AWS Certified AI Practitioner AIF-C01 Practice Test 35
Question 340
An internal prompt includes confidential API keys that users might echo back. A pilot uses this approach instead: Rely solely on hiding secret text in a longer prompt. Which limitation MOST DIRECTLY explains why the requirement is not met?
- Sampling randomness does not revoke exposed credentials
- Prompts may be logged or exposed and should not contain embedded credentials
- Obscurity is not credential protection
- Refusal behavior cannot replace credential rotation and least privilege
Explanation
Correct Answer
C:This is the actual failure mode of the proposed approach: Rely solely on hiding secret text in a longer prompt. Obscurity is not credential protection
Incorrect Answers
A:This limitation concerns Raise temperature so the model forgets the secret; it does not diagnose the proposed Rely solely on hiding secret text in a longer prompt approach.
B:Prompts may be logged or exposed and should not contain embedded credentials This explains the recommended design, but it is not the requested diagnosis of Rely solely on hiding secret text in a longer prompt.
D:This limitation concerns Assume model refusal alone makes a leaked key invalid; it does not diagnose the proposed Rely solely on hiding secret text in a longer prompt approach.
Question 341
A company needs to compare two Bedrock FMs on accuracy, safety and human preference. A pilot uses this approach instead: Choose by provider brand alone. Which limitation MOST DIRECTLY explains why the requirement is not met?
- Capacity does not establish safety or correctness
- Bedrock evaluation supports task-specific model comparisons
- Cheap unsafe output can fail the business requirement
- Brand identity does not prove task performance
Explanation
Correct Answer
D:This is the actual failure mode of the proposed approach: Choose by provider brand alone. Brand identity does not prove task performance
Incorrect Answers
A:This limitation concerns Select by context length alone; it does not diagnose the proposed Choose by provider brand alone approach.
B:Bedrock evaluation supports task-specific model comparisons This explains the recommended design, but it is not the requested diagnosis of Choose by provider brand alone.
C:This limitation concerns Use cost only and omit quality measures; it does not diagnose the proposed Choose by provider brand alone approach.
Question 342
A team repeatedly edits instructions used by three Bedrock applications and needs a common revision history. Before approving a rollout, an architect suggests Write prompts into CloudTrail as the only prompt-template system. Which mismatch should be documented for this precise design?
- Ad hoc prompts lack shared version governance
- Sampling settings do not track who approved a prompt
- A managed prompt resource supports controlled prompt iteration and versions
- Audit logs record API activities rather than manage reusable prompts
Explanation
Correct Answer
D:This is the actual failure mode of the proposed approach: Write prompts into CloudTrail as the only prompt-template system. Audit logs record API activities rather than manage reusable prompts
Incorrect Answers
A:This limitation concerns Store only untracked text snippets in personal chat histories; it does not diagnose the proposed Write prompts into CloudTrail as the only prompt-template system approach.
B:This limitation concerns Change model temperature to create an audit history; it does not diagnose the proposed Write prompts into CloudTrail as the only prompt-template system approach.
C:A managed prompt resource supports controlled prompt iteration and versions This explains the recommended design, but it is not the requested diagnosis of Write prompts into CloudTrail as the only prompt-template system.
Question 343
A document search app needs managed search across embedding vectors generated from product descriptions. Before approving a rollout, an architect suggests Use AWS Config compliance timelines as a vector store. Which mismatch should be documented for this precise design?
- Audit logs are not designed as the search similarity index
- Vector search supports similarity-based candidate retrieval
- Encryption key storage cannot evaluate nearest-neighbor queries
- Resource-configuration history is not a vector similarity engine
Explanation
Correct Answer
D:This is the actual failure mode of the proposed approach: Use AWS Config compliance timelines as a vector store. Resource-configuration history is not a vector similarity engine
Incorrect Answers
A:This limitation concerns Put only prose in CloudTrail audit logs for semantic lookup; it does not diagnose the proposed Use AWS Config compliance timelines as a vector store approach.
B:Vector search supports similarity-based candidate retrieval This explains the recommended design, but it is not the requested diagnosis of Use AWS Config compliance timelines as a vector store.
C:This limitation concerns Use KMS keys as the embedding database; it does not diagnose the proposed Use AWS Config compliance timelines as a vector store approach.
Question 344
An assistant has a strict 300 millisecond response budget and a fixed monthly inference allowance. During a design review, someone proposes Select by price without testing task accuracy. Which technical objection is specifically relevant to this proposal?
- Larger parameter count does not guarantee lower latency
- Cheap incorrect answers fail the application requirement
- Training corpus size does not establish served latency
- A model that meets the task with less compute can be preferable
Explanation
Correct Answer
B:This is the actual failure mode of the proposed approach: Select by price without testing task accuracy. Cheap incorrect answers fail the application requirement
Incorrect Answers
A:This limitation concerns Assume largest available model is always most responsive; it does not diagnose the proposed Select by price without testing task accuracy approach.
C:This limitation concerns Choose solely by advertised training dataset size; it does not diagnose the proposed Select by price without testing task accuracy approach.
D:A model that meets the task with less compute can be preferable This explains the recommended design, but it is not the requested diagnosis of Select by price without testing task accuracy.
Question 345
A large FM is too costly to serve, but a smaller model could learn some of its task behavior. A pilot uses this approach instead: Assume distillation means indexing source PDFs in a vector DB. Which limitation MOST DIRECTLY explains why the requirement is not met?
- Distillation can transfer useful behavior to a more efficient model
- A larger teacher alone does not solve serving-cost constraints
- Retrieval indexing is not teacher-student model training
- More context does not reduce model parameters
Explanation
Correct Answer
C:This is the actual failure mode of the proposed approach: Assume distillation means indexing source PDFs in a vector DB. Retrieval indexing is not teacher-student model training
Incorrect Answers
A:Distillation can transfer useful behavior to a more efficient model This explains the recommended design, but it is not the requested diagnosis of Assume distillation means indexing source PDFs in a vector DB.
B:This limitation concerns Replace the small student model with a larger untested teacher in production; it does not diagnose the proposed Assume distillation means indexing source PDFs in a vector DB approach.
D:This limitation concerns Expand the context window as the entire model compression method; it does not diagnose the proposed Assume distillation means indexing source PDFs in a vector DB approach.
Question 346
A model fine-tuned on a narrow support sample produces poor answers for a neglected customer segment. During a design review, someone proposes Use only model size as a fairness measure. Which technical objection is specifically relevant to this proposal?
- More duplicates do not add missing segment coverage
- Removing evaluation hides generalization problems
- Parameter count does not measure group performance
- Coverage of real user scenarios reduces harmful specialization gaps
Explanation
Correct Answer
C:This is the actual failure mode of the proposed approach: Use only model size as a fairness measure. Parameter count does not measure group performance
Incorrect Answers
A:This limitation concerns Repeat the same narrow examples more often; it does not diagnose the proposed Use only model size as a fairness measure approach.
B:This limitation concerns Remove the holdout set to make training accuracy look better; it does not diagnose the proposed Use only model size as a fairness measure approach.
D:Coverage of real user scenarios reduces harmful specialization gaps This explains the recommended design, but it is not the requested diagnosis of Use only model size as a fairness measure.
Question 347
A customer service bot needs accurate answers from an internal policy collection that changes weekly. A pilot uses this approach instead: Fine-tune on one static snapshot as the only update mechanism. Which limitation MOST DIRECTLY explains why the requirement is not met?
- Static weights may not reflect weekly document changes
- Knowledge Bases can retrieve relevant documents to ground generated replies
- Randomness does not provide current policy evidence
- Pretrained data does not contain confidential current policies by default
Explanation
Correct Answer
A:This is the actual failure mode of the proposed approach: Fine-tune on one static snapshot as the only update mechanism. Static weights may not reflect weekly document changes
Incorrect Answers
B:Knowledge Bases can retrieve relevant documents to ground generated replies This explains the recommended design, but it is not the requested diagnosis of Fine-tune on one static snapshot as the only update mechanism.
C:This limitation concerns Increase temperature until the model recalls current internal clauses; it does not diagnose the proposed Fine-tune on one static snapshot as the only update mechanism approach.
D:This limitation concerns Rely only on public pretraining for private documents; it does not diagnose the proposed Fine-tune on one static snapshot as the only update mechanism approach.
Question 348
A document search app needs managed search across embedding vectors generated from product descriptions. Which option BEST meets the business requirement?
- Put only prose in CloudTrail audit logs for semantic lookup
- Use AWS Config compliance timelines as a vector store
- Store embeddings in a suitable vector-enabled search datastore such as Amazon OpenSearch Service
- Use KMS keys as the embedding database
Explanation
Correct Answer
C:Vector search supports similarity-based candidate retrieval The specified scenario makes this the most appropriate approach.
Incorrect Answers
A:Audit logs are not designed as the search similarity index This prevents the approach from meeting the stated requirement.
B:Resource-configuration history is not a vector similarity engine This prevents the approach from meeting the stated requirement.
D:Encryption key storage cannot evaluate nearest-neighbor queries This prevents the approach from meeting the stated requirement.
Question 349
An assistant has a strict 300 millisecond response budget and a fixed monthly inference allowance. Before approving a rollout, an architect suggests Choose solely by advertised training dataset size. Which mismatch should be documented for this precise design?
- A model that meets the task with less compute can be preferable
- Cheap incorrect answers fail the application requirement
- Larger parameter count does not guarantee lower latency
- Training corpus size does not establish served latency
Explanation
Correct Answer
D:This is the actual failure mode of the proposed approach: Choose solely by advertised training dataset size. Training corpus size does not establish served latency
Incorrect Answers
A:A model that meets the task with less compute can be preferable This explains the recommended design, but it is not the requested diagnosis of Choose solely by advertised training dataset size.
B:This limitation concerns Select by price without testing task accuracy; it does not diagnose the proposed Choose solely by advertised training dataset size approach.
C:This limitation concerns Assume largest available model is always most responsive; it does not diagnose the proposed Choose solely by advertised training dataset size approach.
Question 350
An evaluation team wants scalable subjective feedback on FM answers before human sampling. During a design review, someone proposes Use the price per token as the sole quality score. Which technical objection is specifically relevant to this proposal?
- Cost does not judge whether the response is correct
- Self-reported confidence may be poorly calibrated
- A judging model can scale rubric-based scoring but may have systematic bias
- Automated judges can be biased and should be calibrated
Explanation
Correct Answer
A:This is the actual failure mode of the proposed approach: Use the price per token as the sole quality score. Cost does not judge whether the response is correct
Incorrect Answers
B:This limitation concerns Use only the model's confidence in its own answer; it does not diagnose the proposed Use the price per token as the sole quality score approach.
C:A judging model can scale rubric-based scoring but may have systematic bias This explains the recommended design, but it is not the requested diagnosis of Use the price per token as the sole quality score.
D:This limitation concerns Treat judge outputs as infallible ground truth; it does not diagnose the proposed Use the price per token as the sole quality score approach.
