Reference · Cheatsheet

Prompt vs RAG vs fine-tune

Pick the smallest technique that solves the job. Upgrade only when the simpler option fails.

Decision table

| You need… | Start with | Why |

|-----------|------------|-----|

| Better wording / format | **Prompt engineering** | Free, instant, reversible |

| Answers from your docs | **RAG** | Grounds replies in your data |

| Consistent brand tone/style | **Fine-tune / LoRA** | Bakes style into the model |

| New facts that change weekly | **RAG** (not fine-tune) | Update docs, not weights |

| Tool use / actions | **Tools + agent loop** | Model calls APIs safely |

Prompt  →  RAG  →  Fine-tune  →  Train from scratch
 cheaper ────────────────────────────────► more expensive

Fine-tuning to "teach" 50 PDFs of policy → use RAG

RAG for "write like our brand" with no docs → try prompt, then fine-tune

Giant prompt with 200 pages pasted → chunk + RAG

**Related lessons:** `fine-tuning-vs-rag`, `what-is-rag`, `add-context`