Treat it as a ladder of cost and commitment. Start on the cheapest rung, and climb only when the one below clearly fails.
Prompting changes nothing lasting, only each call's instructions. Always start here, and stay when the behavior can be put into words.
RAG changes the facts in the context window, for private, fresh or changing knowledge.
Fine-tuning changes the weights, so behavior sticks. Use it when prompts work but are long, slow or unreliable.
Continued pretraining changes base knowledge and language, for a model that barely understands your domain's text or notation.
As a rough rule, RAG changes what the model answers with, and fine-tuning changes how it answers.
These stack rather than compete. Most production systems are prompting plus RAG, with a fine-tune added later for cost or consistency. Pretraining is rare, because it needs billions of tokens.
Rewriting in plainer words…
This answer doesn't lend itself to a diagram - it reads best . No credits were charged.
How would you know prompting has clearly failed and it's time to move up a rung?Build a small eval set first and measure against it. Then name the actual failure, such as a missing fact, drifting format, or a prompt too long and slow.
How do fine-tuning and RAG work together in one system?RAG supplies the facts and the fine-tune shapes how the model uses them, for example citing sources, sticking to retrieved context, or keeping a fixed output format.
When would you pick continued pretraining over supervised fine-tuning, and what does it take?Point to a base model that can't read the domain's text or notation. Then name the cost, a large unlabeled domain corpus, and the risk of forgetting.
Your prompt-plus-RAG system works but is slow and costly. How could fine-tuning help, and how do you prove it?Fine-tuning moves long instructions and examples into the weights, so prompts shrink or a smaller model can do the job. Check it on the same eval first.
What you can say
I think of it as a ladder of cost and commitment. You start on the cheapest rung and climb only when the one below clearly fails.
Prompting changes nothing lasting, only each call's instructions. I always start there, and I stay there if I can put the behavior into words.
RAG changes the facts in the context window. That's what I use for knowledge that's private, fresh or always changing.
Fine-tuning changes the weights, so the behavior sticks. I reach for it when the prompts work but are long, slow or unreliable.
Continued pretraining changes base knowledge and language. It's for a model that barely understands your domain's text or notation, and it's rare because it needs billions of tokens.
RAG changes what the model answers with, and fine-tuning changes how it answers. They stack, so most systems are prompting plus RAG, with a fine-tune added later for cost or consistency.
Weak answers to avoid
Jumps straight to fine-tuningFine-tuning is costly and slow to iterate on. Say you'd start with prompting and an eval set, and climb only when the cheaper rung measurably fails.
Fine-tunes to teach the model factsPrivate or changing facts belong in retrieval. Fine-tuning changes how the model answers, not reliably what it knows, and it goes stale after every update.
Treats the options as either/orThey stack. Most production systems are prompting plus RAG, with a fine-tune added later, so describe the combination instead of picking one winner.
Defines each option but never says whenThe question asks when. Tie each rung to its trigger, such as behavior you can describe, private or changing knowledge, unreliable prompts, or notation the model can't read.
Interview lens
Likely follow-ups, what you can say, and the weak answers to avoid.
