Which Amazon Comprehend filtering design meets privacy and latency requirements?

AIP-C01 By islever Editorial Last updated

Implement input and output safety controls. Implement data security and privacy controls.

Short answer

D Run Amazon Comprehend toxicity detection, prompt safety classification, and PII redaction in parallel before the Bedrock FM so all filters finish first.

Question

A company is developing a customer communication platform that uses an AI assistant powered by an Amazon Bedrock foundation model (FM). The AI assistant summarizes customer messages and generates initial response drafts.
The company wants to use Amazon Comprehend to implement layered content filtering. The layered content filtering must prevent sharing of offensive content, protect customer privacy, and detect potential inappropriate advice solicitation. Inappropriate advice solicitation includes requests for unethical practices, harmful activities, or manipulative behaviors. The solution must maintain acceptable overall response times, so all pre-processing filters must finish before the content reaches the FM.
Which solution will meet these requirements?
  1. Use parallel processing with asynchronous API calls. Use toxicity detection for offensive content. Use prompt safety classification for inappropriate advice solicitation. Use personally identifiable information (PII) detection without redaction.
  2. Use custom classification to build an FM that detects offensive content and inappropriate advice solicitation. Apply personally identifiable information (PII) detection as a secondary filter only when messages pass the custom classifier.
  3. Deploy a multi-stage process. Configure the process to use prompt safety classification first, then toxicity detection on safe prompts only, and finally personally identifiable information (PII) detection in streaming mode. Route flagged messages through Amazon EventBridge for human review.
  4. Use toxicity detection with thresholds configured to 0.5 for all categories. Use parallel processing for both prompt safety classification and personally identifiable information (PII) detection with entity redaction. Apply Amazon CloudWatch alarms to filter metrics. Correct Answer

Key Takeaways

The question tests Amazon Comprehend's layered moderation capabilities and latency trade-off; the trap is choosing parallel latency in option A while ignoring that PII detection without redaction does not protect customer privacy.

Amazon Comprehend layered content filtering for a Bedrock-powered assistant requires toxicity detection, prompt safety classification, and PII redaction to run before inference. This page establishes why option D satisfies all three checks with parallel pre-processing while option A leaves PII unredacted.

Common mistake: The most common wrong answer is A: it runs the Comprehend filters in parallel, but PII detection without redaction still lets sensitive customer data reach the Bedrock FM, so the privacy requirement is not met.

Expert Analysis

Why the Answer Is Correct

Option D is the only choice that combines the three Amazon Comprehend capabilities the scenario requires: toxicity detection for offensive content, prompt safety classification for inappropriate advice solicitation, and PII detection with entity redaction for customer privacy. Running the independent checks in parallel is the right latency pattern because the requirement says every pre-processing filter must finish before the draft request reaches the Bedrock FM. Redaction is essential: merely detecting PII does not stop sensitive data from being included in the FM prompt. The 0.5 toxicity threshold is a tuning detail, not a disqualifier in a question that does not specify per-category thresholds, and CloudWatch alarms add useful visibility into filter behavior.

Why the Other Options Are Wrong

Option A has the right parallel latency idea and covers toxicity and prompt safety, but its PII detection is without redaction, so customer data can still reach the FM and the privacy requirement fails. Option B misuses Comprehend custom classification as if it could build an FM and gates PII detection behind the custom classifier, so not every message receives the privacy filter and the design is not a clean layered pre-processing pipeline. Option C is the serial version: prompt safety, then toxicity, then PII in streaming mode adds cumulative latency and EventBridge human review without clearly redacting PII before inference. Option D's fixed 0.5 threshold is less flexible than a tuned threshold, but it still supplies all three filters plus redaction and parallel execution.

Community Comment Notes

The comments show the main split in the thread. Naaser supports A for "Lowest Latency" and complete coverage, but OlaFemi points out the fatal gap: "a redaction is needed" for PII, which A omits. Chibuzo1 backs D because it pairs toxicity, prompt safety, and PII with redaction and parallel execution, while mosun and jyrajan69 object to D's fixed 0.5 threshold; that is a tuning concern, not a failure of the privacy requirement. dbcert87 argues that "layered filtering which rules our parellel processing" favors C, but the latency requirement and the need to finish all filters before the Bedrock FM make the parallel design more defensible. The comments are useful for spotting the redaction trap, but the exam's decisive criterion is complete pre-processing with PII protection.

Official Reference

These links open the official vendor documentation in a new tab.

Exam Strategy

Read the privacy requirement as a binary control: detecting PII is not the same as preventing it from reaching the Bedrock FM. When the question says all filters must finish before inference, prefer parallel independent checks that include redaction, and treat fixed thresholds as secondary tuning details.

Frequently Asked Questions

Why is PII detection without redaction insufficient in option A?

Detecting PII only flags it; unless the message is blocked or entities are redacted, sensitive customer data can still reach the Bedrock FM, failing the privacy requirement.

Does layered content filtering require sequential Comprehend calls?

No. Toxicity detection, prompt safety classification, and PII redaction can run in parallel as long as every filter finishes before the Bedrock FM receives the draft request.

More AIP-C01 FAQ →

Discussion 7 contributions

Compiled from public community threads and reader submissions, lightly edited for clarity. Handles and vote tallies are shown as recorded, and comments reflect the views of individual contributors.

From public community threads 7

  1. Naaser 👍 3 Answered A
    Lowest Latency: Using parallel processing with asynchronous API calls ensures that all three Amazon Comprehend filters (Toxicity, Prompt Safety, and PII) run simultaneously. This prevents cumulative "sequential" latency, meeting the requirement for acceptable response times.Complete Coverage:Toxicity Detection: Directly addresses the requirement to filter offensive content.Prompt Safety Classification: Specifically designed to detect inappropriate advice solicitation (unethical, harmful, or manipulative requests).PII Detection: Protects customer privacy by identifying sensitive data before it reaches the foundation model.
  2. Chibuzo1 👍 2 Answered D
    The answer is D because it correctly pairs all three required Comprehend capabilities with the right configurations — toxicity detection with a calibrated threshold, prompt safety classification, and PII detection with entity redaction — running in parallel to minimize pre-processing latency, ensuring all filters complete before the FM receives any content, and maintaining operational visibility through CloudWatch alarms.
  3. OlaFemi 👍 2 Answered D
    A is not correct, it says Use personally identifiable information (PII) detection without redaction. and a redaction is needed
  4. dbcert87 👍 1 Answered C
    1/ Question asked layered filtering which rules our parellel processing 2/ Hard coding .5 for all categories is not sage and AWS best practice 3/Sequential filtering is best if complete before inference.
  5. jyrajan69 👍 1 Answered A
    Option D is not the right answer as it required hard coding the 0.5 as toxicity thresh hold, better option is A
Show 2 more comments
  1. mosun 👍 1 Answered A
    Answer is A, because option D has fixed value - 0.5 for all categories
  2. de1612d 👍 1 Answered D
    Answer is D

Reader comments & corrections

No reader corrections yet — spotted an error or have a note? Share it below.

Log in to comment, report an error, or add a note about this question.

How Can AgentCore Query Cross-Account DynamoDB While Preserving User Identity?

S-grade

Which API Gateway setup streams Bedrock responses with least operational overhead?

S-grade

Resilient Retry and Streaming Strategies for Amazon Bedrock

S-grade

Bedrock Prompt Management & Guardrails Detection

S-grade

Practice All AIP-C01 Questions

Access 110 questions with complete answers and detailed explanations.

View Full AIP-C01 Practice Test →

← Back to AIP-C01 Study Guide