AWS AIP-C01 Practice Exam

AWS Certified Generative AI Developer Professional AIP-C01

The AWS Certified Generative AI Developer - Professional certification validates the ability to integrate foundation models into applications and run generative AI solutions in production on AWS. The exam covers five domains: Foundation Model Integration, Data Management, and Compliance (31%), Implementation and Integration (26%), AI Safety, Security, and Governance (20%), Operational Efficiency and Optimization for GenAI Applications (12%), and Testing, Validation, and Troubleshooting (11%). This practice exam mirrors its scenario style with single-answer and multiple-answer questions.

827 questions in the bank

Question bank updated : questions, answers, explanations and chapters.

Certification Overview

  • Exam name: AIP-C01
  • Vendor: AWS
  • Exam code: AIP-C01
  • Duration: 180 minutes
  • Total questions: about 75
  • Official passing score: 750 on a scale of 100-1,000 (source)
  • Our practice exam pass mark: 75%

Who Should Take This Exam?

Generative AI Developer, AI Engineer, Machine Learning Engineer, Software Engineer, Solutions Architect, Cloud Developer, Data Engineer, Technical Lead

Prerequisites

No prior certification is required. AWS recommends 2 or more years of building production applications on AWS or open-source technologies and 1 year of hands-on generative AI implementation.

Topics Covered

  • Foundation Model Integration, Data Management, and Compliance
  • Implementation and Integration
  • AI Safety, Security, and Governance
  • Operational Efficiency and Optimization for GenAI Applications
  • Testing, Validation, and Troubleshooting

Question Types

  • Multiple Choice (Single Answer)
  • Multiple Choice (Multiple Answers)

AIP-C01 Practice Questions

Our question bank contains 827+ practice questions for this certification. Sample questions from each exam chapter. Expand a question to see the answer choices. With full access, you get unlimited practice exams with randomized questions from our full question bank.

Foundation Model Integration, Data Management, and Compliance

A retail company runs a support assistant in eu-central-1 that calls an Amazon Bedrock model. During promotions the calls are throttled, and an earlier Regional disruption left the assistant unusable. Customer data must stay inside the EU, and the team does not want to manage several model endpoints. Which two actions improve resilience? (Choose two.)
  • A. Raise Lambda reserved concurrency so more model calls run in parallel during promotions
  • B. Deploy a second API Gateway stage in us-east-1 behind Route 53 failover routing
  • C. Invoke the model through an EU cross-Region inference profile Correct answer
  • D. Enable model invocation logging to Amazon S3 and alarm on throttle counts
  • E. Route calls through a Step Functions circuit breaker that falls back to a smaller model Correct answer

Options C and E are correct. An EU cross-Region inference profile is a single profile the application calls, and it stays inside EU geography, so customer data stays in the EU without the team managing several endpoints, while spreading requests across more than one Region. A Step Functions circuit breaker that falls back to a smaller model keeps the assistant answering when the primary model is throttled or unavailable. Option A only raises how many calls Lambda makes in parallel, which does not change what Bedrock accepts. Option B sends calls through us-east-1, outside the EU, and adds a second stack to manage. Option D observes throttling but does not change how the assistant behaves when it happens.

A team stores document embeddings in Amazon Aurora PostgreSQL and wants to run nearest-neighbor similarity queries over them with SQL. Which extension must be enabled on the cluster?
  • A. pg_cron
  • B. pgvector Correct answer
  • C. PostGIS
  • D. pg_trgm

Option B is correct because pgvector is the PostgreSQL extension that adds a vector column type and similarity operators, and it supports HNSW and IVFFlat indexes, so nearest-neighbor queries run in plain SQL. pg_cron schedules jobs, PostGIS handles geospatial data, and pg_trgm does trigram text matching, so none of them search embeddings.

An insurer's claims assistant calls Amazon Bedrock from AWS Lambda. Product owners want to change the model ID and inference parameters in production during business hours, ramp the change over 20 minutes, and roll it back automatically if the function's error alarm fires. No function code may be redeployed. Which solution meets these requirements?
  • A. Put the model ID in an API Gateway stage variable and redeploy the stage with a canary release setting that sends a share of requests
  • B. Keep the model ID and parameters in an AWS AppConfig profile, deployed with a linear strategy and a CloudWatch alarm monitor Correct answer
  • C. Save the model ID as an AWS Systems Manager Parameter Store parameter that the function reads on every invocation
  • D. Store the model ID in a Lambda environment variable and change it with the UpdateFunctionConfiguration API during the change window

Option B is correct. AWS AppConfig holds the model ID and inference parameters as configuration the function reads, so no code is redeployed. A linear deployment strategy ramps the change over the 20 minutes, and a CloudWatch alarm attached as a monitor makes AppConfig roll the deployment back automatically if the error alarm fires. Option A changes an API Gateway stage variable, which does not reach the Lambda function's model call. Option C has no gradual ramp and no alarm-driven rollback. Option D changes the function configuration all at once and has no ramp or automatic rollback.

A fintech team serves a fine-tuned open-weight model from an Amazon SageMaker AI real-time endpoint. A retrained model version must replace the current one with no downtime. Traffic should move to the new version in two steps, and a rise in the endpoint's 5XX error alarm must return all traffic to the old version automatically. The team wants the LEAST operational overhead. Which approach should it use?
  • A. Deploy the new model to a second endpoint behind Route 53 weighted records, with a Lambda function that shifts and reverts weights
  • B. Create a shadow variant for the new model and promote it to production once its latency and error metrics match the current variant
  • C. Update the endpoint using a blue/green deployment with canary traffic shifting and the alarm set as the automatic rollback trigger Correct answer
  • D. Add the new model as a second production variant weighted 95/5 and run a Lambda function that resets the weights when the alarm fires

Option C is correct. A SageMaker AI blue/green deployment with canary traffic shifting moves traffic to the new fleet in two steps with no downtime, and an alarm configured for automatic rollback returns all traffic to the old fleet if the 5XX alarm trips. It is configured on the endpoint update itself, so it has the least operational overhead. Option A needs a second endpoint plus a Lambda function to shift and revert weights. Option B uses a shadow variant, which receives mirrored copies of requests and is not a two-step traffic shift with rollback. Option D relies on custom Lambda code to reset the weights.

Implementation and Integration

After a Strands Agents agent is deployed on AWS Lambda, each invocation starts with an empty conversation because the execution environment is frequently new. Which change preserves the conversation across invocations?
  • A. Attach a session manager that persists agent state to Amazon S3 Correct answer
  • B. Increase the function memory so the conversation stays in process
  • C. Write the message history to the /tmp directory of the function
  • D. Turn on provisioned concurrency so one environment serves each user

Option A is correct because a session manager that persists the agent's messages and state to Amazon S3 reloads the conversation at the start of each invocation, whichever execution environment serves it. Option B is wrong because more memory does not stop Lambda from creating new environments. Option C is wrong because /tmp is local to one environment and is lost with it. Option D is wrong because provisioned concurrency does not tie an environment to a user.

A support assistant built on Amazon Bedrock AgentCore must remember each customer's stated preferences from sessions held weeks earlier, and the prompt must not carry old transcripts. What should the developer do?
  • A. Enable prompt caching so earlier transcripts persist between separate sessions
  • B. Replay the complete stored transcripts into the system prompt of each new session
  • C. Use AgentCore Memory long-term strategies to extract and retrieve preferences Correct answer
  • D. Fine-tune the model on each customer's past conversations every night

Option C is correct because AgentCore Memory long-term strategies extract preferences from sessions and retrieve them in later sessions, so the prompt carries only the relevant facts. Option A is wrong because prompt caching only reduces cost within a short cache window and does not store memory. Option B is wrong because replaying full transcripts is exactly what the stem rules out. Option D is wrong because nightly fine-tuning per customer is costly and cannot keep up with new preferences.

Follow-up messages such as 'yes, book it' reach the wrong specialist in a retail assistant built with AWS Agent Squad on Lambda, because the classifier sees no earlier turns. Which change fixes the routing?
  • A. Replace the classifier model with one that has a larger output limit
  • B. Configure DynamoDB-backed chat storage keyed by user ID and session ID Correct answer
  • C. Increase the Lambda memory so in-memory chat storage retains more turns
  • D. Add few-shot examples of follow-up phrases to each specialist's system prompt

Option B is correct because Agent Squad's classifier needs the conversation history to route a short follow-up, and DynamoDB chat storage keyed by user ID and session ID keeps that history across stateless Lambda invocations. Option A is wrong because output limit has nothing to do with missing context. Option C is wrong because in-memory storage is lost when the environment changes. Option D is wrong because few-shot phrases cannot tell the classifier which specialist handled the previous turn.

A developer's IAM role must be able to call both the Converse and ConverseStream operations of Amazon Bedrock Runtime for one foundation model. Which TWO IAM actions must the identity policy allow? (Choose two.)
  • A. bedrock:GetFoundationModel
  • B. bedrock:InvokeModel Correct answer
  • C. bedrock:ListInferenceProfiles
  • D. bedrock:InvokeModelWithResponseStream Correct answer
  • E. bedrock:CreateModelInvocationJob

Options B and D are correct because Converse is authorized by bedrock:InvokeModel and ConverseStream by bedrock:InvokeModelWithResponseStream. GetFoundationModel and ListInferenceProfiles only read metadata, and CreateModelInvocationJob starts batch inference, so none of them permit either runtime call.

AI Safety, Security, and Governance

Which Amazon Bedrock Guardrails filter is designed to detect user inputs that try to override the application's instructions or bypass its safety behavior?
  • A. Denied topics
  • B. Prompt attack content filter Correct answer
  • C. Contextual grounding check
  • D. Word filters

Option B is correct because the prompt attack content filter in Amazon Bedrock Guardrails detects jailbreak and prompt injection attempts in user input. Denied topics and word filters match subject matter or terms rather than override attempts, and the contextual grounding check evaluates responses against source material.

A chatbot calls the Converse API with a system prompt and the user's message. The prompt attack filter must evaluate the user's text but not the application-written instructions. What should the developer do?
  • A. Increase the prompt attack filter strength to HIGH on the guardrail
  • B. Mark the user's text as guard content so the guardrail evaluates just that part Correct answer
  • C. Add a denied topic that describes the system prompt wording
  • D. Call ApplyGuardrail with the source set to OUTPUT before each request

Option B is correct because the guardrail evaluates only the content tagged as guard content when it is supplied in the Converse request, so the application's own system prompt is not scanned for prompt attacks. A higher filter strength, a denied topic or an OUTPUT-source ApplyGuardrail call does not limit evaluation to the user's text.

An application invokes a model hosted on an Amazon SageMaker AI endpoint, and the security organization wants the same input moderation policy already defined in an Amazon Bedrock guardrail. How should the application apply it?
  • A. Pass the guardrail identifier in the endpoint's InvokeEndpoint request header
  • B. Call the ApplyGuardrail API with the user text before invoking the endpoint Correct answer
  • C. Associate the guardrail with the endpoint through a resource-based policy
  • D. Attach the guardrail to the SageMaker AI endpoint configuration

Option B is correct because the ApplyGuardrail API evaluates text independently of any model, so the application can run the existing guardrail on the input before calling the SageMaker AI endpoint. Endpoints have no header, resource policy or endpoint configuration setting for attaching a Bedrock guardrail.

Auditors of a retailer's credit-scoring model on Amazon SageMaker AI ask which dataset version and processing job produced the model deployed last quarter. The evidence must survive deletion of the endpoint and its training jobs. Which two actions provide it? (Choose two.)
  • A. Keep the training jobs' CloudWatch log streams for one year
  • B. Enable Model Monitor data capture on the endpoint
  • C. Register each model in the SageMaker Model Registry as a versioned model package Correct answer
  • D. Tag the endpoint with the dataset's S3 prefix
  • E. Build the workflow with SageMaker Pipelines so lineage links datasets, jobs and model Correct answer

Options C and E are correct because the Model Registry keeps each model as a versioned model package that outlives the endpoint and training jobs, and SageMaker Pipelines records lineage linking datasets, processing jobs and the model. Log streams, Model Monitor data capture and an endpoint tag do not link a specific dataset version and processing job to the deployed model, and the tag is lost with the endpoint.

Operational Efficiency and Optimization for GenAI Applications

A support assistant calls Anthropic Claude Sonnet 4 on demand in Amazon Bedrock. A typical request has 2,000 input tokens and 400 output tokens, and for this model each output token counts five times toward the tokens-per-minute quota. The quota is 400,000 tokens per minute. About how many typical requests per minute can the assistant sustain?
  • A. About 91
  • B. 200
  • C. 100 Correct answer
  • D. About 167

Option C is correct because each request burns 2,000 input tokens plus 400 x 5 = 2,000 output-equivalent tokens, which is 4,000 quota tokens, and 400,000 / 4,000 = 100 requests per minute. Option B ignores the five-times output burndown, and options A and D come from miscounting the burndown.

Which service adjusts the number of instances behind an Amazon SageMaker AI real-time endpoint variant by tracking a CloudWatch metric target?
  • A. Amazon SageMaker Inference Recommender
  • B. AWS Compute Optimizer
  • C. Application Auto Scaling Correct answer
  • D. Amazon EC2 Auto Scaling

Option C is correct because SageMaker AI endpoint variants scale through Application Auto Scaling with a target-tracking policy on a CloudWatch metric such as invocations per instance. EC2 Auto Scaling manages EC2 groups, Inference Recommender only benchmarks instance types, and Compute Optimizer only recommends sizes.

A batch summarization job and an interactive chat application share one AWS account and the same on-demand model quota. Nightly batch runs push the chat application into throttling. Which two changes protect the chat application? (Choose two.)
  • A. Raise the memory setting of the Lambda function that runs the batch job
  • B. Turn on model invocation logging in the batch account
  • C. Run the batch job in its own AWS account Correct answer
  • D. Lower the temperature of the batch requests to 0
  • E. Buy Provisioned Throughput for the chat application's model Correct answer

Options C and E are correct. A separate account has its own on-demand quota, so batch runs can no longer throttle the chat application, and Provisioned Throughput gives the chat model dedicated capacity that batch traffic cannot consume. Lambda memory, invocation logging and temperature do not change how the shared quota is consumed.

A synchronous chat workflow built with AWS Step Functions runs a Lambda function that only wraps one Bedrock call, then two short steps, and finishes in about 20 seconds. Which two changes cut orchestration overhead? (Choose two.)
  • A. Run the workflow as an Express workflow with synchronous execution Correct answer
  • B. Place an Amazon SQS queue between each pair of steps
  • C. Call Bedrock through the optimized Step Functions integration and drop the wrapper function Correct answer
  • D. Route every call through one shared Lambda function
  • E. Poll for each step's result in Wait states of one second

Options A and C are correct. An Express workflow with synchronous execution avoids the start-up and state-transition overhead of a standard workflow for a short chat flow, and the optimized Bedrock integration removes the wrapper Lambda function and its cold start and extra hop. Queues between steps, a shared function and one-second Wait polling all add delay.

Testing, Validation, and Troubleshooting

Which metric in an Amazon Bedrock automatic (programmatic) model evaluation job measures how much a model's output changes when the prompts contain typos or random capitalization?
  • A. Context coverage
  • B. Robustness Correct answer
  • C. Toxicity
  • D. Accuracy

Option B is correct. Robustness is the automatic evaluation metric that perturbs prompts (typos, random capitalization, number-word swaps) and measures how much the output changes. Accuracy compares to references, toxicity scores harmful content, and context coverage belongs to RAG evaluation.

A team prepares a custom prompt dataset for an Amazon Bedrock model evaluation job. Which format and location does the job read?
  • A. A CSV file in Amazon S3, with one prompt in each row
  • B. A Parquet table registered in the AWS Glue Data Catalog
  • C. A JSON Lines file in Amazon S3, with one record per line Correct answer
  • D. A JSON array stored in a single Amazon DynamoDB item

Option C is correct. Bedrock evaluation jobs read custom prompt datasets as JSON Lines files stored in Amazon S3, one record per line. CSV, Glue tables and DynamoDB items are not accepted dataset sources.

Which TWO metrics are built-in metrics of an Amazon Bedrock LLM-as-a-judge model evaluation job (not a RAG evaluation job)? (Choose two.)
  • A. Robustness
  • B. Citation precision
  • C. Context coverage
  • D. Coherence Correct answer
  • E. Helpfulness Correct answer

Options D and E are correct because Coherence and Helpfulness are built-in metrics of a Bedrock LLM-as-a-judge model evaluation job. Citation precision (B) and context coverage (C) belong to RAG evaluation jobs, and Robustness (A) is an automatic algorithmic metric, not a judge metric.

A publisher's assistant summarizes 40-page reports, and no reference summaries exist. The team must measure how often summaries contain statements that the report does not support, across 3,000 summaries a week. Which TWO actions build this measurement? (Choose two.)
  • A. Run the automatic accuracy metric with BERTScore, comparing each summary with the full text of its report
  • B. Rate a random sample of about 100 summaries each week in a human-based job to check the judge's ratings Correct answer
  • C. Run the automatic toxicity metric on each summary and treat the low scores as a sign of faithful summaries
  • D. Run a retrieve-only RAG evaluation on the summaries and read the context coverage score for each report
  • E. Define a custom metric in an LLM-as-a-judge job whose judge prompt receives the report and the summary and rates unsupported statements Correct answer

Options B and E are correct because with no reference summaries the team needs a custom LLM-as-a-judge metric that sees the report and the summary and rates unsupported statements, plus a weekly human-rated sample to check that the judge's ratings can be trusted. BERTScore (A) needs references, toxicity (C) does not measure faithfulness, and retrieve-only context coverage (D) scores retrieval, not summaries.

Free Study Guide

Read the free AWS AIP-C01 study guide: every exam topic explained in plain English, with diagrams, exam tips and the mistakes to avoid.

More AWS Practice Exams

All AWS practice exams →

Practice with realistic mock exams to prepare for your AWS certification.