1.3.2. Grounding vs. Fine-Tuning vs. Prompting
💡 First Principle: Three levers change model behavior, and they do different jobs: prompting shapes how it responds (tone, format, reasoning steps), grounding supplies what facts it reasons over (fresh/private data at request time), and fine-tuning adjusts the model's learned tendencies (style, format consistency, narrow task behavior). Choosing the wrong lever is a classic exam trap.
The decisive question is "what's actually missing?" If the model lacks facts, ground it — fine-tuning won't add knowledge and prompting can't supply data the prompt doesn't contain. If the model lacks a behavior or format you keep having to re-specify, fine-tune it so you stop paying for long instructions every call. If you just need to steer a capable model for this one request, prompt it.
| Lever | Changes | Adds new facts? | Best for | Cost/effort |
|---|---|---|---|---|
| Prompting | Response shape, this request | No | Quick steering, few-shot examples | Lowest |
| Grounding (RAG) | Facts available at request time | Yes (retrieved) | Private/fresh data, citations | Medium (index + retrieval) |
| Fine-tuning | Model's default behavior | No | Consistent style/format, narrow tasks | Highest (training + hosting) |
⚠️ Exam Trap: "Fine-tune the model on our documents so it knows our data" is almost always the wrong answer. Fine-tuning teaches patterns, not facts — it won't reliably recall a specific policy number, and it can't include data created after training. The fact-recall job belongs to grounding.
Reflection Question: A bank wants its assistant to (a) always answer in a fixed compliance-approved format and (b) cite the customer's current account terms. Which lever solves (a), which solves (b), and why aren't they the same lever?
