Amazon AWS Certified Generative AI Developer - Professional AIP-C01 Practice Test 3

Question 20

A planned GenAI application has not checked whether its selected model, inference endpoint and expected request volume are supported in its required AWS Region. What should the team validate before committing to the design?

Check exact model and runtime API availability, inference profile destinations, quotas and permissions in the deployment Region.
Assume every model offers identical modalities and capacity in all AWS Regions.
Choose a public model benchmark winner and postpone regional feasibility checks until after launch.
Use S3 cross-Region replication to guarantee that a Bedrock model exists in the target Region.

Explanation

Correct Answer

Catalog/Region, endpoint capability, IAM/SCP and service quota verification establishes that the proposed runtime can actually serve the business workload.

Incorrect Answers

Foundation model support differs by model version, API and Region; a universal-availability assumption is unsafe.

Offline benchmark quality cannot make an unavailable model callable or provide required throughput.

Replicating source objects does not provision or authorize a foundation model in another inference Region.

Question 21

A photo-enabled warranty assistant must inspect a product photo and a typed description together. The proposed model version accepts text but not image content. Which selection check must pass before deployment?

Enable a cross-Region inference profile for the existing text-only model.
Raise the text-only model's maximum output-token setting.
Store the photographs as S3 object URLs in the text prompt.
Verify that a model version offered in the intended Region supports image input and the selected Bedrock inference API before sending image content blocks.

Explanation

Correct Answer

Input modality is a capability of the exact foundation model and API, not of the application prompt. The model card and Region/API compatibility must show image input before an image block can be reasoned over.

Incorrect Answers

A profile changes eligible compute destinations, not the model's supported input modalities; a text-only model still cannot process image pixels.

Output length changes response capacity; it does not add image encoders or support the required image content type.

A URL string is still text; without an image-capable model and supported content block, the model has not examined the photographed evidence.

Question 22

An order-support chatbot must look up a current order and, with the caller's authorization, change its shipping address. The selected model only emits prose. What capability should the integration require?

Use a Bedrock Knowledge Base to make the model write the updated order to the source database.
Ask the model to print a JSON representation of an address-change request and display it as confirmation.
Choose a tool-capable model and API, pass a typed tool definition, then run the authorized backend operation and return its tool result for the model to incorporate.
Increase temperature until the model decides to modify the order.

Explanation

Correct Answer

Bedrock tool use is an interaction contract: the model requests a tool, while an application or supported server-side service executes a validated and authorized operation and supplies the matching result.

Incorrect Answers

Knowledge Bases retrieval supplies evidence; generated text does not execute or authorize a database mutation.

A syntactically plausible JSON request is not an executed order update; the system still needs an authenticated tool boundary and actual operation result.

Sampling settings affect generated tokens, not tool invocation support or backend authorization.

Question 23

An Arabic-language support desk requires production quality on Arabic dialect queries. A vendor advertises generic multilingual ability but has only been tested internally on English. What must the selection gate establish?

Verify only that the model can tokenize Unicode strings without an API error.
Test the exact available model version with representative Arabic prompts and labeled dialect-specific answers, checking supported language performance and failure rates.
Choose the largest listed context window instead of evaluating Arabic quality.
Translate every Arabic request into English without testing translation loss or data-processing permissions.

Explanation

Correct Answer

A generic multilingual description does not prove task-level Arabic fidelity. Use localized held-out evaluation on the specified model and Region rather than extrapolating from English or a different model revision.

Incorrect Answers

Successful tokenization proves input acceptance, not accurate comprehension or correct dialect-sensitive responses.

Context length is independent of multilingual semantic accuracy.

Unverified translation can change names, negation and regulatory wording; it does not establish the required Arabic performance.

Question 24

A claims classifier must return JSON containing category and one of three allowed status values. Free-form text has caused parser failures. What is the best supported integration design?

Use a prompt instruction that says "return only JSON" and remove downstream validation.
Set temperature to zero and accept any returned status string.
Use a model/API that supports constrained JSON Schema output for the response shape, and separately validate allowed claim states and authorization in application code.
Switch the classifier to a vector embedding model and interpret similarity scores as status fields.

Explanation

Correct Answer

Bedrock structured output can enforce a supported JSON Schema subset for compatible model/API pairs. Schema validity is distinct from business correctness, so status transitions still require deterministic downstream rules.

Incorrect Answers

Natural-language instructions alone do not provide the same schema-conformance guarantee and cannot validate business authorization.

Low temperature neither guarantees schema conformance nor restricts status values to authorized transitions.

Embedding similarity is not an application JSON contract or a business-state validator.

Question 25

One provider changes the model-specific request property for sampling. Other providers still accept the old request structure. How should a multi-provider application avoid repeated call-site edits?

Translate common application settings into the provider-specific inference request fields through tested per-model adapters.
Send both the old and new parameter names in every request to all models.
Increase the S3 read timeout for documents supplied to the model.
Change the vector-index metric so the provider interprets new parameters.

Explanation

Correct Answer

Bedrock inference parameters and provider-native payloads are not universally identical. Adapter tests isolate each model's accepted request schema from the app's stable interface.

Incorrect Answers

Unsupported or duplicate provider fields can be rejected, and broadening every payload does not isolate schema changes.

Storage timeouts do not address a provider inference-parameter schema mismatch.

Vector similarity indexing is unrelated to decoding parameter names accepted by a model inference API.

Question 26

A multimodal model responds to a prototype in us-east-1, but the deployment must serve from an approved caller Region where the model ID is not directly available. What should the team verify before deciding if that exact version can be invoked?

Check exact model/version availability in the caller Region and inspect supported inference profile destination Regions and invocation modes.
Assume an ARN copied from the working Region can be invoked directly in every other AWS Region.
Increase the model context window to make a Region-limited model globally available.
Move input files to the approved S3 bucket and assume the model deployment automatically follows them.

Explanation

Correct Answer

Model-version listings and inference-profile routing determine whether and where a request can execute; region and organization-policy checks are separate prerequisites.

Incorrect Answers

Regional catalogs and supported invocation types vary, so copying an identifier cannot create availability.

Context capacity does not change an unavailable model/endpoint or its geographic routing.

Moving stored data does not alter model availability or the compute Region selected by an inference profile.

Question 27

A foundation-model version in Bedrock is marked Legacy, and the company relies on that exact version in a customer workflow. What is the required lifecycle action?

Assume the original model provider's public retirement date overrides Bedrock's lifecycle notice.
Confirm the model's published Bedrock-specific EOL date and regional state, regression-test an Active successor, and migrate before invocation stops working.
Keep the Legacy model indefinitely by setting temperature to zero.
Create new Provisioned Throughput for the Legacy model as the default migration path.

Explanation

Correct Answer

AWS distinguishes Active, Legacy and EOL lifecycle states; new adoption and new Provisioned Throughput can be restricted in Legacy. Endpoints do not automatically migrate when a model reaches EOL.

Incorrect Answers

Bedrock version/Region EOL notices govern availability in the service, and those dates can differ from the underlying provider's announcements.

Decoding settings cannot extend service lifecycle or preserve an endpoint once the model reaches EOL.

AWS restricts creating new Provisioned Throughput for models in Legacy; a new Active target is the appropriate migration plan.

Question 28

A financial audit asks whether temperature=0 guarantees byte-identical model answers after a Bedrock model/version change. How should the architect evaluate the claim?

Measure exact repeatability on controlled inputs and versions, and do not promise identical responses across model upgrades based solely on sampling settings.
Assert temperature=0 makes different provider models return identical bytes for the same prompt.
Increase maximum tokens until all responses become identical.
Store previous output in the prompt and treat it as proof of model determinism.

Explanation

Correct Answer

Temperature controls sampling behavior, not changes in model weights, backend serving behavior or tokenizer/version implementations. Exact reproducibility must be measured for the model and API environment.

Incorrect Answers

Model architecture, weights and vocabulary differ; zero temperature is not a cross-model identity contract.

A larger generation budget may increase variability and does not enforce deterministic content.

Copying a past answer is not a controlled repeatability test and can conceal a new model's changed behavior.

Question 29

A legal summarization request contains 180,000 input tokens, while the chosen model's accepted context budget is smaller. What is the correct design response?

Set maxTokens to the entire input length and send the 180,000-token document unchanged.
Disable request-size validation to make the endpoint accept the oversized prompt.
Switch from cosine to Euclidean distance on an unrelated vector index without reducing context.
Check that version's supported input/output token limits and retrieve or segment the material into bounded context, reserving space for the response.

Explanation

Correct Answer

Model context capacity is a hard request constraint. Retrieval, segmentation and response-token reservation allow the application to stay within the actual model/API limit while retaining relevant evidence.

Incorrect Answers

maxTokens controls generated output, not permission to exceed the model's input/context window.

The service/model context limit still applies; bypassing local checks usually produces an API error.

A similarity metric change does not by itself reduce the number of input tokens submitted to inference.

Question 30

Two Bedrock foundation models have very different parameter counts. The proposed procurement decision is based only on which has more parameters. What evaluation is needed?

Select the model with the largest published parameter count and skip task testing.
Use different prompts and workloads for each model and compare the cheapest request.
Compare the number of available model cards without invoking either version.
Compare both available model versions on the same representative labeled tasks, using quality, p95 latency, token cost, model restrictions and permitted Regions.

Explanation

Correct Answer

Parameter count does not directly rank accuracy, cost or latency on a particular workload. A comparable evaluation population and regional/model constraints determine feasibility and business utility.

Incorrect Answers

Size is not evidence of superior task accuracy or an acceptable latency/cost envelope.

Changing evaluation inputs confounds the comparison and hides accuracy, throughput and completion-cost differences.

Catalog breadth does not demonstrate whether either candidate meets the application acceptance thresholds.

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

Question 20

Explanation

Correct Answer

Incorrect Answers

Question 21

Explanation

Correct Answer

Incorrect Answers

Question 22

Explanation

Correct Answer

Incorrect Answers

Question 23

Explanation

Correct Answer

Incorrect Answers

Question 24

Explanation

Correct Answer

Incorrect Answers

Question 25

Explanation

Correct Answer

Incorrect Answers

Question 26

Explanation

Correct Answer

Incorrect Answers

Question 27

Explanation

Correct Answer

Incorrect Answers

Question 28

Explanation

Correct Answer

Incorrect Answers

Question 29

Explanation

Correct Answer

Incorrect Answers

Question 30

Explanation

Correct Answer

Incorrect Answers

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