A company wants an AI assistant to answer questions about its internal documents. Another wants a model to respond in a more consistent specialized format. Those requests may sound similar because both mention customization, but they emphasize different problems.
For AIF-C01 preparation, ask whether the application needs access to changing information, an adjustment to model behavior, or both. Start with the requirement before choosing the most sophisticated-sounding technique.
A policy assistant with frequently updated documents
Consider an original scenario: a support team updates its policies every week. Staff need answers grounded in the current approved documents and want to see where an answer came from.
Retrieval-augmented generation is a relevant approach because the application can retrieve pertinent material and supply it as context when generating a response. Amazon Bedrock Knowledge Bases documentation describes retrieval from connected data sources and the use of that information in responses.
The deciding requirement is access to maintained source information. Fine-tuning is not automatically the best way to make a model remember every new policy revision. A model trained on yesterday’s examples does not acquire today’s document changes merely because the training was expensive.
A consistent output behavior
Now consider a separate team with a well-defined response style and representative examples. The team wants the model to follow that behavior more consistently. Fine-tuning may be a relevant customization option, depending on the model, the task and available support.
That does not mean every formatting problem requires fine-tuning. A clearer prompt, examples in context or application-level validation may be sufficient. Compare simpler approaches before assuming model training is necessary. AWS discusses customization approaches in its AI Practitioner exam guide.
Retrieval does not guarantee truth
A RAG answer can still be wrong if retrieval selects the wrong document, the source is outdated, important context is missing or the generated answer misstates the evidence. A citation is something a reviewer can inspect; it is not proof that the cited passage supports every sentence.
For the policy-assistant example, write three checks: Is the document current? Is it relevant to the question? Does the answer follow from it? Those checks create a more useful review plan than “add a knowledge base and accuracy is solved.”
Try the AWS AI Practitioner scenario questions after making that distinction. Explain which phrase in the business requirement supports your chosen method and which phrase would point toward an alternative.
A comparison card that avoids false choices
Divide a card into information access, behavior adjustment, maintenance and evaluation. Put the policy scenario under information access and the specialized-output scenario under behavior adjustment. Add a note that an application can combine techniques when its requirements justify them.
The AWS AI Practitioner model-customization guide can help connect the comparison with prompting, foundation models and evaluation. For exam practice, the useful skill is selecting the approach that addresses the stated need while recognizing what that approach does not automatically solve.