Explore / AWS Certified Generative AI Developer - Professional (AIP-C01) Exam Prep / Practice questions
AWS Certified Generative AI Developer - Professional (AIP-C01) practice questions
10 questions from the bank behind this course's timed mock. Free to read and to answer, no account needed, and each has its answer and an explanation. They belong to AWS Certified Generative AI Developer - Professional (AIP-C01) Exam Prep.
Answer them one at a time →Read lesson 1 free
Question 1
A developer taking AIP-C01 has 4 minutes left and 8 questions they haven't answered. They remember that 10 of the 75 questions are unscored and decide that some of these 8 are probably trial items. Under the exam guide's rules, what is the BEST action?
- Leave all 8 blank and request a section-level rescore, because compensatory scoring ignores unanswered questions
- Leave the hardest ones blank, because blank answers on unscored trial questions count as neutral
- Answer only the questions that look like standard multiple choice, because unscored trial questions are usually the multiple-response ones
- Pick an answer for all 8 questions, guessing where needed, because a blank is scored as incorrect and guessing isn't penalised
Show the answer
Answer: D. Pick an answer for all 8 questions, guessing where needed, because a blank is scored as incorrect and guessing isn't penalised
The unscored questions aren't identified, so any of the 8 could count, and a blank is always marked wrong while a guess costs nothing. The 'neutral blank' option is wrong because you can't tell which questions are unscored, and on a scored question a blank is simply an incorrect answer.
Question 2
A startup's internal document summarisation tool calls Amazon Bedrock from us-east-1 using an on-demand foundation model. During busy periods the tool is throttled. The data is not regulated and has no residency rules, and leadership wants the lowest cost and the most available throughput. They want to make as few changes as possible. What should the developer do?
- Replace the model ID with the global. inference profile ID for the same model
- Replace the model ID with the us. geographic inference profile ID for the same model
- Deploy the application and a SageMaker AI endpoint in a second Region and use DNS failover
- Buy Provisioned Throughput for the model and route requests through an inference profile
Show the answer
Answer: A. Replace the model ID with the global. inference profile ID for the same model
With no residency rules, a global profile can route to any supported commercial Region. That gives the most throughput, at about 10% less than standard pricing, and the only change is the ID. A us. geographic profile would also reduce throttling, but it is limited to US Regions and charges standard pricing, so it is not the cheapest or highest-throughput choice.
Question 3
A GenAI developer runs a RAG application on an Amazon OpenSearch Service k-NN index that uses the faiss engine. Every query must return 5 chunks that belong to a single business unit. Right now the application runs the k-NN search with k=5 and then drops hits from other business units in application code. Some queries now return only 1 or 2 chunks. Which change best guarantees 5 relevant results without searching the whole index more broadly?
- Increase number_of_replicas so that more copies of the graph each return their own top 5 results
- Put the business-unit condition in the filter parameter inside the knn clause, so OpenSearch applies efficient k-NN filtering during the vector search
- Switch the k-NN field to the nmslib engine, which applies metadata filters before the graph search
- Keep filtering after the search, but raise k to 50 and keep the first 5 hits that match the business unit
Show the answer
Answer: B. Put the business-unit condition in the filter parameter inside the knn clause, so OpenSearch applies efficient k-NN filtering during the vector search
Efficient k-NN filtering (faiss and lucene engines) applies the filter during the vector search, so the query still returns k matching results. Raising k and filtering afterwards only makes a shortfall less likely. It costs extra work and still can't guarantee 5 matches, and nmslib is the deprecated engine, not a fix.
Question 4
A compliance officer at an insurance company says the existing CloudTrail trail must record every call that renders the ClaimsSummary managed prompt for inference, including which IAM principal made each call. The trail still uses its default settings, and the officer reports that no render events appear. What should the GenAI developer do?
- Add an advanced event selector to the trail that captures Data events for the resource type AWS::Bedrock::Prompt
- Nothing, because RenderPrompt is a management event that the trail already records by default
- Turn on Amazon Bedrock model invocation logging and send the records to Amazon S3
- Create a CloudWatch alarm on the ClaimsSummary prompt that fires whenever the prompt is invoked
Show the answer
Answer: A. Add an advanced event selector to the trail that captures Data events for the resource type AWS::Bedrock::Prompt
RenderPrompt is logged as a CloudTrail data event. Data events are off by default, so you need an advanced event selector for AWS::Bedrock::Prompt. Model invocation logging records the prompt and response text, but the requirement here is for the trail itself to record the renders, and invocation logging doesn't add anything to the trail.
Question 5
A support agent built on the Bedrock Converse API calls a Lambda-backed get_order_status tool. CloudWatch logs show that when an order ID doesn't exist, the agent calls the tool with the same ID eight times in a row. Each time the Lambda function returns {"ok": false, "error": "Lookup failed"}. The team still wants the agent to recover on its own when the order service briefly times out. Which change BEST stops the loop and keeps that recovery?
- Add a system prompt rule telling the model never to call any tool more than once per conversation
- Add exponential-backoff retries in the orchestrator for every tool result whose status is error, so the model never sees the failure
- Change toolChoice from auto to any, so the model has to pick a tool on every turn rather than retrying freely
- Return a different error for each failure: for a missing order, say the order doesn't exist and tell the model to inform the user and not retry; for a timeout, say that retrying once is safe
Show the answer
Answer: D. Return a different error for each failure: for a missing order, say the order doesn't exist and tell the model to inform the user and not retry; for a timeout, say that retrying once is safe
Retry loops happen when the error doesn't say whether a retry can help. A specific, actionable error stops pointless retries of a missing order and still allows a safe retry after a timeout. Blanket backoff retries are wrong because a missing order never succeeds on retry, and hiding the failure from the model creates a silent failure.
Question 6
A GenAI team ships through a CodePipeline V2 pipeline. Its DeployProd stage has automatic rollback turned on. The code and model ID live in Git, but the production system prompt is stored in a DynamoDB item that product managers edit through an admin console. After a manager edits that item, output quality drops, and the next pipeline run fails in DeployProd and rolls back. The prompt problem is still there. What should a GenAI developer change so that a rollback actually restores the prompt?
- Turn on automatic rollback for the Source stage as well, so that the DynamoDB item is reverted with the source revision
- Move the prompt template into the Git repository, or into a versioned store that the pipeline promotes, so that it ships as a pipeline artifact
- Turn on Amazon Bedrock model invocation logging so that the previous prompt can be found and pasted back after a failure
- Add a manual approval action before DeployProd so that a reviewer can check the DynamoDB item before each release
Show the answer
Answer: B. Move the prompt template into the Git repository, or into a versioned store that the pipeline promotes, so that it ships as a pipeline artifact
CodePipeline rollback only redeploys what the pipeline deployed, so the prompt has to be a versioned artifact that goes through the pipeline. Turning on rollback for the Source stage is wrong because CodePipeline doesn't allow rollback on the Source stage, and it still wouldn't touch a database item edited by hand.
Question 7
A bank's chatbot already uses a numbered Amazon Bedrock guardrail version in production. The team wants to add a denied topic for tax advice. The product owner worries that the new topic will block legitimate account questions. The team wants to measure how often the topic would fire on real production traffic before it affects any customer. What should the team do?
- Lower every content filter's input strength to LOW to offset the stricter denied topic.
- Add the denied topic with its input action set to NONE, publish a new version, and review the guardrail trace and assessments for detections before switching the action to BLOCK.
- Add the denied topic to the DRAFT guardrail and point production Converse calls at DRAFT, so the topic can be edited quickly if users complain.
- Add the tax terms as custom words in the word filter, because exact matching cannot produce false positives.
Show the answer
Answer: B. Add the denied topic with its input action set to NONE, publish a new version, and review the guardrail trace and assessments for detections before switching the action to BLOCK.
With an action of NONE, the guardrail detects and reports the topic in the trace without blocking, so the team can measure false positives safely on live traffic. Pointing production at DRAFT doesn't make anything safer: DRAFT is mutable and still blocks, so customers would be affected right away.
Question 8
A lending team builds SageMaker AI model cards for its Bedrock-based chatbot from a deployment pipeline. Some releases fail at the CreateModelCard step because engineers add keys such as "intended_use" or a custom "business_unit" field to the card content. The governance lead wants these errors caught before deployment, without dropping the standard card structure. What should the team do?
- Pass Content to CreateModelCard as a Python dict instead of a serialized string so that extra keys are ignored
- Validate the content JSON against the model card JSON schema in CI before calling CreateModelCard
- Create the card with ModelCardStatus set to Approved so that the schema is not checked for extra fields
- Put all the card content into resource Tags on the model card instead of the Content field
Show the answer
Answer: B. Validate the content JSON against the model card JSON schema in CI before calling CreateModelCard
The model card schema sets additionalProperties to false, so misspelt or custom keys are rejected. Checking the JSON against the schema in CI catches these errors before the API call. Passing a dict does not help, because Content must be a JSON string that follows the schema, and the status you choose does not relax validation.
Question 9
A GenAI developer is designing a two-tier router for a support assistant. Measured costs from the team's own invocation logs are $0.20 per 1,000 requests for a small model, $2.00 per 1,000 for a mid model and $8.00 per 1,000 for a large model. Evaluation shows that 70% of traffic is simple and the small model meets the quality floor on it. On the remaining 30%, both the mid and large models meet the floor. The current design sends everything to the mid model. The goal is the lowest cost that still meets the floor. What should the developer do?
- Send all traffic to the small model, because it has by far the highest score per dollar.
- Keep sending all traffic to the mid model, because a router adds misrouting risk that is never worth the saving.
- Route the simple 70% to the small model and the complex 30% to the large model, because tiering always costs less than sending all traffic to one model.
- Route the simple 70% to the small model and the complex 30% to the mid model, then re-evaluate quality on each tier's share of traffic before claiming the saving.
Show the answer
Answer: D. Route the simple 70% to the small model and the complex 30% to the mid model, then re-evaluate quality on each tier's share of traffic before claiming the saving.
The small plus mid blend costs 0.7 × 0.20 + 0.3 × 2.00 = $0.74 per 1,000 requests, against $2.00 for sending everything to mid, and both tiers meet the floor. Escalating to the large model instead costs 0.7 × 0.20 + 0.3 × 8.00 = $2.54, which is more than all-mid, so tiering is not automatically cheaper.
Question 10
An ML engineer at an insurance company hosts a fine-tuned open-weight model on an Amazon SageMaker AI endpoint. The engineer wants to score it with the same Amazon Bedrock automatic evaluation report card that the team uses for its Bedrock models, using the existing JSONL dataset in Amazon S3. The engineer does not want to build a separate evaluation harness. What should the engineer do?
- Put the SageMaker AI endpoint ARN in the bedrockModel modelIdentifier field, because that field accepts any inference ARN
- Copy the evaluation dataset to the SageMaker AI endpoint and compute NLP-F1 with a custom AWS Lambda function, because Bedrock Model Evaluations only scores Bedrock-hosted models
- Create a human-based evaluation job that compares the SageMaker AI endpoint with a Bedrock model, because only human jobs accept models hosted outside Bedrock
- Generate the model's responses to the dataset prompts, then create the automatic evaluation job with a precomputedInferenceSource in place of a bedrockModel so the job scores the supplied responses
Show the answer
Answer: D. Generate the model's responses to the dataset prompts, then create the automatic evaluation job with a precomputedInferenceSource in place of a bedrockModel so the job scores the supplied responses
With a precomputedInferenceSource you supply the responses yourself, so Bedrock Model Evaluations can score a model hosted outside Bedrock and still produce the usual report card. The Lambda option is wrong because it assumes Bedrock can only score its own hosted models, and it means building the separate harness the engineer wants to avoid.
The full timed mock
65 questions in 170 minutes, with a pass mark of 75%. Each paper is drawn from a bank of questions written for the mock alone, and these 10 are left out of it. Full access opens it, with every level of the course.
LevelUpWith is independent of the organisation that sets this exam. These questions were written for this course; they are not taken from the exam.