Amazon AWS Certified Generative AI Developer - Professional AIP-C01 Practice Test 22

Question 210

A deployment uses an end-of-life model ID. Which deployment or capacity decision is most appropriate?

Continue invoking a retired model indefinitely and assume Bedrock silently maps it to another version.
Replace the ID in configuration without testing output contract and user-case correctness.
Plan supported model-version upgrade and decommission according to Bedrock lifecycle.
Archive the application's prompt code and treat that as a substitute for a supported model deployment.

Explanation

Correct Answer

An EOL model requires planned migration to a published supported version with regression testing, not a silent identifier substitution.

Incorrect Answers

Lifecycle transitions can end availability rather than transparently preserving API behavior.

A new version can change semantics or unsupported settings.

Source archiving does not keep an EOL inference resource operational.

Question 211

A validated small FM handles most simple FAQ classes, while harder questions exceed its capability. Select TWO complementary model-routing behaviors: the low-cost path for supported cases and an escalation path for the others.

Send all tasks to the smallest model because it has the lowest input-token price.
Route bounded simple cases to validated smaller models.
Escalate unsupported/complex cases to tested capable models.
Escalate only when output JSON fails to parse, ignoring semantically incorrect classifications.
Use the largest model for every trivial class without measuring whether the smaller model is sufficient.

Explanation

Correct Answer

Route bounded FAQ classification to a tested smaller model when its actual confusion matrix meets required quality.

Escalate harder or low-confidence cases to an evaluated capable model instead of forcing every input onto the cheap path.

Incorrect Answers

A cheaper failure can cost more in remediation and user harm.

Syntactic success does not prove task quality.

Unnecessary escalation wastes cost with no established quality improvement.

Question 212

An organization needs safe canary rollout of model change. What should be verified before choosing this model-serving path?

Route every production user to the new model before checking its real error rate.
Track only infrastructure availability and ignore answer correctness after promotion.
Route small traffic to new version and rollback on quality/latency regression.
Rollback to the same new model ID while calling it the prior stable version.

Explanation

Correct Answer

Canary rollout sends a measured fraction of requests to a qualified new model and reverts on predefined safety, quality or latency regressions.

Incorrect Answers

A full cutover removes the canary containment boundary.

A healthy endpoint may produce unacceptable business content.

A reversal must target an actually validated previous artifact/version.

Question 213

An endpoint handles confidential prompts but logs raw full content. Which deployment or capacity decision is most appropriate?

Enable unredacted prompt logging into a broadly shared CloudWatch group to simplify debugging.
Redact only text visible in the web UI but still store full raw prompts in backend logs.
Configure privacy-aware invocation logging before production deployment.
Disable all service metadata and audit correlations while retaining full document bodies.

Explanation

Correct Answer

Invocation logging should capture useful diagnostics without leaking confidential prompt bodies; enforce destination IAM, encryption and retention.

Incorrect Answers

Sensitive prompt contents become accessible outside their permitted purpose.

Server-side invocation logs remain a separate disclosure path.

Removing metadata does not protect the sensitive payload or maintain auditability.

Question 214

A company thinks Bedrock service and Agents Classic share lifecycle. Which operational correction addresses the deployment failure?

Stop all existing foundation-model inference merely because Agents Classic restricts new customers.
Distinguish continued core Bedrock inference from Classic agent creation restrictions.
Assume new accounts can create Classic agents with unchanged availability indefinitely.
Rename Classic configuration as AgentCore without checking orchestration and capability migration.

Explanation

Correct Answer

Agents Classic creation restrictions do not imply the Bedrock core InvokeModel/Converse service itself is discontinued.

Incorrect Answers

Classic's onboarding lifecycle is not the lifecycle of the whole Bedrock runtime.

The documented customer onboarding cutoff applies to the legacy agent product.

Different runtimes/services do not guarantee feature parity.

Question 215

A custom hosted model has a new LoRA adapter. What should be verified before choosing this model-serving path?

Attach the LoRA adapter to any same-size model regardless of base architecture.
Replace the base model after validation without retesting the adapter.
Store only the adapter file while omitting base version and serving configuration.
Verify compatible adapter artifact and base model before serving production traffic.

Explanation

Correct Answer

Deploy a compatible adapter with its exact base model and preserve both artifact versions and tested rollback state.

Incorrect Answers

Adapter weight compatibility is tied to its supported base model and training setup.

Changing the base invalidates the previously tested pairing.

A reproducible deployment requires the complete artifact/version relationship.

Question 216

An ERP app requires LLM assistance but cannot expose credentials to clients. What should the engineer implement at the enterprise service boundary?

Include an IAM access key in shipped mobile application code so clients can call Bedrock directly.
Put model invocation behind authenticated service/API integration layer.
Grant every browser user the same administrator role for model access.
Allow anonymous direct requests and filter unsafe outputs only after generation.

Explanation

Correct Answer

Place Bedrock inference behind a trusted, authenticated backend that holds IAM credentials and enforces the application's resource policies.

Incorrect Answers

Embedded static secrets can be extracted and reused outside authorization controls.

Shared broad privileges cannot enforce app-specific user entitlements.

Output filtering is not authentication of the request or its resource scope.

Question 217

A legacy API only supports synchronous calls while GenAI jobs take minutes. Which integration architecture change best addresses this failure?

Increase the legacy client timeout beyond its documented integration maximum and hope each call completes.
Bridge through queued asynchronous workflow with status callbacks.
Return HTTP 200 success as soon as the job is queued without a pending/status field.
Poll the GenAI model for a partial answer every millisecond while keeping one open legacy HTTP call.

Explanation

Correct Answer

Use SQS/Step Functions or a supported asynchronous job with correlation/status callbacks, not a synchronous legacy request waiting minutes.

Incorrect Answers

Service timeout constraints and disconnects remain; long work needs an asynchronous contract.

Accepted work is not proof that the GenAI processing succeeded.

Aggressive polling increases load without a durable completion boundary.

Question 218

A CRM update event must trigger a post-call summary. Which change satisfies the specified security or reliability requirement?

Use event-driven EventBridge or queue consumers rather than manual polling.
Require a support employee to start a nightly manual summary export for each updated CRM case.
Embed a CRM administrator password in the generation prompt so the model can watch for updates.
Treat every retrieved CRM search result as proof that a new change event occurred.

Explanation

Correct Answer

Publish a CRM event on update and let a controlled EventBridge/SQS subscriber trigger summarization with authenticated source identifiers.

Incorrect Answers

Manual polling misses near-real-time update triggers and increases operational error.

A model prompt is neither an event subscription nor safe credential storage.

Search data does not establish whether a specific update was committed.

Question 219

A downstream billing API is temporarily unavailable. What should the engineer implement at the enterprise service boundary?

Buffer events and retry with idempotent effect; isolate poison messages in DLQ.
Discard billing events after the first 503 response without an acknowledged retry path.
Retry with a new charge ID on every attempt and no deduplication at billing.
Consume the same malformed message indefinitely, blocking later valid events.

Explanation

Correct Answer

Buffer billing events with retry/backoff, idempotent writes and a DLQ for poison messages, preserving event IDs and recovery evidence.

Incorrect Answers

Transient service outages would permanently lose required work.

Repeated deliveries can create duplicate financial effects.

A poison message requires bounded retries and isolation, not a stuck queue.

Question 220

Multiple business systems need the same GenAI endpoint. Which integration architecture change best addresses this failure?

Give every downstream app a different unversioned Bedrock wrapper with incompatible response formats.
Put model API credentials into each client package to avoid a centralized service.
Publish a central API contract and shared authorization/observability gateway.
Route all callers through an unauthenticated endpoint because the underlying model is read-only.

Explanation

Correct Answer

Publish a versioned shared API contract and caller-aware gateway so multiple apps reuse a consistent authenticated model service.

Incorrect Answers

Divergent contracts cause integration breakage and inconsistent policy enforcement.

Distributed client secrets expand access exposure and cannot enforce a common boundary.

Prompts and model usage still need identity, quotas and data-access authorization.

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

Question 210

Explanation

Correct Answer

Incorrect Answers

Question 211

Explanation

Correct Answer

Incorrect Answers

Question 212

Explanation

Correct Answer

Incorrect Answers

Question 213

Explanation

Correct Answer

Incorrect Answers

Question 214

Explanation

Correct Answer

Incorrect Answers

Question 215

Explanation

Correct Answer

Incorrect Answers

Question 216

Explanation

Correct Answer

Incorrect Answers

Question 217

Explanation

Correct Answer

Incorrect Answers

Question 218

Explanation

Correct Answer

Incorrect Answers

Question 219

Explanation

Correct Answer

Incorrect Answers

Question 220

Explanation

Correct Answer

Incorrect Answers

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