AWS AI Practitioner Applications of Foundation Models: 311 practice questions

12 of the 311 Applications of Foundation Models questions in the Certsqill AWS AI Practitioner bank, shown in full below. Each one carries an explanation for every option, not just the correct one — the wrong answers are where the marks go.

Preparing for AWS AI Practitioner? Take the free 5-min readiness check →

1. Prefer a smaller suitable model for lower response latency: Which approach is most appropriate?

1. Add retrieval so every response performs an additional document lookup.

   Retrieval can improve grounding, but unnecessary lookups may add latency when the team has no document requirement.
2. Use an agent that plans several tool calls before answering each request.

   Planning and tool calls can increase latency, making an agent unsuitable when fast direct responses are the priority.
3. Use a larger model for broader capability despite the latency priority.

   A larger model may add latency without providing needed value when quality is already adequate and complex reasoning is unnecessary.
4. Prefer a smaller suitable model for lower response latency. ✓

   A smaller suitable model can reduce response latency when its quality meets the interactive requirement.

When quality is adequate and complex context is unnecessary, a smaller suitable model often better meets interactive latency needs.

2. Choose a model with a context window that fits: Which choice best addresses the requirement?

1. Fine-tune the model on contract examples to expand context

   Fine-tuning can adapt behavior or style, but does not ordinarily expand the context window.
2. Lower temperature to accept more contract text

   Temperature affects sampling variability, not the maximum context the model can process.
3. Choose a model with a context window that fits the contract ✓

   The complete contract must fit within the model’s usable context window for one-request processing.
4. Use embeddings alone instead of supplying the contract context

   Embeddings support similarity tasks but do not let a generative model process the complete contract directly.

Choose a model whose context window can accommodate the complete contract.

3. Reduce temperature when generating responses: What is the best approach?

1. Fine-tune the model on additional factual documents

   Fine-tuning can adapt behavior or style, but adding facts does not specifically provide the requested variability control.
2. Use retrieval to supply changing reference information

   Retrieval improves access to external information but does not directly control sampling variability between otherwise similar responses.
3. Add an agent that selects different verification tools

   Tool selection may improve task completion, but agent planning does not directly reduce variation in model sampling.
4. Reduce temperature when generating responses ✓

   Lower temperature usually reduces sampling variability, making repeated outputs more consistent without guaranteeing truth or identical results.

Reducing temperature generally lowers sampling variability, but it does not guarantee factual accuracy or fully deterministic outputs.

4. Use a supported maximum output-length control: What is the best approach?

1. Retrieve only the most relevant article passages

   Retrieving fewer passages can reduce input context, but it does not directly impose a maximum generated response length.
2. Use a supported maximum output-length control ✓

   A supported output-length limit directly constrains generated content, making it appropriate when downstream systems require a defined maximum.
3. Fine-tune the model using short caption examples

   Examples may encourage concise style, but learned behavior alone does not reliably enforce a strict response-length boundary.
4. Increase the model's context window for longer articles

   A larger context window accommodates more input and possible output context, but does not itself limit generated length.

A supported maximum output-length control directly addresses a strict response boundary; retrieval and model capacity address different constraints.

5. Use RAG with refreshed policy sources: What is the best approach?

1. Use RAG with refreshed policy sources ✓

   RAG retrieves current external information at inference time without ordinarily updating model weights.
2. Use an agent to determine whether policies are correct

   An agent can plan and use tools, but agentic behavior does not inherently provide current documents or verify policy truth.
3. Lower temperature to make policy answers more reliable

   Lower temperature can reduce sampling variability, but it does not provide current policy information or guarantee factual answers.
4. Fine-tune the model whenever a policy changes

   Fine-tuning can adapt behavior, but frequently changing factual policy content is generally better handled through refreshed retrieval.

RAG supplies refreshed policy information during inference without repeatedly updating model weights.

6. Retrieve the approved policies as inference context: Which approach best fits?

1. Provide a few policy examples in every prompt

   In-context examples can guide responses temporarily, but they are less systematic for maintaining a growing authorized policy collection.
2. Use an agent framework to permanently teach the model

   Agent frameworks coordinate models, tools, and workflows; they do not inherently update foundation-model weights with policies.
3. Fine-tune the model on every approved policy

   Fine-tuning changes parameters and can suit behavior adaptation, but monthly factual updates are usually better handled through retrieval.
4. Retrieve the approved policies as inference context ✓

   Retrieval supplies current policies during inference without ordinarily changing the model's learned parameters or behavior.

Retrieval provides changing policy facts at inference time, whereas fine-tuning is intended mainly for persistent behavior or task adaptation.

7. Similarity search over vectors: Which capability should it use?

1. Similarity search over vectors ✓

   Vector similarity search compares numerical representations to identify content that is semantically related to the query.
2. Temperature adjustment during generation

   Temperature affects response sampling variability after retrieval or prompting, not the comparison of embedded document meanings.
3. Supervised fine-tuning with labeled examples

   Fine-tuning changes model behavior using examples, but it does not itself perform similarity retrieval across stored document vectors.
4. Agent planning across business tools

   Agent planning selects actions and tools, but it is not the underlying capability for comparing stored embeddings.

Vector similarity search compares embeddings to retrieve semantically related content from a vector-capable storage system.

8. Use smaller, coherent chunks while preserving needed: What adjustment is most appropriate?

1. Increase temperature to compensate for broad retrieved passages

   Temperature changes sampling variability and cannot remove irrelevant material from retrieved context or improve chunk boundaries.
2. Fine-tune on the unrelated passages to improve retrieval

   Fine-tuning may adapt behavior, but it does not correct poorly sized retrieval units or remove irrelevant retrieved content.
3. Expand every chunk to include complete documents

   Larger chunks generally add more unrelated content, worsening the stated relevance problem and consuming available context.
4. Use smaller, coherent chunks while preserving needed context ✓

   Smaller coherent chunks can improve relevance by reducing unrelated material, while preserving enough context for interpretation.

Chunking balances relevance and context; smaller coherent units can reduce unrelated retrieval material without discarding necessary meaning.

9. Enforce authorization during retrieval: What should the company prioritize?

1. Add citations after unrestricted retrieval

   Citations may identify sources, but adding them after retrieval does not prevent unauthorized content from being exposed.
2. Lower temperature for safer course responses

   Temperature affects sampling variability and cannot determine whether a learner may access particular course material.
3. Rely on embeddings to hide restricted passages

   Embeddings support similarity tasks but do not establish permissions, encryption, or authorization boundaries.
4. Enforce authorization during retrieval ✓

   Access checks must restrict which source content can be retrieved, preventing unauthorized passages from entering the model's context.

Authorization must constrain retrieval itself; embeddings, citations, and temperature do not establish user permissions.

10. Refresh the index after approved source updates: What action best addresses the problem?

1. Reduce temperature to make summaries more current

   Temperature changes response variability but cannot add articles or refresh outdated retrieval content.
2. Increase the model's context window

   A larger context window permits more content, but it does not update a stale index or introduce new articles.
3. Fine-tune the model on yesterday's summaries

   Training on old summaries cannot provide newly approved articles and may reinforce outdated information.
4. Refresh the index after approved source updates ✓

   Updating ingestion and retrieval data makes newly approved articles available as current inference context.

Refreshing ingestion and the retrieval index makes newly approved articles available; generation settings and context capacity cannot update stale sources.

11. Verify each cited passage against the generated claim: What practice is best?

1. Assume every citation proves the answer is correct

   Citations aid checking but do not prove source accuracy, interpretation, completeness, or generated-answer correctness.
2. Lower temperature and skip source review

   Lower temperature may reduce variability, but it does not verify evidence or justify removing human or automated review.
3. Verify each cited passage against the generated claim ✓

   Checking cited source content against the answer tests whether the evidence actually supports the generated statement.
4. Fine-tune the model until citations appear more often

   More frequent citations do not ensure that cited passages support claims or that the underlying documents are accurate.

Citations support verification, but the team must compare each cited passage with the generated claim rather than trust citation presence.

12. Customize the model with representative support examples: What is the best approach?

1. Use an agent to select support tools for each response

   Agents can plan tool use, but selecting tools does not inherently establish a persistent writing style or support format.
2. Customize the model with representative support examples ✓

   Supervised customization can adapt task behavior, style, and formatting when representative examples and stable requirements are available.
3. Use retrieval of the stable policies alone

   Retrieval supplies policy information, but it does not necessarily create consistent adoption of a distinctive tone and response format.
4. Lower temperature during every support interaction

   Lower temperature may reduce variation, but it does not specifically teach the model the desired brand style or format.

Customization best fits stable style and format adaptation, while retrieval, temperature, and agents address knowledge, variability, or orchestration.

Other AWS AI Practitioner domains

• Fundamentals of GenAI — 269 questions →
• Fundamentals of AI and ML — 223 questions →
• Security, Compliance, and Governance for AI Solutions — 157 questions →
• Guidelines for Responsible AI — 156 questions →
• All 1116 AWS AI Practitioner questions →