8.3. Practice Questions (Mixed Domain)

Question 1: A financial services company must process 10,000 customer documents nightly to generate compliance summaries. Documents contain PII. Processing must complete within 4 hours. Data must not leave us-east-1. What is the most cost-effective architecture?

Key constraints: offline/nightly (→ async/batch), PII (→ Comprehend + Guardrails), data residency (→ VPC endpoint, disable cross-region inference), 4-hour window (→ concurrent Lambda + SQS, or Bedrock Batch Inference)

Question 2: A RAG chatbot returns correct answers 90% of the time, but for questions about regulatory changes published last month, it consistently responds "I don't have information about that." The knowledge base is configured with Bedrock Knowledge Bases. What is the most likely cause?

Key signal: recent information missing → sync delay. Answer: Knowledge Base sync job not triggered after new documents were added → configure event-driven sync via S3 event notification → Lambda → StartIngestionJob.

Question 3: A GenAI customer service application has a P50 latency of 1.8s and a P99 latency of 31s. The operations team must reduce P99 to under 10s. X-Ray shows generation subsegment = 29s at P99. CloudWatch shows ThrottledRequests spikes at 10am and 2pm. What is the most targeted fix?

Key signals: throttling at peak times → provisioned throughput. Generation is the problem, not retrieval. Answer: purchase provisioned throughput model units sized for peak load.

Question 4: You need to build an application where: (1) the FM must only discuss topics in company documentation, (2) all responses must cite source documents, (3) PII must never appear in responses. What combination of features addresses all three requirements?

Requirement 1: Bedrock Guardrails topic denial + grounding check. Requirement 2: Bedrock Knowledge Bases retrieve_and_generate (returns citations automatically). Requirement 3: Bedrock Guardrails PII redaction on output.

Question 5: Your Bedrock Agent successfully retrieves customer order data in step 1 of a reasoning loop, but then makes 8 additional tool calls trying to retrieve the same data before timing out. What is the most likely root cause?

Key signal: agent retrieved data successfully but didn't recognize it → tool return format not matching agent's expectation. Answer: Lambda action group returning data in incorrect format → agent doesn't recognize it as a valid observation → retries. Fix: correct the response format in the Lambda function to match the action group OpenAPI schema.