1
Curious builder0 XP earned · 300 to level 2
0 daysFinish a lesson to begin
Badge collection0 of 6 unlocked
51 small wins to finish your pathNext question →
How do you decide between prompting, RAG, PEFT (LoRA), and full fine-tuning?
30-second answerSay your answer out loud first, then reveal.
| Approach | Use when | Data needed | Cost | Update speed |
|---|---|---|---|---|
| Prompting / few-shot | Task is clear; strong base model | 0–20 examples | Lowest | Instant |
| RAG | Needs specific / private / fresh knowledge | Documents | Low–medium | Re-index |
| LoRA / PEFT | Consistent format, tone, classification, domain jargon, tool-calling style; distilling into a small model | ~500–50K examples | Medium | Hours |
| Full fine-tuning | Large shift (new language/domain), max quality, owning a model | 100K+ examples | High | Days |
| Continued pretraining | Lots of raw unlabelled domain text (e.g. legal corpus) | Billions of tokens | Highest | Days–weeks |
Signals that fine-tuning is worth it
- The prompt has become huge (thousands of tokens of instructions and examples) and expensive on every call.
- Output format consistency failures persist despite good prompting.
- Latency or cost demands a smaller model that must match a bigger one on a narrow task.
- You have high-quality labelled data and a reliable eval.
Signals it isn't
- The problem is factual knowledge (use RAG).
- The data is small, noisy or changing weekly.
- You can't evaluate the result.
Related
Slow is fine. Stopping is the only problem.