Prompting vs RAG vs Fine-tuning

Start cheap, escalate only when evals show you must.

	Prompting	RAG / tools	Fine-tuning
Adds new knowledge	Only what fits in the prompt	Yes, fresh and private data	Poorly; not for facts that change
Changes behavior/style	Yes, mostly	No	Yes, strongly
Effort	Minutes	Days (pipeline + index)	Weeks (data + training + evals)
Update cost	Edit text	Re-index docs	Re-train
Citations / traceability	No	Yes	No
Typical wins	Most tasks	Q&A over docs, support, search	Narrow high-volume tasks, distilling a big model into a small one

Rules of thumb

Prompting first - most "we need fine-tuning" cases are fixed with clearer instructions and examples.
Knowledge problem → RAG, behavior problem → fine-tuning.
Combine them: a fine-tuned model can still use RAG.
Fine-tuning needs data: hundreds to thousands of high-quality examples plus an eval set.
Check availability: not every vendor/model offers fine-tuning; open-weight models can be fine-tuned yourself (e.g. LoRA/QLoRA).
New base models often beat your old fine-tune - re-evaluate on upgrades.