Three options, three different problems
RAG, fine-tuning and agents are not interchangeable; each solves a different problem. Picking the wrong one is the most common way to burn an AI budget.
- RAG: "the model lacks the knowledge" → retrieve from your documents.
- Fine-tuning: "the model has the knowledge but its behaviour does not fit" → tune style, format and tone.
- Agents: "the answer requires performing several real actions" → tools and a decision loop.
A practical comparison
- Time to results: days for RAG, weeks for fine-tuning, days-to-months for agents depending on scope.
- Cost: RAG adds inference and index maintenance; fine-tuning is a one-off training cost plus version upkeep; agents multiply token spend through iteration loops.
- Risk: low for RAG (answers trace back to sources), medium for fine-tuning (harder to evaluate), high for agents (real actions carry real risk).
- Maintenance: RAG stays current as documents change; fine-tuning needs retraining when requirements shift; agents need continuous monitoring.
The decision tree
1. If documents can answer the question → RAG.
2. If format, tone or terminology is still wrong after RAG → add a light adapter.
3. If the job is about *doing* rather than *telling* → agent.
4. Serious projects usually combine all three: agent + RAG + a fine-tuned model for one specific stage.
The golden rule
Choose the simplest option that passes your success metric, and decide with real evaluation — not vibes. You can always add complexity later; removing it rarely works.
