Photo by Vishnu Mohanan on Unsplash
Prompting, RAG or fine-tuning: which one is the scenario asking for?
by Miles Corran · October 22, 2026
Domain 3, Applications of Foundation Models, is the largest part of the AWS Certified AI Practitioner (AIF-C01) exam: 28% of scored questions, according to AWS's exam guide. Many of its scenarios come down to one decision: a company has a foundation model that isn't quite doing what it needs. How should it fix that?
There are three main options, and they get more expensive in this order: better prompts, retrieval-augmented generation (RAG) and fine-tuning. The trick is matching the problem in the scenario to the cheapest option that solves it.
Option 1: Prompt engineering
What it is: changing what you send the model, without changing the model. That includes clear instructions, a role ("You are a support agent for..."), examples of good answers, and asking the model to work through a problem step by step.
A few prompting terms the exam uses:
- Zero-shot: the prompt gives instructions but no examples.
- Few-shot: the prompt includes a handful of worked examples for the model to follow.
- Chain-of-thought: the prompt asks the model to reason step by step before answering.
Reach for it when the model has the knowledge it needs but the format, tone or focus of its answers is wrong. It's the fastest and cheapest option, so it's usually the first thing to try.
It won't help when the model simply doesn't know the information, because it was never in its training data.
Option 2: Retrieval-augmented generation (RAG)
What it is: before the model answers, the system searches your own documents for relevant passages and adds them to the prompt. The model then answers using that retrieved text. The model itself doesn't change.
Under the hood, documents are usually split into chunks and turned into embeddings (lists of numbers that capture meaning) and stored in a vector database, so the system can find passages that are similar in meaning to the question, not just ones that share keywords.
On AWS, Amazon Bedrock Knowledge Bases is the managed way to build this: you point it at your data, and it handles the chunking, embeddings and retrieval.
Reach for it when the scenario mentions:
- company-specific or private information the model was never trained on (internal policies, product manuals, support tickets);
- information that changes often, so the answers must stay current;
- a need to cite sources or reduce hallucinations by grounding answers in real documents.
Watch out for the phrase "without retraining the model". That's a strong hint the answer is RAG.
Option 3: Fine-tuning
What it is: training an existing foundation model further on your own labelled examples, so the model itself changes. Amazon Bedrock supports customising some models this way, and Amazon SageMaker AI offers more control for teams that want to manage training themselves.
A related idea is continued pre-training, where you train the model further on a large amount of unlabelled text from your domain, to make it more familiar with that domain's language.
Reach for it when the scenario needs the model to behave differently in a consistent way: a specialised style, a narrow task it does repeatedly, or domain language that prompting alone doesn't fix.
The trade-offs: fine-tuning needs good training data, costs more and takes longer than the other two options. And the model only knows what it was trained on, so if the information changes next week, a fine-tuned model is out of date until you train it again.
The decision in one table
| The scenario says... | Most likely answer |
|---|---|
| The answers are in the wrong format or tone | Prompt engineering |
| The model needs our internal documents to answer | RAG |
| The information changes often and must be current | RAG |
| We need answers to cite their sources | RAG |
| The model must consistently follow a specialised style or task | Fine-tuning |
| We need the cheapest, fastest improvement | Prompt engineering |
When two options could both work, the exam usually rewards the least complex and lowest-cost one that meets the requirement.
Test yourself
- A retailer wants its chatbot to answer questions using its returns policy, which changes every quarter. Which option fits best?
- A company's model gives correct answers, but they're too long and too informal for customer emails. What should it try first?
- A legal team wants a model that always drafts clauses in its firm's particular house style, across thousands of documents. Which option fits best?
(Answers: RAG; prompt engineering; fine-tuning.)
CertBrain's Domain 3 course covers all three in more depth, with scenario questions in the exam's style. See the courses, or take the free diagnostic to find out which domain needs the most work.
CertBrain is an independent study resource. It isn't affiliated with or endorsed by Amazon Web Services. AWS and AWS Certified AI Practitioner are trademarks of Amazon.com, Inc. or its affiliates.