AIP-C01 sample questions with answers

Choose 2.

1. A.

   Measure each model's perplexity on the brand's past catalogue text and select the model with the lowest perplexity, since it best predicts the brand's writing.

2. B.

   Score each model's outputs with an exact-match accuracy metric against one reference description per brief, because exact match measures brand voice objectively.

3. C.

   Compare the three models by their scores on a public benchmark for reasoning, and assume that the best reasoner also writes the most persuasive brand copy.

4. D.

   Run a model evaluation job that uses an LLM as a judge with a custom metric built from the written style rubric, scoring all three models on the 600 briefs.

5. E.

   Run a human-based evaluation job with the five copywriters as a private work team on a sample of briefs, rating each model's outputs against the same rubric.

1. A.

   Keep the prompts unchanged and raise the new model's temperature, because a warmer temperature compensates for differences between the two models' prompt styles.

2. B.

   Simple prompt optimization in Prompt Management, run on each template once, because it evaluates each of its rewrites against the ground-truth answers and the Lambda scorer for up to five models.

3. C.

   Amazon Bedrock Advanced Prompt Optimization, supplying the templates, sample inputs, ground truth and the Lambda scorer, with the current model as the baseline and the new model as a target.

4. D.

   A fine-tuning job on the new model with the 480 samples, because fine-tuning always replaces the need to adapt prompt templates when migrating between models.

1. A.

   Fixed-size chunking with a much smaller maximum token size and zero overlap, so that each chunk is short enough to contain only one topic from the conversation.

2. B.

   Semantic chunking, which places boundaries where the meaning of consecutive sentences diverges, tuned with the buffer size and breakpoint percentile threshold.

3. C.

   No chunking, with each transcript stored as a single chunk, so that no topic is ever split across two chunks during ingestion into the knowledge base.

4. D.

   Hierarchical chunking with a large parent size and a small child size, so that the parent chunks capture each topic regardless of where the topic shifts occur.

1. A.

   Switch the embedding request to 256 output dimensions, because a smaller output vector allows the model to accept a proportionally longer input text for each chunk.

2. B.

   Split oversized tables into smaller chunks below the model's input limit of 8,192 tokens (and 50,000 characters), repeating the table header in each piece so that the rows stay interpretable.

3. C.

   Base64-encode each large table before sending it to the embedding model, because the encoding compresses the text enough to stay below the token limit.

4. D.

   Request a service quota increase for the maximum input tokens of the embedding model, so that very large tables can be embedded as a single chunk without splitting any of the rows across chunks.

Choose 2.

1. A.

   Give the gateway's service role the AWSLambda_FullAccess managed policy, so that new tools can be added later without any further changes to the role.

2. B.

   Give the gateway's service role an IAM policy that allows lambda:InvokeFunction only on the ARN of the retrieval function.

3. C.

   Configure the gateway's inbound authorization as JWT, using the Cognito user pool's discovery URL and the allowed client IDs, so that each MCP request must carry a valid token.

4. D.

   Set inbound authorization to No Authorization and rely on the Lambda function to check a shared API key that the partners embed in each tool call's arguments.

5. E.

   Make the Lambda function's URL public and list it in the MCP server description, so that partners can call either the function URL or the gateway.

1. A.

   An Amazon SQS queue for each step, with Lambda consumers that pass partial results to the next queue, and a DynamoDB table that tracks which steps each query has completed.

2. B.

   A single AWS Lambda function that calls every step in sequence, with retries written in the function code and progress logged to CloudWatch Logs for later troubleshooting.

3. C.

   One large prompt that asks the generation model to classify, decompose, retrieve and answer in a single call, with the knowledge base contents pasted into the prompt.

4. D.

   An AWS Step Functions workflow: a classification task with Bedrock, a Choice state that routes to rewrite or decomposition tasks, a Map state that runs one retrieval per sub-query, a merge task, and a final generation task.

1. A.

   Give the model the knowledge base ID in the system prompt and instruct it to call the Retrieve API itself, because models on Amazon Bedrock can call AWS APIs directly with the permissions of the caller's role.

2. B.

   Define a searchproductdocs tool in toolConfig with a JSON input schema for query and product; when the model returns a toolUse block, run Retrieve with those inputs and send the results back as a toolResult.

3. C.

   Call Retrieve before every model invocation and always paste the top results into the system prompt, so that the model can ignore them when a question does not need product facts.

4. D.

   Fine-tune the model on the knowledge base content so that it can answer product questions without retrieval and never needs to search the knowledge base.

1. A.

   Put the photo's base64 text inside the system field, and send the question as the only item in the messages array with the role set to assistant.

2. B.

   Append the photo's file path to the question text, such as 'see /tmp/claim.jpg', and let Amazon Bedrock read the file from the machine that is running the application.

3. C.

   Send one user message whose content array holds an image block with format jpeg and the image bytes as the source, followed by a text block containing the question.

4. D.

   Send the photo in a first request and the question in a second request, relying on the model to remember the photo from the earlier call when it answers.

Choose 2.

1. A.

   Ask the foundation model to guess the intended drug name from context whenever a transcribed word sounds like a medicine but is spelled as a common word.

2. B.

   Increase the audio sample rate of the recordings, because Transcribe transcribes words with unusual spellings correctly only when the audio is at least 48 kHz.

3. C.

   Enable automatic language identification for the jobs, with English and Spanish as language options and the matching custom vocabulary supplied for each language.

4. D.

   Enable speaker partitioning so that Transcribe separates the pharmacist from the customer, because brand names are mis-transcribed when two speakers overlap.

5. E.

   Create a custom vocabulary of the drug brand names for each of the two languages, with pronunciation hints where brand names are spelled unusually.

Choose 2.

1. A.

   Use a Lambda function to strip quoted reply chains and signatures, and normalise all detected dates to ISO 8601 before the email reaches the prompt.

2. B.

   Raise the temperature to 0.9 so the model explores more interpretations of ambiguous dates.

3. C.

   Append the entire raw mailbox thread history for the claimant to each request so the model has maximum context.

4. D.

   Increase the model's max tokens setting and add a request for step-by-step reasoning so that the model has more room to work through the noisy thread before answering.

5. E.

   Use Amazon Comprehend entity recognition to extract person, organisation and date entities from the thread and pass them to the model as a structured, labelled context block alongside the cleaned email.