (how to respond) — they solve different problems
- 6Fine-tuning can reduce per-call cost at high volume by letting a smaller tuned model match a larger prompted one — a cost play once volume justifies the investment
- 7The decision is sequenced: prove the use case with prompting first, then fine-tune only if consistency, cost, or format needs justify the added complexity
Why It Matters for PMs
Prompting-first is usually right: Most commerce needs are met by prompting and RAG without fine-tuning's cost and rigidity. PMs should default here and justify fine-tuning explicitly.
RAG vs fine-tuning clarity: Confusing knowledge (RAG) with behavior (fine-tuning) leads to wrong architecture. Knowing which solves which is a foundational PM distinction.
Maintenance burden: Fine-tuned models must be re-tuned as base models improve. PMs must weigh this ongoing cost against the consistency or unit-cost gains.
Test Your Understanding
Question 1 (Multiple Choice)
What should usually be the default approach for commerce use cases?
Question 2 (Multiple Choice)
What is the key distinction between RAG and fine-tuning?
Question 3 (Multiple Choice)
What is a real ongoing burden of fine-tuning?
Question 4 (Short Answer)
How would you decide whether a commerce use case justifies fine-tuning over prompting?
Think about your answer before revealing the explanation.
Question 5 (Short Answer)
Why is RAG the right tool for catalog knowledge while fine-tuning suits response behavior?
Think about your answer before revealing the explanation.