AIF-C01 - Applications of foundation models (28% of the exam) - Section 3.2
Explain retrieval-augmented generation and when to use it over fine-tuning.
Use retrieval-augmented generation to ground answers in current documents without retraining, and recognise that a frequently changing source needs a re-index rather than a fine-tuning run. Contrast this with fine-tuning, which bakes knowledge into weights.
Retrieval-augmented generationFine-tuning vs RAGVector store
Practice question for this objective
lock_openFree sampleApplications of foundation modelsmedium
A travel firm runs a customer assistant on a fluent general-purpose foundation model. Its fares and route timetables change several times a week, and every answer must reflect the live schedule rather than stale figures baked in at training time. The team is choosing between retrieval-augmented generation and fine-tuning. Which two factors most strongly favour retrieval-augmented generation over fine-tuning for this use case? Select TWO.
AThe source data changes several times a week, so answers must reflect content that can be updated without retraining the model.check_circle Correct
BThe assistant must adopt a distinctive brand voice and reply format that the base model does not produce on its own.
CThe team wants the timetable knowledge stored permanently inside the model weights so no external store is queried at run time.
DEach answer should be grounded in retrievable source passages the team can update independently of the model itself.check_circle Correct
EThe base model already lacks fluency on travel topics and needs its core language ability rebuilt from scratch.
Retrieval-augmented generation suits knowledge that changes frequently because answers can be grounded in an external source updated independently of the model. RAG retrieves relevant passages from an external store at query time and conditions the answer on them, so updating the store immediately changes future answers. Fine-tuning instead writes knowledge into the weights, which becomes stale the moment the underlying data changes and demands a retrain to refresh. Frequently changing source data is therefore the classic signal to prefer RAG.
Why A is correct: RAG fetches current passages at query time, so refreshing the index keeps answers current, whereas fine-tuning would need a costly retrain on every schedule change.
Why B is wrong: Shaping tone and structure is a behaviour-adaptation goal better served by fine-tuning or prompt design, not by retrieving fresh documents.
Why C is wrong: Baking knowledge into weights describes fine-tuning, and it suits static facts, so it argues against RAG rather than for it.
Why D is correct: Grounding replies in an external, separately maintained document set is the defining benefit of RAG and decouples knowledge updates from the model.
Why E is wrong: The stem states the model is already fluent, so rebuilding language ability is unnecessary and would point to pre-training, not RAG.
Answers that look right on this material and are not. Each one is a distractor from a different question in the AIF-C01 bank for this domain.
Schedule a periodic retrain of the foundation model's weights so the latest policy text is absorbed into the model.
errorWhy it is wrong: Retraining bakes knowledge into the weights, the slow and costly path RAG is meant to avoid, and it leaves the model changed rather than untouched.
The need to reshape the model's output style to match the firm's internal report format.
errorWhy it is wrong: Tempting because formatting is a real requirement, but output style is exactly what fine-tuning excels at - teaching the model a new tone or structure through labelled examples. It is not a RAG strength.
Add a retrieval-augmented generation layer so the model fetches the latest house-style examples from a vector store on every request.
errorWhy it is wrong: RAG injects retrieved knowledge into the prompt, which suits changing facts; it does not reliably reshape the model's default formatting behaviour, and the requirement here is structural style rather than fresh information.