Amazon AWS Certified Generative AI Developer - Professional AIP-C01 Practice Test 5

Question 40

A just-deployed custom model begins producing unsafe finance recommendations; the prior approved model package passed the same tests. What immediate release action is appropriate?

Increase serving instances to make the new model respond faster.
Mark the failing package as approved because its artifact can load.
Route production traffic back to the last validated model package and investigate the failing safety-regression cohort before another promotion.
Raise response-token limits to suppress unsafe advice.

Explanation

Correct Answer

A model-quality regression requires restoring the verified version rather than merely scaling. Versioned SageMaker deployment and evaluation evidence allow safe rollback and root-cause analysis.

Incorrect Answers

Additional capacity does not correct unsafe generated content.

Loading successfully is not the same as passing safety acceptance; a failing model should not be promoted.

Output length does not enforce safe recommendations or resolve a detected behavioral regression.

Question 41

A real-time SageMaker model endpoint must absorb busy-hour growth; customer messages range from very short to very long. What metric-driven capacity strategy should the engineer use?

Load-test tokens per request and request concurrency, measure latency/utilization, and configure supported endpoint autoscaling around actual serving capacity.
Assume 1,000 requests per minute always consume the same compute independently of prompt/output length.
Increase embedding index shard count to autoscale the separate inference endpoint.
Select any cross-Region inference profile without checking whether it supports this SageMaker endpoint.

Explanation

Correct Answer

SageMaker real-time endpoints serve deployed models with configured instances and production variants. Scaling needs empirical request sizes and p95 latency, not a raw request count that ignores token demand.

Incorrect Answers

Token volume and generation length can dominate throughput and create tail-latency breaches at equal request rates.

Index sharding does not add instances to the deployed generative inference variant.

Bedrock cross-Region inference profiles do not directly manage SageMaker real-time endpoint instance autoscaling.

Question 42

A proprietary fine-tuned model artifact is not listed among Bedrock foundation models. Before committing to SageMaker real-time hosting, which compatibility test matters?

Change the Bedrock model ID to a guessed string that resembles the proprietary artifact.
Upload the weight files to an S3 bucket and infer that every default container can deserialize them.
Use a cross-Region inference profile to route to an unregistered proprietary model.
Verify that the model artifact, inference container, accelerator/runtime requirements and endpoint serving contract are supported by the chosen SageMaker deployment.

Explanation

Correct Answer

SageMaker can host custom containers/artifacts, but successful hosting depends on compatible runtime, hardware and inference handlers. Bedrock catalog absence does not by itself prove any artifact can be served.

Incorrect Answers

An invented Bedrock identifier cannot make an unregistered model available for invocation.

S3 storage does not supply a compatible model server, libraries or hardware capability.

A Bedrock inference profile routes supported Bedrock models, not arbitrary unregistered SageMaker artifacts.

Question 43

An application currently hardcodes provider-specific model calls in five microservices. Procurement plans to change vendors without redeploying those services. What architecture meets the requirement?

Let every microservice select its preferred vendor from a local static list.
Make the old model's outputs longer so they resemble those of another vendor.
Expose a stable app-facing inference contract through a model-neutral gateway with tested provider/version adapters behind it.
Skip acceptance tests because the replacement provider advertises the same general modality.

Explanation

Correct Answer

A model-neutral gateway can translate supported parameters, normalize error handling and preserve the application contract while routing only to approved model implementations.

Incorrect Answers

Model switches still require service-level code/config changes and provide no centrally validated contract.

Output length does not establish request/response schema or version compatibility.

Models vary in supported APIs, tokenization, tool behavior and output schema; each adapter needs regression evaluation.

Question 44

Following a model version upgrade, a payment service receives a different JSON field layout. The older parser accepts only the previous contract. Which release control prevents silent corruption?

Treat a successful HTTP 200 inference response as proof that the JSON fields are unchanged.
Contract-test the upgraded model's supported structured output against the consuming schema and gate production routing on the tested version.
Skip schema validation and pass unknown JSON fields directly into the payment database.
Only increase the response token limit and keep the old parser untouched.

Explanation

Correct Answer

A JSON schema is a transport/shape contract, while downstream business validation enforces permitted values. Model revisions must be regression-tested with the actual parser before route promotion.

Incorrect Answers

Transport success does not establish schema compatibility or accepted business semantics.

Unvalidated fields can change records or cause downstream failures; an upgraded provider must meet the existing contract.

Generation size cannot reconcile renamed or incompatible schema fields.

Question 45

A team proposes reserved Bedrock throughput for an experiment with low, unpredictable daily traffic. How should the financial decision be made?

Choose the highest available provisioned allocation because unused capacity improves answer quality.
Compare only the peak number of requests without accounting for token sizes and quiet periods.
Compare observed on-demand cost and request distribution with the minimum committed provisioned capacity and actual expected utilization.
Assume a cross-Region inference profile automatically discounts every model by the same percentage.

Explanation

Correct Answer

Provisioned capacity has commitment economics; sparse or unpredictable demand may not justify reserved throughput. Decision requires actual token distribution, utilization and eligible model capacity options.

Incorrect Answers

More committed throughput does not independently improve model accuracy and may create idle expense.

Peak requests hide cost and utilization of committed resources across the billing period.

Inference-profile pricing and deployment availability vary; generic routing is not a substitute for measured provisioned-vs-on-demand economics.

Question 46

A Glue ETL feed contains null customer_id values in records that must later be used as trusted GenAI context. Which pre-ingestion gate should reject the invalid rows?

Apply an IsComplete/Completeness rule to customer_id in Glue Data Quality and quarantine failed records before contextual indexing.
Check only that the customer_id column exists in the schema.
Use a vector similarity threshold to remove any rows with null customer_id.
Replace every missing customer_id with a single synthetic value and mark the feed clean.

Explanation

Correct Answer

A DQDL completeness test detects missing values in a required field; routing failures to quarantine prevents a model from presenting invalid rows as vetted customer records.

Incorrect Answers

A column can be present in every record definition and still hold null values; schema presence does not establish completeness.

Embedding similarity does not guarantee required scalar fields are populated, and missing identifiers can still enter the index.

A common placeholder destroys record identity and masks upstream completeness defects.

Question 47

An orders CSV has two distinct rows with the same transaction_id, causing retrieval to return contradictory updates. What data-quality test should run before indexing?

Run only an IsComplete check on transaction_id and accept all non-null duplicates.
Check transaction_id uniqueness using a DQDL IsUnique or equivalent primary-key constraint and isolate duplicate records.
Increase retrieval top-k until one of the duplicate transaction rows appears first.
Normalize transaction_id to lowercase and assume the duplicate collision disappears.

Explanation

Correct Answer

A uniqueness rule detects repeated business keys. The check belongs before embedding/indexing, where semantic similarity cannot distinguish a legitimate transaction from conflicting duplicate IDs.

Incorrect Answers

Completeness catches nulls, not repeated values; both rows can pass despite violating transaction identity.

Returning more candidate rows does not resolve which duplicate is authoritative.

Case conversion may make additional IDs collide and cannot prove unique ownership of a transaction.

Question 48

A support export mixes timestamps with Z offsets, local offsets, and missing zones. The GenAI incident timeline currently puts events in the wrong order. What should preprocessing enforce?

Sort all timestamps alphabetically without interpreting their offsets.
Parse timestamps with explicit source time-zone evidence, normalize to a canonical instant such as UTC, and retain the original offset and provenance.
Strip every time-zone suffix and treat the remaining local times as UTC.
Ask the model to infer the correct time zone from the incident description.

Explanation

Correct Answer

A wall-clock string alone is ambiguous across zones and daylight-saving transitions. Conversion with preserved source offset supports chronological ordering and subsequent evidence audit.

Incorrect Answers

Lexical ordering of differently offset local times does not represent chronological ordering.

This silently shifts real instants and can corrupt the incident sequence.

An FM guess cannot replace an explicit source offset or an accepted missing-zone remediation process.

Question 49

An ingestion pipeline reports success, but a ten-page contract PDF has extracted spans only from pages one through eight. What acceptance test catches the loss?

Compare parsed pages and source-span coverage with the PDF page inventory, failing ingestion when expected pages have no extracted evidence.
Check that the embedding index contains at least one vector for the contract.
Increase the retrieval result count while retaining the same incomplete parsed document.
Rewrite the model prompt to request summaries of all ten pages without checking extraction.

Explanation

Correct Answer

A completed job is not proof of complete source extraction. Document/page coverage and source offsets reveal silent truncation before incomplete legal context becomes retrievable.

Incorrect Answers

A single stored chunk can coexist with two entirely missing pages.

Retrieval can only return indexed content, so top-k changes cannot restore unparsed pages.

A prompt cannot generate reliable evidence for pages that were never ingested.

Question 50

An invoice archive mixes USD, EUR and GBP amounts that are labeled inconsistently before semantic indexing. What normalization avoids misleading comparisons?

Strip all currency symbols and embed only the number as text.
Parse the numeric amount with its original ISO currency and unit, store explicit normalized fields, and apply approved conversion only with a dated rate.
Assume every non-dollar amount is automatically USD at a one-to-one conversion.
Raise the embedding dimension until it represents currency conversions exactly.

Explanation

Correct Answer

An amount without its currency is not an equivalent quantity. Keeping source denomination and an auditable conversion date avoids comparing unrelated numerical values in retrieval-generated totals.

Incorrect Answers

The same digits in different currencies have different value; dropping currency metadata destroys their meaning.

An arbitrary exchange assumption changes financial values without a source or effective date.

Embedding dimensionality does not provide deterministic FX conversion or attach missing monetary units.

Pass Your Amazon AWS Certified Generative AI Developer - Professional AIP-C01 Exam Easy!

Question 40

Explanation

Correct Answer

Incorrect Answers

Question 41

Explanation

Correct Answer

Incorrect Answers

Question 42

Explanation

Correct Answer

Incorrect Answers

Question 43

Explanation

Correct Answer

Incorrect Answers

Question 44

Explanation

Correct Answer

Incorrect Answers

Question 45

Explanation

Correct Answer

Incorrect Answers

Question 46

Explanation

Correct Answer

Incorrect Answers

Question 47

Explanation

Correct Answer

Incorrect Answers

Question 48

Explanation

Correct Answer

Incorrect Answers

Question 49

Explanation

Correct Answer

Incorrect Answers

Question 50

Explanation

Correct Answer

Incorrect Answers

AWS Certified Generative AI Developer - Professional AIP-C01 Practice Tests