Don’t Sit the AWS Generative AI Developer — Professional Exam Until You Try This

Master GenAI on AWS with Real Exam-Style Questions
Are you preparing for the AWS Certified Generative AI Developer — Professional certification? This challenging exam validates your ability to design, develop, and manage enterprise-grade Generative AI solutions on AWS. If you’re serious about passing on your first attempt, you need more than just theoretical knowledge — you need practice with real, scenario-based questions that mirror the actual exam.

The Solution
I’ve created a comprehensive 150+ question practice guide specifically designed to bridge the gap between theoretical knowledge and exam-day success. The guide includes:

•150 high-fidelity practice questions covering all five exam domains

•Detailed explanations for every answer, clarifying why the correct choice is best and why distractors are suboptimal

•Real-world scenarios with fictional companies across diverse industries (retail, telecom, healthcare, finance, manufacturing, and more)

•Professional-level complexity that matches the actual exam difficulty

•Domain-by-domain organization aligned with the official exam guide

🔥 Below, I’ve included 20 sample questions from the guide. Try them out, check your answers against the answer key, and see how well you understand the material. If you find these challenging and educational, the complete guide with all 150 questions is available on Amazon.

Domain 1: Design and Development (Questions 1–4)
Q1. A major retail chain, “OmniMart,” wants to build a personalized shopping assistant that analyzes real-time video feeds from in-store cameras to identify products a customer is looking at. The assistant should then provide product details, suggest complementary items, and answer questions. The solution must scale to thousands of stores with minimal on-site infrastructure and operational overhead. Which architecture should a GenAI developer propose?

A) Use Amazon SageMaker to train a custom model and deploy it to a fleet of Amazon EC2 G5 instances, with a central API for all stores.

B) Stream video to Amazon Kinesis Video Streams, use Amazon Rekognition for object detection, and an Amazon Bedrock agent for conversational AI, with results cached on Amazon ElastiCache.

C) Deploy custom computer vision models on AWS Snowball Edge devices in each store for local processing.

D) Install a self-hosted inference server in each store, connected to a central Amazon RDS database for product information.

Q2. A telecommunications giant, “AetherLink Telecom,” is using an Amazon Bedrock Knowledge Base to power a customer support chatbot. The knowledge base ingests millions of technical manuals from an Amazon S3 bucket. During ingestion, some complex PDF documents with intricate diagrams are failing to process correctly, leading to incomplete answers. The operations team needs a robust way to monitor and automatically retry failed ingestion jobs for specific documents. Which solution is the MOST resilient?

A) Manually monitor the Amazon Bedrock console for failed jobs and re-initiate them through the AWS CLI.

B) Set up an Amazon S3 Event Notification on the source bucket to trigger a Lambda function that calls the Bedrock StartIngestionJob API.

C) Create an Amazon EventBridge rule that listens for Bedrock Ingestion Job Failed events, sends the event to an Amazon SQS dead-letter queue (DLQ), and triggers a Lambda function to process and retry from the DLQ.

D) Use Amazon CloudWatch Logs to filter for ingestion failures and trigger an AWS Lambda function that manually re-ingests the failed document.

Q3. An investment bank, “Apex Financial Group,” is developing a RAG application to summarize financial reports. The system retrieves relevant sections from 10-K filings, but the final summaries from the Anthropic Claude v2 model on Amazon Bedrock are often too generic and miss critical nuances. The retrieved text chunks are highly relevant, but the model struggles to synthesize them effectively. Which combination of prompt engineering techniques should be used to improve the quality of the generated summaries? (Select TWO.)

A) Implement a chain-of-thought prompting strategy that instructs the model to first extract key financial figures, then identify trends, and finally generate the summary.

B) Increase the temperature parameter to 1.0 to make the model more creative.

C) Use few-shot prompting by providing high-quality examples of report sections and their corresponding ideal summaries within the prompt.

D) Decrease the number of retrieved documents to reduce the context size.

E) Switch to a smaller, faster model to reduce latency.

Q4. A healthcare analytics firm, “Veridian Health Analytics,” is building a platform for researchers to find similar medical images (e.g., MRIs, X-rays) across a massive, petabyte-scale archive of 100 million images. Searches are infrequent but must be highly accurate and responsive when they occur. The solution must be HIPAA-compliant and minimize costs, as the company only wants to pay for the queries that are actually run. Which vector database solution is the MOST appropriate?

A) A self-managed FAISS index on a large Amazon EC2 instance with EBS storage.

B) Amazon OpenSearch Serverless, configured with a vector search index and VPC endpoint for security.

C) Amazon RDS for PostgreSQL with the pgvector extension on a provisioned cluster.

D) Amazon DynamoDB with a custom secondary index designed to approximate similarity search.

Domain 2: Enterprise Integration and Application Development (Questions 5–9)
Q5. A food delivery company, “InstaFeast,” needs to deploy existing Python agent code to Amazon Bedrock AgentCore Runtime. The agent must handle quick data lookups requiring sub-second responses and comprehensive research report generation with streaming responses over several minutes. The solution must automatically manage HTTP server configuration, endpoint routing, and health monitoring. Which deployment approaches will meet these requirements with MINIMAL operational overhead? (Select TWO.)

A) Amazon Bedrock Agents (Serverless) with automatic scaling

B) Self-managed agent infrastructure on Amazon EC2

C) AWS Lambda with Bedrock Agent Runtime for flexible, serverless execution

D) Amazon ECS with manual configuration and health monitoring

E) On-premises server deployment with AWS Direct Connect

Q6. A news media company, “The Daily Chronicle,” wants to develop a content conformance tool that automatically reviews and adjusts articles to ensure compliance with a style guide. Journalists need a web-based article editor that provides real-time analysis of content upon request. When journalists click “analyze,” the system should immediately begin providing suggested revisions through the editor interface. Which architecture will meet these requirements with the LEAST operational overhead?

A) Amazon Bedrock Streaming API with WebSocket/Server-Sent Events for real-time delivery

B) Amazon Bedrock batch processing with scheduled analysis

C) Custom-built model serving infrastructure

D) Amazon SageMaker Batch Transform jobs

Q7. A ride-hailing app, “SwiftRide,” needs secure authentication for a third-party application that uses Amazon Bedrock. The solution must integrate with the company’s existing identity provider (IdP), maintain comprehensive audit logs, and eliminate long-lived credentials. Which solutions will meet these requirements? (Select TWO.)

A) AWS IAM Identity Center (SSO) with SAML/OIDC integration

B) AWS Security Token Service (STS) with AssumeRole for temporary credentials

C) Long-lived IAM user access keys stored in the application

D) AWS Secrets Manager for credential storage

E) Manual credential rotation processes

Q8. A real estate tech company, “UrbanScape Properties,” uses an AI assistant to answer customer questions based on internal company documents. The company wants to include new documents in the assistant’s responses as soon as possible and exclude deleted documents immediately. Documents are stored in Amazon S3, and the AI assistant uses Amazon Bedrock Knowledge Bases. A GenAI developer must create a scalable, event-driven, and resilient solution. Which solution will meet these requirements?

A) Amazon S3 Event Notifications triggering AWS Lambda to update the Bedrock Knowledge Base

B) Manual document uploads to the knowledge base

C) Scheduled batch jobs to sync documents daily

D) Amazon S3 Select for document filtering

Q9. A gaming company, “PixelStorm Entertainment,” is implementing AI governance policies requiring all FM interactions to be secured with guardrails. The company configures Amazon Bedrock guardrails and must ensure all InvokeModel and Converse API calls apply the guardrails. Which solution will enforce guardrail compliance for the API calls in the MOST operationally efficient way?

A) Configure guardrails directly on the Bedrock model or use a centralized API Gateway

B) Implement guardrails in each application code

C) Manually review all API calls for compliance

D) Use AWS WAF to filter requests

Domain 3: Security and Compliance (Questions 10–13)
Q10. A financial services company, “Prosperity Bank,” is developing a mobile app to help users with account inquiries. The company has email exchange data between customers and support staff to use as source material. The data is stored in Amazon S3 and contains personally identifiable information (PII) that should not appear in search results. Which solution will meet these requirements?

A) Use Amazon Comprehend for PII detection and Amazon Bedrock Guardrails for output filtering

B) Manually review all documents for PII

C) Store PII data in a separate, unencrypted location

D) Disable search functionality for sensitive data

Q11. A telecommunications provider, “Nexus Telecom,” is designing a GenAI solution to predict customer churn. The solution must validate feasibility, performance characteristics, and business value before full-scale deployment. The provider wants to test with a subset of customer data first. Which approach should be taken?

A) Develop a technical proof-of-concept using Amazon Bedrock and SageMaker

B) Deploy directly to production without testing

Write on Medium
C) Use only theoretical analysis without any implementation

D) Hire external consultants for design

Q12. A bank, “Keystone Trust Bank,” is implementing data validation and processing pipelines for FM consumption. The bank needs to ensure data meets quality standards before being used for model training. Which services can be used for data validation? (Select TWO.)

A) AWS Glue Data Quality

B) SageMaker Data Wrangler

C) Manual spreadsheet reviews

D) Custom Lambda functions

E) No validation needed

Q13. A financial services company, “Quantum Capital,” operates a RAG application that answers user questions using internal market analysis reports. The application uses Amazon Bedrock for the embedding model and Amazon OpenSearch Service as the vector store. After a recent Lambda function code update, the application returns generic responses even for questions that previously returned accurate answers. CloudWatch Logs shows no errors, AWS X-Ray confirms successful FM invocation, and OpenSearch is healthy. What is the cause of this issue?

A) The embedding model or search query logic was changed in the code update

B) The vector store is corrupted

C) The FM model was changed

D) The network connection is unstable

Domain 4: Operations and Support (Questions 14–17)
Q14. An e-commerce company, “LuxeLane,” has deployed a product recommendation engine using a fine-tuned model on Amazon SageMaker. The company has noticed that the inference costs are significantly higher than projected. Analysis shows that the endpoint is often idle during off-peak hours but is provisioned for peak capacity. Which solution would be the MOST cost-effective for this workload?

A) Use a SageMaker Serverless Inference endpoint that automatically scales to zero when not in use.

B) Manually scale down the number of instances during off-peak hours using a scheduled script.

C) Switch to a smaller instance type, even if it increases latency during peak hours.

D) Purchase a SageMaker Savings Plan to cover the cost of the provisioned instances.

Q15. An analytics company, “CogniData,” provides a service that summarizes large documents for its clients. The service uses Amazon Bedrock with the Anthropic Claude model. To manage costs, the company wants to implement a system that estimates the number of tokens a document will consume before it is sent to the model. This will allow them to provide upfront pricing to their clients. What is the MOST accurate way to implement this?

A) Use a simple character-to-token ratio (e.g., 4 characters per token) to estimate the token count.

B) Use the official Anthropic tokenizer library in a Lambda function to precisely calculate the token count for a given text.

C) Send the first 1,000 characters of the document to the model and extrapolate the token count.

D) Assume a fixed token count for all documents regardless of their length.

Q16. A gaming company, “PixelPlay Games,” wants to develop high-performance FM systems to maximize resource utilization and throughput. Which strategies should be used? (Select TWO.)

A) Batching strategies for request optimization

B) Capacity planning and utilization monitoring

C) Using a single large instance to handle all traffic

D) Auto-scaling configurations

E) Manually scaling instances based on daily traffic patterns

Q17. A content delivery network, “SwiftStream CDN,” needs to create intelligent caching systems to reduce costs and improve response times. Which caching strategies should be used?

A) Semantic caching, result fingerprinting, edge caching, prompt caching.

B) Caching all results indefinitely to maximize hit rate.

C) Using a single, centralized cache for all applications.

D) Manually clearing the cache on a daily basis.

Domain 5: Monitoring and Troubleshooting (Questions 18–20)
Q18. A media company, “Chronicle News,” uses a RAG system to answer questions about its vast archive of news articles. Recently, users have complained that the answers are becoming outdated. For example, when asked “Who is the CEO of ExampleCorp?”, the system gives a name from two years ago. The vector database is being updated correctly. What is the MOST likely cause of this issue?

A) The embedding model used for the query is different from the one used to embed the documents.

B) The reranker model is prioritizing older, more detailed documents over newer, more concise ones.

C) The document chunking strategy is creating chunks that are too large, causing the model to lose focus.

D) The prompt is not explicitly instructing the model to prioritize the most recent information.

Q19. A retail company, “StyleSphere,” has a GenAI application that generates marketing copy. The marketing team wants to evaluate two different foundation models from Amazon Bedrock (e.g., Anthropic Claude and AI21 Labs Jurassic) to see which one produces more engaging content. They want to run a live experiment with a small percentage of users. Which approach is the MOST appropriate for this evaluation?

A) A/B testing, where 1% of users are shown copy from Model A and another 1% are shown copy from Model B, with engagement metrics (e.g., click-through rate) being compared.

B) Offline evaluation, where a set of prompts is run through both models and the outputs are manually scored by the marketing team.

C) Shadow testing, where both models generate copy for all requests, but only the output from the current production model is shown to users.

D) Manually switching between the two models in the production environment on a daily basis.

Q20. A quality assurance platform, “QA-Pro,” needs to create systematic quality assurance processes to maintain consistent performance standards. Which mechanisms should be used?

A) Continuous evaluation workflows, regression testing, automated quality gates.

B) Not implementing any QA processes to speed up development.

C) Manually testing the system before each release.

D) Relying on user complaints to identify quality issues.

Answer Key
Q1: B — This solution is fully managed and serverless, minimizing operational overhead. Kinesis Video Streams is designed for real-time video ingestion, Rekognition for scalable analysis, and Bedrock for conversational AI, which is more efficient than managing EC2 fleets or on-site hardware.

Q2: C — This is the most resilient and automated solution. Using EventBridge to capture the specific failure event and an SQS DLQ ensures that failed ingestion jobs are not lost and can be retried systematically by a Lambda function, preventing data loss and manual intervention.

Q3: A, C — Chain-of-thought prompting guides the model through a logical reasoning process, improving synthesis. Few-shot prompting provides concrete examples, which is highly effective for teaching the model the desired output format and quality for a specific task like summarization.

Q4: B — Amazon OpenSearch Serverless is ideal for infrequent, large-scale similarity searches. Its pay-per-use model is the most cost-effective for this workload, and as a managed service, it supports HIPAA compliance and eliminates infrastructure management.

Q5: A, C — Amazon Bedrock Agents (Serverless) and AWS Lambda with Bedrock Agent Runtime both automatically manage infrastructure, routing, and health monitoring.

Q6: A — Amazon Bedrock Streaming API with WebSocket/Server-Sent Events enables real-time delivery of suggestions to the web editor interface.

Q7: A, B — AWS IAM Identity Center with SAML/OIDC and AWS STS with AssumeRole both provide temporary credentials, IdP integration, and comprehensive audit logging.

Q8: A — Amazon S3 Event Notifications trigger AWS Lambda to update the Bedrock Knowledge Base in real-time for scalable, event-driven synchronization.

Q9: A — Configuring guardrails directly on the Bedrock model or using a centralized API Gateway ensures all API calls apply guardrails with minimal operational overhead.

Q10: A — Amazon Comprehend for PII detection combined with Amazon Bedrock Guardrails for output filtering protects sensitive information from appearing in search results.

Q11: A — Developing a technical proof-of-concept using Amazon Bedrock and SageMaker validates feasibility, performance, and business value before full-scale deployment.

Q12: A, B — AWS Glue Data Quality and SageMaker Data Wrangler validate data quality for FM consumption.

Q13: A — The embedding model or search query logic was likely changed in the code update, causing the retrieval system to return irrelevant documents.

Q14: A — Amazon SageMaker Serverless Inference automatically scales to zero during idle periods, making it the most cost-effective solution for this workload.

Q15: B — Using the official Anthropic tokenizer library provides the most accurate token count estimation, enabling precise upfront pricing.

Q16: A, B, D — Batching strategies, capacity planning, utilization monitoring, and auto-scaling maximize resource utilization.

Q17: A — Semantic caching, result fingerprinting, edge caching, and prompt caching reduce costs and improve response times.

Q18: B — The reranker model is prioritizing older, more detailed documents over newer, more concise ones, causing the outdated information to be retrieved.

Q19: A — A/B testing with live users and engagement metrics is the most appropriate approach for evaluating which model produces more engaging marketing copy.

Q20: A — Continuous evaluation workflows, regression testing, and automated quality gates maintain consistent performance standards.