Question 31
Section 3When should fine-tuning be preferred over RAG?
Correct answer: A
Explanation
Fine-tuning is preferred when the goal is to bake in a stable behavior, such as a repeated task style or output format, because it changes the model’s parameters rather than supplying external documents at runtime. RAG is better for adding or updating factual knowledge, but it does not reliably enforce a consistent style or structure when simple prompting fails.
Why each option is right or wrong
A. When the model needs to learn a stable, repeated task style or format that simple prompting cannot reliably produce
Fine-tuning is the right choice when the requirement is to make the model consistently produce the same task behavior, style, or output structure across many runs, especially after prompt-only attempts still vary. RAG is designed to retrieve external context at inference time, so it helps with up-to-date or domain-specific facts, but it does not alter the model’s weights and therefore cannot reliably lock in a repeated format or style the way training can.
B. When facts change every hour and must be reflected immediately
Rapidly changing facts are usually better handled by retrieval than retraining or fine-tuning.
C. When the goal is to cite source documents for every answer
Source citation aligns with retrieval workflows that return grounded documents alongside answers.
D. When the goal is to ground on a private knowledge base
Private knowledge base grounding is a classic RAG use case, not primarily a fine-tuning goal.