AIF-C01 applications of foundation models
Applications of Foundation Models is worth 28% of AIF-C01’s scored content — around 14 of the 50 scored questions and the largest domain. This is where the exam moves from “what is a foundation model” to “how would you actually use one”.
The central decision: RAG or fine-tuning
The most tested distinction on the exam. Both adapt a general model to your needs, and they solve different problems.
| Retrieval-augmented generation | Fine-tuning | |
|---|---|---|
| What it does | Retrieves your content at query time and gives it to the model | Adjusts the model’s weights with your examples |
| Good for | Current facts, changing content, citing sources | Style, tone, format, domain-specific behaviour |
| When content changes | Just update the index | Retrain |
| Cost | Lower to start, cost per query | Higher upfront, cheaper per query |
| Transparency | You can see what was retrieved | Opaque |
The rule to carry into the exam: if the requirement involves information that changes, or answers that must be grounded in your documents, it is RAG. If it involves the model behaving differently — a consistent tone, a house format, domain vocabulary — it is fine-tuning.
A scenario mentioning monthly updates, current data, or “must not be out of date” is a RAG question every time.
Embeddings and vector stores
The mechanism under RAG, at the level AIF-C01 asks:
- Content is split into chunks and converted into embeddings — numeric representations of meaning.
- Those are stored in a vector store.
- A query is embedded the same way, and semantically similar chunks are retrieved.
- Those chunks go into the prompt so the model answers from them.
The consequence worth knowing: when a RAG system gives a wrong answer, the usual cause is retrieval, not the model. It answered faithfully from the wrong content.
Choosing a model
A real decision with real trade-offs:
- Capability against cost against latency. Bigger is not better if a smaller model meets the requirement.
- Modality — does it need images or audio, or just text?
- Context window — how much must it consider at once?
- Whether a task needs generative AI at all, or whether a pre-trained service or traditional ML is cheaper and more reliable.
AWS’s expected instinct: the cheapest option that meets the stated requirement.
Evaluating output
Harder than evaluating a classifier, because there is often no single right answer. Concepts at this level:
- Human evaluation — people rate output against criteria.
- Automated metrics — comparison against reference answers.
- Grounding checks — is the answer supported by the retrieved content?
- Task-specific measures — did it do what was asked?
The conceptual point: “accuracy” is not a sufficient measure for generative output, and this is examinable.
Agents
Foundation models that go beyond answering: they use tools, call APIs and take multi-step actions. At foundational level, know that an agent can act rather than only respond, and that acting introduces risk — which is why human approval for consequential actions keeps appearing as the right answer.
Sample questions
Question 1. A law firm wants an assistant that answers from its own case documents, which are added to weekly, and must cite which document each answer came from. Which approach fits?
- A. Retrieval-augmented generation over an index of the case documents
- B. Fine-tune a foundation model on the case documents each week
- C. Select a model with a larger context window
- D. Increase the temperature to encourage detailed answers
Show answer
Answer: A
Retrieval-augmented generation retrieves current documents at query time and makes the source visible, satisfying both the weekly updates and the citation requirement. Fine-tuning would need retraining as documents are added and cannot cite sources, a larger context window does not supply the documents, and raising temperature is unrelated.
Question 2. A bank wants its assistant to consistently reply in its formal house style and use its internal terminology. The underlying facts come from a live system. Which combination is most appropriate?
- A. RAG only
- B. Fine-tuning only
- C. Lower the temperature and use a detailed system prompt
- D. Fine-tuning for style and terminology, with RAG for the live facts
Show answer
Answer: D
Style and terminology are behaviour, which fine-tuning addresses, while live facts must be retrieved at query time through RAG, so the requirement calls for both. Using only RAG leaves style inconsistent, only fine-tuning leaves facts stale, and lowering temperature does not teach house style.
Question 3. A RAG-based assistant produces fluent answers that cite the wrong product manual. Where should the team look first?
- A. The model’s size — switch to a larger foundation model
- B. The retrieval step — how content is chunked and matched
- C. The temperature setting
- D. The maximum response length
Show answer
Answer: B
Fluent answers drawn from the wrong source point at retrieval returning the wrong chunks, not at the model’s generation. A larger model would summarise the wrong content just as fluently, temperature affects variability, and more tokens in the response changes length rather than sourcing.
What to practise
Write down five business scenarios and decide RAG or fine-tuning for each, with a one-line reason. Then take the two that could plausibly need both and say why. That exercise covers the single most tested distinction in the largest domain of the exam.