Amazon AWS Certified Generative AI Developer - Professional AIP-C01 Practice Test 20

Question 190

A long-lived agent retains unrelated user memory indefinitely. Which agent/tool architecture control directly resolves the situation?

Apply actor scoping, retention and deletion policies to stored memories.
Save every tool result and user message forever under a global memory namespace.
Treat absence of visible chat history as proof the persistent memory record was deleted.
Give the model unilateral authority to decide which confidential user memories to retain.

Explanation

Correct Answer

Use actor-scoped memories with purpose-limited retention, deletion and expiry so unrelated historical content does not persist forever.

Incorrect Answers

This violates minimization and can mix unrelated actors' data.

User interface state does not prove backend data deletion.

Retention policy must be enforced in governed storage, not natural-language discretion.

Question 191

A tool returns a URL to a private document for an anonymous caller. Which change preserves actual tool authorization and trustworthy state?

Convert a private S3 locator into a public presigned link for any anonymous chat visitor.
Enforce signed-in document authorization before passing link or content to model.
Return restricted document snippets to the model and only hide the URL in the UI.
Trust the tool's 'document found' status as evidence the requester may read it.

Explanation

Correct Answer

Check the requester against the private document's actual ACL before returning its content or a usable link to a tool/model.

Incorrect Answers

A broadly issued signed URL can expose data without verifying reader rights.

The disclosure has already happened through the model context.

Existence of a document is not an entitlement.

Question 192

An agent asks a tool to perform an unsupported function. What should the developer implement before allowing the agent to act?

Return a typed capability error instead of allowing invented fulfillment.
Synthesize a successful tool response JSON without ever invoking a supported operation.
Try a similar tool with broader permissions without checking whether it implements the user request.
Hide the error from the user and mark the agent workflow complete.

Explanation

Correct Answer

An unsupported function should return a typed capability/error state and route to approved fallback or human help, not invented completion.

Incorrect Answers

A fabricated tool result misrepresents the actual backend state.

A different tool may perform an unauthorized or unrelated action.

Unsupported capability does not constitute task success.

Question 193

A workflow must choose small and large models by complexity. Which agent/tool architecture control directly resolves the situation?

Route every task to the cheapest model without checking complexity or failure cost.
Dispatch bounded subtasks using recorded routing criteria and measure escalation accuracy.
Route every question to the largest model irrespective of measured adequacy.
Allow model-generated labels alone to authorize high-risk tool escalation without policy checks.

Explanation

Correct Answer

Select capability-bounded subtasks using measured routing criteria; log escalation/quality so simple cases use smaller models only when sufficient.

Incorrect Answers

Price cannot replace task-specific quality thresholds.

This wastes cost and can fail capacity targets without quality benefit.

Task routing confidence is not business permission.

Question 194

An agent completes a purchase but omits invoice confirmation. Which change preserves actual tool authorization and trustworthy state?

Use the model's generated invoice number as evidence of a completed purchase.
Require receipt validation from purchasing system prior to success response.
Interpret an HTTP 202 accepted response as proof the downstream payment was committed.
Close the transaction on the frontend while backend invoice creation remains pending.

Explanation

Correct Answer

Wait for a receipt or authoritative purchase status and validate invoice/account details before telling the user the order succeeded.

Incorrect Answers

A model can invent plausible identifiers independent of the purchasing system.

Queue acceptance is different from final business success.

User-visible completion must match authoritative commit state.

Question 195

A team adopts third-party MCP servers with broad scopes. What should the developer implement before allowing the agent to act?

Connect any advertised public MCP endpoint with an administrator token and skip security review.
Forward customer documents to the server to test it before verifying data-sharing approval.
Accept arbitrary tool schema updates from the server with no contract tests or scope review.
Threat-model server provenance, OAuth scopes and outbound data exposure before enabling.

Explanation

Correct Answer

Assess MCP server provenance, requested OAuth scopes, outbound destinations and sensitive-data flow before enabling third-party tools.

Incorrect Answers

Server discovery is not proof of trustworthiness or least privilege.

Testing itself can leak sensitive information.

A changed schema may expand permitted behavior unexpectedly.

Question 196

A low-latency application invokes Bedrock foundation models as-needed. Which operational correction addresses the deployment failure?

Purchase provisioned throughput before measuring load because the app needs low latency.
Assume every Bedrock model allows on-demand requests in every AWS Region.
Compare supported on-demand inference against provisioned throughput by measured load.
Evaluate only request count and ignore input/output token volume.

Explanation

Correct Answer

Select on-demand only for a model/Region/deployment supporting it, and compare measured arrivals, utilization and cost with offered provisioned throughput.

Incorrect Answers

Committed capacity can be wasteful or insufficient without traffic and model-specific throughput evidence.

Inference deployment availability is model-version/Region specific.

Token mix drives throughput, queueing and price beyond raw HTTP calls.

Question 197

A customer needs dedicated predictable model capacity. What should be verified before choosing this model-serving path?

Assume all foundation-model versions accept the same provisioned throughput allocation.
Reserve the largest available capacity without forecasting actual token demand.
Use a regular S3 lifecycle rule to guarantee dedicated inference capacity.
Evaluate available Bedrock provisioned throughput capacity with utilization economics.

Explanation

Correct Answer

Provisioned capacity can provide predictable throughput for supported models, but only a measured steady workload justifies its commitment and expense.

Incorrect Answers

Availability and capacity purchase depend on the exact model and Region.

An oversized commitment wastes budget and still may not address other bottlenecks.

Object retention policy does not allocate or reserve model serving throughput.

Question 198

A company must summarize one million independent documents overnight. Users do not wait for per-document responses, and each summary needs no tool calls. Which Bedrock serving approach best fits?

Use a supported asynchronous Bedrock batch inference job with independent JSONL records and S3 input/output locations.
Open a persistent interactive Converse session for all million records and wait for one live answer.
Require a toolUse/toolResult exchange within each Bedrock batch record to complete the summary.
Provision a synchronous public API for every document regardless of the deadline and measured economics.

Explanation

Correct Answer

Batch jobs handle independent prompts asynchronously and write results to S3 for later retrieval, subject to supported model and Region prerequisites.

Incorrect Answers

A single interactive session is unsuitable for large offline independent records and couples availability to a long-running request.

Batch inference does not support interactive tool calling within a record; this workload intentionally does not need it.

Synchronous serving can be valid for other tasks but is not required by a next-morning asynchronous deadline.

Question 199

A live chat client waits for a reply after each user turn, but the application has mistakenly submitted every turn as a Bedrock batch inference job that produces asynchronous S3 output. What invocation-path correction meets conversational latency?

Route each chat turn to a new S3 batch job and block the browser until the output appears.
Use a supported synchronous or streaming Bedrock runtime inference API for live chat turns rather than waiting for an asynchronous S3 batch job.
Poll the batch output bucket for each streamed response token.
Convert each chat request into a batch record and assume it supports conversational tool calls.

Explanation

Correct Answer

Batch inference is asynchronous and designed for queued work with S3 input/output, not an interactive request/response dialogue. Live chat requires a supported low-latency invoke path with appropriate model/Region eligibility.

Incorrect Answers

Asynchronous batch scheduling is incompatible with interactive chat turn deadlines.

Batch output is delivered after job processing, not as an interactive token stream.

Batch does not implement live multi-turn tool execution.

Question 200

A proposed Bedrock batch inference JSONL file expects every record to call a live inventory tool and then continue the model conversation with the returned tool result. What should the developer do?

Enable a larger S3 batch output bucket so toolUse/toolResult exchanges become available.
Use batch inference records as a direct substitute for an interactive multi-turn tool dialogue.
Put the inventory access token in the prompt and assume the batch worker executes the tool itself.
Use a supported interactive tool-use API and application/tool orchestrator, because Bedrock batch inference does not support that back-and-forth tool calling.

Explanation

Correct Answer

Client-side Converse tool use, for example, supports a model tool request, an application-executed tool, and a toolResult message. Batch inference cannot perform that loop.

Incorrect Answers

Output storage capacity does not add a feature that the batch record protocol does not support.

Batch records are independent asynchronous inference operations, not ongoing interactive exchanges that wait for client tool execution.

A token in a prompt is neither safe nor an instruction for AWS batch inference to invoke an external business API.

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

Question 190

Explanation

Correct Answer

Incorrect Answers

Question 191

Explanation

Correct Answer

Incorrect Answers

Question 192

Explanation

Correct Answer

Incorrect Answers

Question 193

Explanation

Correct Answer

Incorrect Answers

Question 194

Explanation

Correct Answer

Incorrect Answers

Question 195

Explanation

Correct Answer

Incorrect Answers

Question 196

Explanation

Correct Answer

Incorrect Answers

Question 197

Explanation

Correct Answer

Incorrect Answers

Question 198

Explanation

Correct Answer

Incorrect Answers

Question 199

Explanation

Correct Answer

Incorrect Answers

Question 200

Explanation

Correct Answer

Incorrect Answers

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