CoachNed
Sign up free

Method

RAG looks things up. Fine-tuning changes the intern’s habits. Do not mix them up.

This is the AI question generalist interviewers can actually grade. Get it right in one minute and you sound like you have done the work.

Ned comparing RAG and fine-tuning on a whiteboard
Read the 12-min lessonSit the AI case

The two moves

RAG (retrieval-augmented generation) means: search the client’s documents, paste the relevant chunks into the prompt, then let the model write. The model stays generic. The facts come from the filing cabinet. When the policy updates, you update the cabinet, not the model.

Fine-tuning means: extra training on your examples so the model more often produces your format, your tone, or a narrow skill. It does not magically know this morning’s endorsement. It is closer to tutoring than to a searchable archive.

  • Need citations and changing facts? RAG.
  • Need a stable format on a stable task, and you have graded examples? Fine-tune can help.
  • Need a first draft of an email? Prompting is enough.
  • Need the system to pay a claim? Neither RAG nor fine-tune is the whole answer — you need a harness.

What interviewers want to hear

“I would RAG over the policy library because the facts move. I would not fine-tune until I have a labelled set and a task that does not depend on last week’s wording. I would still keep a human on payout until evals are green.” That is a partner-safe paragraph.

If they push on cost: retrieval has an infrastructure cost; fine-tuning has a data and retraining cost. Both beat “we’ll just use a bigger model” when the miss is a wrong clause.

The AI lesson covers this in context with agents. Clearclaim Mutual makes you choose it under a clock.

FAQ

Can I do both RAG and fine-tuning?

Yes, later. Interviews reward sequencing: prompt, then RAG, then tune if the task is stable. Both at once is how projects get expensive before they get good.

Does a bigger context window replace RAG?

It reduces the need to be clever about search. It does not replace having the right document, access control, or a way to cite.

Is fine-tuning the same as training from scratch?

No. From-scratch training is rare in these cases. Fine-tuning starts from a model that already speaks. Still expensive compared to RAG if your problem is “find the clause.”

Keep going