View Full Amazon AWS Certified Generative AI Developer – Professional AIP-C01 Exam Dumps and Practice Test Dumps

 

Question 301

Which component is responsible for representing documents as numerical vectors for storage in a vector search system?

1. Embedding model
2. API Gateway
3. AWS CloudTrail
4. AWS Secrets Manager

Correct Answer: 1

Explanation

An embedding model converts text or other supported content into numerical vector representations that capture semantic characteristics. These vectors can be stored in a vector index and compared with query embeddings to identify semantically related information. Embeddings are commonly used in RAG architectures to retrieve relevant passages before generation. API Gateway manages APIs, CloudTrail records AWS activity, and Secrets Manager protects credentials. Developers should ensure that document and query embeddings use compatible models and configurations so that similarity calculations produce meaningful retrieval results.

Question 302

A company wants to restrict an AI assistant so that it does not discuss specific prohibited subjects. Which capability should the developer configure?

1. Amazon S3 lifecycle policies
2. Bedrock Guardrails topic restrictions
3. AWS KMS key rotation
4. CloudWatch dashboards

Correct Answer: 2

Explanation

Bedrock Guardrails can help organizations define policies that restrict certain topics and control undesirable content in generative AI interactions. Topic restrictions can be useful when business or regulatory requirements specify subjects that an assistant should refuse to discuss. Guardrails should be combined with authentication, authorization, input validation, and other application-level controls. S3 lifecycle policies manage stored objects, KMS handles encryption keys, and CloudWatch provides monitoring. Guardrails specifically address safety and content-control requirements within supported generative AI workflows.

Question 303

A RAG application retrieves documents correctly, but responses frequently contain unsupported claims. Which area should the developer investigate?

1. S3 storage class
2. IAM username naming
3. Grounding instructions and response evaluation
4. API Gateway domain configuration

Correct Answer: 3

Explanation

Correct retrieval does not guarantee that a model will use retrieved information appropriately. Developers should investigate whether the prompt clearly instructs the model to ground factual responses in the supplied context and avoid unsupported claims. Output evaluation should measure groundedness and identify statements that are not supported by retrieved evidence. Context formatting, document quality, and model behavior may also need review. Storage classes, IAM username names, and API domain configuration do not directly address unsupported model claims. Groundedness should be treated as an important quality dimension.

Question 304

A developer wants an AI workflow to process several independent documents simultaneously and then combine their results. Which approach is most appropriate?

1. Sequential processing only
2. Manual processing
3. One extremely large prompt
4. Parallel workflow execution followed by aggregation

Correct Answer: 4

Explanation

Independent document-processing tasks can often execute in parallel, reducing the overall workflow duration compared with sequential processing. After the parallel branches complete, an aggregation step can combine or summarize their outputs. AWS Step Functions can support parallel workflow patterns and coordinate downstream tasks. Sending every document in one prompt can create context-size, cost, and processing challenges. Manual processing does not scale effectively. Parallel execution should still account for service quotas, concurrency limits, error handling, and the requirements of downstream aggregation.

Question 305

A developer wants to reduce the cost of an AI application without significantly changing its response quality. Which action should be considered first?

1. Use larger prompts
2. Reduce unnecessary input and output tokens
3. Increase model temperature
4. Retrieve every available document

Correct Answer: 2

Explanation

Reducing unnecessary input and output tokens can lower inference costs while preserving the information needed for useful responses. Developers can remove redundant prompt instructions, retrieve only relevant context, use appropriate chunk sizes, limit unnecessary output, and avoid sending duplicate information. Model selection and caching can also contribute to cost optimization. Increasing prompt size or retrieving every available document usually increases processing costs. Temperature affects generation behavior rather than directly reducing token usage. Any optimization should be evaluated to ensure that response quality remains within required thresholds.

Question 306

A company wants to allow an AI agent to call only two approved internal APIs. Which approach best follows least-privilege principles?

1. Give the agent administrator access
2. Allow unrestricted network access
3. Expose only the required tools with restricted permissions
4. Give the agent access to every available API

Correct Answer: 3

Explanation

Least privilege means providing only the capabilities required for an agent to perform its intended tasks. If an agent needs two internal APIs, those APIs should be explicitly exposed and protected with appropriate authorization and validation. Unrelated tools and services should remain inaccessible. Administrator access or unrestricted network connectivity creates unnecessary risk if the agent is manipulated or compromised. Tool permissions should be enforced by trusted application and infrastructure controls rather than relying solely on model instructions. Monitoring and audit logging can provide additional protection.

Question 307

A developer wants a workflow to retry a failed external API call three times and then follow an alternative path. Which capability is most appropriate?

1. Prompt expansion
2. Vector indexing
3. Controlled retry and error-handling logic
4. Increasing model temperature

Correct Answer: 3

Explanation

Controlled retry and error-handling logic allows a workflow to attempt recovery from transient failures while preventing endless execution. The workflow can specify a maximum number of retries, use exponential backoff, and then transition to a fallback or failure-handling path when attempts are exhausted. AWS Step Functions can implement retry and catch behavior for orchestrated workflows. Prompt size and temperature do not provide workflow-level failure handling. Retry policies should distinguish transient errors from permanent failures so that inappropriate retries do not increase latency or cost.

Question 308

A company wants to ensure that a newly released AI application version can be quickly replaced if serious problems appear. Which deployment practice is most useful?

1. Maintain a rollback-capable deployment strategy
2. Delete the previous version immediately
3. Disable monitoring
4. Change prompts manually after deployment

Correct Answer: 1

Explanation

A rollback-capable deployment strategy allows a team to return to a previously validated application version if the new release causes serious errors or quality degradation. This is particularly valuable for AI applications because changes to models, prompts, retrieval logic, or agent tools can have unexpected effects. Canary or blue/green deployment approaches can further reduce risk by maintaining a known-good version while the new release is evaluated. Removing the previous version immediately eliminates an important recovery option. Monitoring should remain active throughout deployment and rollback operations.

Question 309

A developer needs to store application credentials securely and retrieve them at runtime without placing them in source code. Which AWS service is most appropriate?

1. Amazon CloudWatch
2. Amazon S3
3. AWS Secrets Manager
4. AWS Step Functions

Correct Answer: 3

Explanation

AWS Secrets Manager provides secure storage and retrieval of credentials such as passwords, API keys, and other sensitive values. Applications can retrieve secrets at runtime while IAM policies control which workloads are permitted to access them. This reduces the risk of credentials being exposed through source code repositories or configuration files. Secrets can also be rotated according to organizational requirements. CloudWatch focuses on monitoring, S3 provides object storage, and Step Functions orchestrates workflows. Secrets should also be prevented from appearing in application logs or model prompts.

Question 310

A RAG application retrieves 20 passages for every question, but only the top few are usually relevant. Which change could improve efficiency?

1. Increase the number of retrieved passages
2. Retrieve and rerank a smaller candidate set
3. Add duplicate passages
4. Increase output-token limits

Correct Answer: 2

Explanation

Retrieving and reranking a smaller candidate set can reduce unnecessary context while preserving the most relevant information. The retrieval pipeline can use metadata filtering, similarity thresholds, appropriate top-k values, and reranking to select useful passages before generation. Sending 20 passages for every query may increase token consumption and latency when many results are irrelevant. Adding duplicates worsens context efficiency, while increasing output limits affects generated content rather than retrieval. Retrieval changes should be evaluated for both efficiency and answer quality to ensure important evidence is not removed.

Question 311

Which AWS service provides centralized management of cryptographic keys that can be used to encrypt sensitive application data?

1. AWS KMS
2. Amazon API Gateway
3. Amazon CloudWatch
4. Amazon DynamoDB Streams

Correct Answer: 1

Explanation

AWS Key Management Service provides centralized management of cryptographic keys used to protect data. AWS services and applications can integrate with KMS to encrypt sensitive information while controlling who or what can use the relevant keys. IAM policies and key policies can be used together to restrict key operations. CloudWatch provides monitoring, API Gateway manages APIs, and DynamoDB Streams captures database changes. KMS is particularly relevant to generative AI workloads that process confidential, regulated, or proprietary information requiring encryption controls.

Question 312

A developer wants to determine whether a new prompt version produces better results than the previous version. Which testing approach provides the strongest comparison?

1. Compare results using the same evaluation dataset and criteria
2. Test each version with unrelated datasets
3. Change the model for every prompt version
4. Evaluate only one example

Correct Answer: 1

Explanation

Using the same evaluation dataset and consistent scoring criteria provides a controlled comparison between prompt versions. Keeping other important variables constant allows developers to determine whether observed changes are attributable primarily to the prompt modification. Evaluation can measure relevance, groundedness, accuracy, safety, latency, or cost depending on application requirements. Testing unrelated datasets or changing the model at the same time introduces additional variables and makes conclusions less reliable. A representative evaluation set should contain the types of queries and edge cases expected in production.

Question 313

A developer wants to monitor whether model invocation latency exceeds an acceptable threshold. Which capability should be configured?

1. S3 object versioning
2. CloudWatch alarm
3. KMS key rotation
4. DynamoDB Streams

Correct Answer: 2

Explanation

A CloudWatch alarm can monitor a relevant metric and generate an alert when a configured threshold is exceeded. For an AI application, metrics may include model invocation latency, errors, throttling, request counts, and application-specific performance indicators. Alerts allow teams to investigate problems before they significantly affect users. S3 versioning protects object versions, KMS manages encryption keys, and DynamoDB Streams captures database changes. Monitoring thresholds should be based on expected workload behavior and should be reviewed as the application’s performance requirements evolve.

Question 314

A company needs to protect a production database from AI-generated write operations because the application only requires read access. Which control provides the strongest independent protection?

1. A longer system prompt
2. More few-shot examples
3. A larger model
4. A read-only database identity

Correct Answer: 4

Explanation

A read-only database identity creates an independent authorization boundary that prevents write operations even if the AI model generates an unsafe or unexpected query. This is stronger than relying only on prompt instructions because the database itself enforces the permission restriction. Additional SQL validation can provide another layer of protection by checking generated statements before execution. Developers should apply least privilege to database credentials and avoid giving an AI workload permissions it does not require. Sensitive AI-generated actions should always be constrained by trusted backend controls.

Question 315

A developer wants to ensure that new documents uploaded to Amazon S3 are eventually available for RAG retrieval. Which pipeline is most appropriate?

1. Upload files and never update the index
2. Automatically ingest, embed, and index new documents
3. Store documents only in prompts
4. Manually rewrite the foundation model

Correct Answer: 2

Explanation

An automated ingestion pipeline can detect newly uploaded documents, extract their content, divide them into appropriate chunks, generate embeddings, and update the retrieval index. This allows the RAG system to incorporate new information without requiring manual intervention or model retraining. Monitoring should verify that each stage completes successfully and that the index reflects the expected source-data version. Storing entire documents directly in prompts is inefficient and does not scale well. Foundation models generally do not need to be retrained whenever a knowledge source changes.

Question 316

A developer wants to protect an AI API from excessive request traffic while maintaining access for legitimate clients. Which capability should be used?

1. API throttling
2. Larger vector dimensions
3. Higher temperature
4. More document chunks

Correct Answer: 1

Explanation

API throttling controls the rate at which clients can send requests to an API. It can help protect backend resources, manage service quotas, and prevent individual clients from overwhelming the application. Amazon API Gateway provides API-management capabilities that can support controlled request traffic. Throttling should be combined with authentication, authorization, monitoring, and appropriate retry behavior. Increasing vector dimensions affects embeddings, temperature affects model generation, and document chunking affects retrieval. None of those directly controls incoming API traffic.

Question 317

A developer wants to test whether an AI application can resist attempts to bypass its safety restrictions. Which testing method is most appropriate?

1. Adversarial testing
2. S3 lifecycle testing only
3. UI layout testing
4. Database backup testing

Correct Answer: 1

Explanation

Adversarial testing intentionally exposes an AI system to malicious or manipulative inputs to identify weaknesses in safety and security controls. Tests can include attempts to bypass guardrails, inject instructions, extract sensitive information, manipulate tools, or cause unauthorized actions. Findings can be used to improve prompts, guardrails, authorization, validation, and monitoring. Conventional functional testing alone may not reveal AI-specific vulnerabilities. Adversarial testing should be repeated whenever important changes are made to models, prompts, retrieval systems, tools, or application architecture.

Question 318

A RAG system uses an embedding model that was replaced with a newer model. What should the developer consider before using the existing vector index with the new model?

1. The compatibility of existing embeddings and query embeddings
2. The CloudTrail retention period
3. The API Gateway domain name
4. The S3 storage class

Correct Answer: 1

Explanation

Embedding models can produce vector representations with different characteristics and dimensions. Existing document embeddings may therefore not be directly compatible with query embeddings generated by a new model. Developers should verify compatibility and, when necessary, re-embed source documents and rebuild the vector index. Retrieval quality should then be evaluated using a representative benchmark. CloudTrail retention, API Gateway domain names, and S3 storage classes do not determine whether vectors from different embedding models can be meaningfully compared. Embedding consistency is fundamental to reliable semantic retrieval.

Question 319

A company discovers that a smaller foundation model provides acceptable quality for simple requests while a larger model performs better on complex requests. Which architecture can optimize this workload?

1. Use the larger model for every request
2. Use model routing based on task requirements
3. Disable evaluation
4. Randomly select a model

Correct Answer: 2

Explanation

Model routing can direct requests to different foundation models according to task complexity, quality requirements, cost, or latency objectives. Simple tasks can potentially use a lower-cost model when its quality is sufficient, while complex tasks can be directed to a more capable model. This can reduce overall inference costs without sacrificing quality where higher capability is required. Routing decisions should be validated through representative evaluation datasets and monitored in production. Random selection provides no reliable quality or cost strategy, while always using the largest model may be unnecessarily expensive.

Question 320

A production AI application experiences higher latency after traffic increases significantly. Which investigation is most appropriate?

1. Increase the prompt length
2. Change the model temperature
3. Analyze concurrency, throttling, resource utilization, and downstream dependencies
4. Remove monitoring to reduce overhead

Correct Answer: 3

Explanation

A significant traffic increase can expose bottlenecks in model invocation, retrieval services, APIs, compute resources, databases, or downstream dependencies. Developers should examine concurrency, request rates, throttling, resource utilization, queueing, and dependency latency to identify the actual bottleneck. CloudWatch metrics and application telemetry can provide evidence for this investigation. Increasing prompt length can worsen latency, while temperature does not address infrastructure capacity. Removing monitoring makes diagnosis more difficult. Performance optimization should be based on measured bottlenecks rather than assumptions about which component is responsible.