Applied AI / Forward Deployed Engineer · Fine-Tuning & Model Adaptation · Lesson 1 of 2
The adaptation ladder: prompt → RAG → fine-tune
There's an ordered ladder of ways to make a model do what you want. Climb it in order — each rung is more expensive and less flexible than the last.
- Prompt engineering — cheapest, instant to change. Try first.
- Few-shot / dynamic examples — teach by demonstration in-context.
- RAG / tools — add knowledge and abilities at runtime.
- Fine-tuning — change the weights. Last resort, for when the above plateau.
Fine-tuning does NOT add knowledge (mostly)The classic mistake: 'I'll fine-tune to teach the model our docs.' Fine-tuning is for behavior, format, tone, and narrow skills — not for injecting a knowledge base (that's RAG). Fine-tuning on facts is expensive, forgetful, and un-citable.
- Fine-tune when: you need a consistent style/format prompting won't hold, a specialized narrow task done cheaply/fast, lower latency/cost via a smaller specialized model, or to bake in a hard-to-express behavior.
- Don't fine-tune when: the knowledge changes often (RAG), you have few examples, or you haven't exhausted prompting.
◆ Lock it in
- Climb the ladder in order: prompt → few-shot → RAG/tools → fine-tune.
- Fine-tuning changes behavior/format/skill, not knowledge — use RAG for facts.
- Fine-tune only after prompting/RAG plateau, when you have enough quality examples.
Feynman drill — say it out loudA teammate wants to fine-tune to make the bot know this quarter's prices. Talk them out of it.
Step 1 rate your confidence · Step 2 pick your answer
Which is the BEST reason to fine-tune (rather than prompt/RAG)?
Step 1 — how sure are you?