How does fine-tuning differ from RAG internally?
Change the brain, or hand it a book.
Picture it first
Fine-tuning is like sending a student to coaching classes for months — they actually learn and change how they think and write. RAG is like allowing the student to bring the textbook into the exam — they didn't learn anything new, but they can look things up. Coaching is great for skills; the textbook is great for facts that keep changing.
Now the same idea, in simple QA words
There are two ways to make an AI fit your company. Fine-tuning = more training — it actually changes the AI, like coaching. Great for teaching a style, tone or fixed format. RAG = giving the AI the right documents at question time — the AI doesn't change, it just looks things up. Great for facts that change and need sources. Many companies use both: fine-tune for style, RAG for facts.
How the story connects to the real thing
Real-world example
A telecom company wants its bot to (a) speak in its brand tone with a fixed answer format and (b) know today's plan prices. They fine-tune a smaller model on 3,000 approved support conversations for tone and format, and use RAG over the live price catalogue for facts. When prices change on Monday, they just re-index — no retraining.
Step by step, in plain words
- Fine-tuning: collect examples → train the AI → test the new version → use it.
- Updating fine-tuned knowledge means training again.
- RAG: store documents → find the right ones for each question → give them to the AI.
- Updating RAG knowledge means just updating documents.
- Combine: fine-tuned AI + RAG documents.
How a tester tests it
- Fine-tuned model: compare against the base model on BOTH the target task and general capabilities (catastrophic forgetting).
- Check for memorised training data leaking (PII in fine-tuning sets).
- RAG: test freshness, citations, and permission filtering.
- For combined systems: isolate which part caused a failure — weights or retrieved context.
Try it yourself
1) List 5 requirements of a chatbot you know and tag each 'fine-tune', 'RAG' or 'prompt only'. 2) Explain one choice in one sentence.
Interview angle
“A client wants to fine-tune a model on their HR policies so it 'knows' them. What do you say?”
Fine-tuning is a poor fit for facts that change and must be cited. Policies change, so they'd need retraining every update, and the model can't point to the clause. I'd recommend RAG for the policy knowledge and consider fine-tuning only if they need a consistent tone or format that prompting can't achieve.