Task 1.6: Implement Prompt Engineering Strategies and Governance for FM Interactions. AIP-C01 Practice Questions (Page 4)

Questions 16–20

1. 16expert · hard

   A company has a governance policy that requires all prompt templates to be versioned, and that any change to a template must be associated with a change request ticket. The team uses an S3 bucket to store templates. They want to enforce this policy automatically. Which solution is the MOST effective?

   • AEnable S3 versioning on the bucket. Require developers to include the change request ticket number in the object metadata when uploading a new version.
   • BUse AWS CodeCommit to store the templates. Require all changes to be made via pull requests that must reference a change request ticket. Use AWS CodePipeline to deploy approved changes to the S3 bucket.
   • CUse a Lambda function triggered by S3 events to check if the new object has the required metadata tag. If the tag is missing, the Lambda function deletes the object and sends an alert to the governance team.
   • DStore the templates in a DynamoDB table with a version attribute. Require developers to include the change request ticket number in the version attribute.
   Reveal Solution Discussion Select an answer first
2. 17expert · hard

   A customer service chatbot is built on Amazon Bedrock. It uses a Lambda function that maintains a conversation history in a DynamoDB table. The team has noticed that for long conversations, the model's responses become less relevant because the full history exceeds the model's context window. They need to maintain context while keeping the prompt within the token limit. Which strategy is the MOST effective?

   • AStore the conversation history in an Amazon S3 bucket and use a Lambda function to retrieve the full history and send it to the model.
   • BTruncate the conversation history to the last 10 messages before sending it to the model.
   • CIncrease the model's context window by using a more expensive foundation model that supports longer inputs.
   • DUse Amazon Comprehend to extract key entities and intents from the entire conversation history and store them in DynamoDB. Send only the extracted summary along with the last few messages to the model.
   Reveal Solution Discussion Select an answer first
4. 18application · medium

   A customer support platform is building a multi-step assistant that helps agents resolve complex issues. The workflow is: 1) Classify the issue type, 2) Retrieve relevant knowledge base articles, 3) Generate a draft response, and 4) Check the response for policy compliance. The team wants to implement this as a series of prompts that can be tested and modified independently. Which approach should they use?

   • ACreate a single large prompt that instructs the model to perform all four steps in a single response, using chain-of-thought reasoning.
   • BUse AWS Step Functions to orchestrate a sequential chain of four separate Lambda functions, each calling Bedrock with a specific prompt for one step. Pass the output of each step as input to the next.
   • CUse Amazon Comprehend to perform the classification step, and then use a single Bedrock prompt to handle the remaining three steps (retrieval, generation, and compliance check).
   • DCreate a single Lambda function that calls Bedrock four times in sequence, embedding all four prompts in the function code.
   Reveal Solution Discussion Select an answer first
5. 19foundation · easy

   A prompt engineer wants to refine a prompt by analyzing the model's outputs and adjusting the prompt based on failures. Which process is this an example of?

   • APrompt versioning
   • BModel fine-tuning
   • CPrompt injection
   • DIterative prompt enhancement
   Reveal Solution Discussion Select an answer first
6. 20application · medium

   A marketing agency uses Amazon Bedrock to generate ad copy. They have a set of 50 prompt templates that are used in production. Before rolling out a change to a template, they want to verify that the new version does not regress on key quality metrics (e.g., tone, length, and inclusion of a call-to-action). They have a golden set of 100 input scenarios with expected outputs. Which approach should they use to automate this regression testing?

   • ADeploy the new prompt to production and monitor CloudWatch Logs for an increase in customer complaints or a decrease in click-through rates over the next two weeks.
   • BCreate a Step Functions state machine that runs the new prompt against the golden set, uses Amazon Comprehend to analyze the sentiment and key phrases of each output, and compares these metrics to the expected values using a configurable threshold. The state machine then publishes the pass/fail results to CloudWatch.
   • CUse Amazon SageMaker Ground Truth to have human annotators manually review the outputs of the new prompt against the golden set, and then manually update a DynamoDB table with the results.
   • DCreate a Lambda function that runs the new prompt against the golden set, compares the outputs to the expected results using exact string matching, and reports any mismatches to CloudWatch Logs.
   Reveal Solution Discussion Select an answer first