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

 

Question 241

A developer wants to improve the quality of retrieved results by considering both exact terms and semantic meaning. Which approach should be used?

1. Hybrid search
2. Random retrieval
3. Image classification
4. Log aggregation

Correct Answer: 1

Explanation

Hybrid search combines lexical or keyword-based retrieval with semantic vector search. This approach is useful when users may search using either exact identifiers or natural-language descriptions. Keyword search can effectively find product codes, names, and exact phrases, while semantic search can identify conceptually similar content even when different terminology is used. Combining both approaches can improve retrieval coverage and relevance. Developers should evaluate weighting and ranking behavior using representative queries to determine whether hybrid retrieval provides better results than either method independently.

Question 242

A company wants to prevent an AI application from answering questions about prohibited subjects defined by its business policy. Which capability is most appropriate?

1. Amazon S3 versioning
2. Bedrock Guardrails topic policies
3. AWS CloudFormation
4. DynamoDB Streams

Correct Answer: 2

Explanation

Bedrock Guardrails can be configured to help restrict specific topics and control the type of content generated or processed by a generative AI application. Topic policies are useful when an organization has clearly defined subjects that the application should refuse to discuss. Guardrails should be combined with application-level authorization, input validation, and other security measures rather than being treated as the only protection. S3 versioning manages object versions, CloudFormation provisions infrastructure, and DynamoDB Streams captures database changes. Guardrails directly address generative AI content controls.

Question 243

An organization has a large collection of PDFs containing tables, headings, and multiple sections. What should the developer prioritize when preparing this content for RAG?

1. Increasing model temperature
2. Removing all metadata
3. Preserving meaningful document structure during ingestion
4. Disabling document validation

Correct Answer: 3

Explanation

Preserving meaningful document structure can improve the quality of information available to a RAG system. During ingestion, developers should consider how headings, sections, tables, and related content are extracted and divided into chunks. Poor parsing can cause important relationships to be lost before embeddings are generated. Depending on the document type, appropriate parsing or extraction methods may be required. Temperature does not improve document ingestion, and removing metadata can reduce useful filtering capabilities. Document validation should remain part of a reliable ingestion pipeline.

Question 244

A developer wants to ensure that a Lambda function can invoke a specific foundation model but cannot access unrelated AWS services. Which security practice should be followed?

1. Grant AdministratorAccess
2. Use a dedicated IAM role with least-privilege permissions
3. Share the developer’s personal credentials
4. Give the function access to every model

Correct Answer: 2

Explanation

A dedicated IAM role with least-privilege permissions allows a Lambda function to perform only the operations required by its workload. The role can grant model invocation permissions for the specific resources needed and restrict access to unrelated services. This reduces the potential impact if the function, its dependencies, or credentials are compromised. Administrator permissions are unnecessarily broad, while personal credentials should never be embedded into workloads. Access should be reviewed regularly as application requirements evolve and additional capabilities are introduced.

Question 245

A RAG system has a very large vector index and retrieval latency has increased significantly. Which optimization should the developer investigate?

1. Retrieval index configuration and search strategy
2. Increasing prompt temperature
3. Adding more unrelated context
4. Disabling vector search

Correct Answer: 1

Explanation

Large vector indexes can introduce retrieval latency depending on index configuration, query patterns, filtering, and search strategy. Developers should evaluate the vector index structure, approximate nearest-neighbor configuration, metadata filtering, top-k values, and overall retrieval architecture. Reducing unnecessary candidates before generation can also improve performance. Increasing temperature affects generation rather than vector retrieval. Adding unrelated context can increase downstream latency and cost, while disabling vector search may remove the application’s primary semantic retrieval mechanism. Performance changes should be validated using realistic workloads.

Question 246

A developer needs to store documents that will later be processed by an asynchronous AI pipeline. Which AWS service is commonly suitable for durable object storage?

1. AWS KMS
2. Amazon API Gateway
3. Amazon S3
4. Amazon CloudWatch

Correct Answer: 3

Explanation

Amazon S3 provides highly durable object storage and is commonly used to store documents, datasets, generated artifacts, and other files used by AI applications. New objects can also trigger event-driven processing workflows, allowing downstream services to extract content, generate embeddings, or perform other processing tasks. KMS provides encryption-key management, API Gateway exposes APIs, and CloudWatch provides monitoring. S3 can therefore serve as a durable source repository within an asynchronous generative AI data pipeline.

Question 247

A developer wants to make an AI workflow automatically process a newly uploaded file without continuously polling for changes. Which architecture is preferable?

1. Event-driven processing
2. Manual processing
3. Periodic spreadsheet updates
4. Static prompt execution

Correct Answer: 1

Explanation

Event-driven processing allows an application to react automatically when an event occurs, such as a new object being uploaded to Amazon S3. The event can trigger Lambda, a workflow, or another processing component without requiring constant polling. This can reduce unnecessary resource usage and provide faster response to new data. For generative AI applications, the triggered workflow might extract text, validate content, create embeddings, and update a retrieval index. Event-driven designs are particularly useful when data arrives unpredictably or continuously.

Question 248

A model generates excessively long responses even though users only need concise answers. Which change can directly help control the response length?

1. Increase the context window
2. Increase temperature
3. Remove output instructions
4. Set an appropriate maximum output-token limit

Correct Answer: 4

Explanation

A maximum output-token limit can restrict how much content the model generates. This can help control response length, reduce unnecessary token consumption, and potentially lower latency and inference costs. The prompt should also clearly communicate the desired response style and level of detail. Increasing the context window does not directly limit generated output, while increasing temperature can increase variability rather than shorten responses. Developers should choose a token limit that provides enough capacity for useful answers without allowing unnecessarily large responses.

Question 249

A company is comparing two AI models for an application where response quality, latency, and cost are all important. What should the evaluation include?

1. Only the model names
2. Quality, latency, and cost measurements
3. Only maximum context length
4. Only developer preference

Correct Answer: 2

Explanation

Model selection should consider the requirements of the actual workload rather than relying on a single characteristic. Developers should evaluate response quality using representative tasks while also measuring latency, token consumption, throughput, and cost. Safety and reliability can also be included where relevant. A model with slightly better quality may not be the best choice if its additional cost or latency is unacceptable for the application’s requirements. Controlled evaluation provides objective evidence for selecting the most appropriate foundation model.

Question 250

An agent repeatedly calls the same tool because it fails to recognize that the required operation has already completed. Which safeguard can reduce this behavior?

1. Remove all tool descriptions
2. Increase tool permissions
3. Implement tool-call limits and state tracking
4. Disable logging

Correct Answer: 3

Explanation

Tool-call limits and explicit state tracking can help prevent agents from repeatedly executing the same operation. The application can record completed actions, reject duplicate requests where appropriate, and impose maximum tool-call counts for a workflow. This reduces unnecessary costs and protects downstream systems from repeated operations. Tool descriptions should remain clear, while broadening permissions does not solve the underlying problem. Logging should also remain enabled so developers can investigate why an agent is entering repetitive loops.

Question 251

A developer needs to keep sensitive credentials separate from application source code while allowing an application to retrieve them at runtime. Which service is most appropriate?

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

Correct Answer: 1

Explanation

AWS Secrets Manager is designed to securely store and retrieve sensitive information such as passwords, API keys, and other application credentials. Applications can obtain secrets at runtime while IAM controls which identities or workloads are allowed to access them. This is safer than embedding credentials directly into source code or configuration files. Secrets can also be rotated according to organizational requirements. CloudWatch provides monitoring, OpenSearch supports search workloads, and Step Functions orchestrates workflows rather than serving as the primary secret-storage service.

Question 252

A RAG application retrieves highly relevant information, but the model sometimes ignores that information and produces unsupported claims. What should the developer improve?

1. Remove the retrieved context
2. Use stronger grounding instructions and evaluate generated responses
3. Increase the number of unrelated documents
4. Disable retrieval evaluation

Correct Answer: 2

Explanation

Strong retrieval alone does not guarantee that a model will use the retrieved evidence correctly. The developer should provide clear grounding instructions that tell the model to base answers on the supplied context and acknowledge when sufficient evidence is unavailable. Generated responses should then be evaluated for groundedness and unsupported claims. Retrieval quality, context formatting, and prompt structure should also be reviewed. Adding unrelated documents can make grounding more difficult, while removing context eliminates the evidence needed to support accurate responses.

Question 253

A company wants to maintain different prompt versions and determine which version produced a particular evaluation result. Which practice is most useful?

1. Version and track prompt templates
2. Randomly change prompts in production
3. Delete previous prompt versions
4. Store prompts only in temporary files

Correct Answer: 1

Explanation

Prompt versioning allows developers to track changes and associate specific prompt templates with evaluation results. This makes it easier to determine which version produced a particular behavior and enables controlled rollback when a change causes regressions. Versioned prompts can also be tested systematically against consistent datasets before production deployment. Random production changes make troubleshooting difficult, while deleting previous versions prevents comparison and rollback. Prompt management is particularly important when multiple applications or teams depend on shared templates.

Question 254

An AI application must perform an expensive operation only after receiving approval from an authorized user. Which design is most appropriate?

1. Allow the model to execute it automatically
2. Add an explicit human approval step
3. Give the model administrator permissions
4. Remove authentication

Correct Answer: 2

Explanation

Sensitive or irreversible actions may require human approval before execution. A human-in-the-loop workflow can allow the AI system to prepare a proposed action while requiring an authorized person to review and approve it. This provides an additional safety boundary for high-impact operations such as financial transactions, account changes, or destructive actions. Model confidence alone should not be treated as authorization. Human approval should be combined with identity verification, authorization, input validation, and audit logging to create a stronger control framework.

Question 255

A developer observes that a model’s responses vary significantly between identical requests. Which parameter should be reviewed first?

1. Temperature and other generation settings
2. S3 bucket encryption
3. IAM policy names
4. CloudTrail event retention

Correct Answer: 1

Explanation

Generation parameters such as temperature can influence the variability of model outputs. If an application requires highly consistent responses, developers should review temperature and other supported inference settings and determine whether more deterministic configuration is appropriate. The exact behavior depends on the foundation model and task. Developers should test settings using representative requests rather than assuming that one configuration is universally optimal. S3 encryption, IAM naming, and CloudTrail retention are important infrastructure controls but do not directly control generation variability.

Question 256

A developer wants to identify which AWS API calls were made by resources in a production AI environment for auditing purposes. Which service should be used?

1. Amazon S3
2. AWS CloudTrail
3. Amazon Bedrock Guardrails
4. AWS Lambda

Correct Answer: 2

Explanation

AWS CloudTrail records AWS API activity and can help organizations understand which identities or services performed actions within an AWS environment. These audit records support security investigations, compliance requirements, and operational troubleshooting. For generative AI applications, CloudTrail can complement application-level logs and model-related monitoring. S3 is object storage, Guardrails provides AI safety controls, and Lambda provides compute execution. Audit logging should be configured with appropriate retention, access controls, and protection against unauthorized modification.

Question 257

A RAG application uses a similarity threshold that is too high, causing relevant documents to be excluded from retrieval. Which problem does this create?

1. Lower retrieval recall
2. Higher encryption strength
3. Improved API authentication
4. Increased storage durability

Correct Answer: 1

Explanation

A similarity threshold determines how closely a retrieved item must match a query before it is accepted. If the threshold is set excessively high, documents that are relevant but not extremely similar may be excluded. This reduces retrieval recall and can leave the model without important evidence. Developers should tune retrieval thresholds using representative evaluation queries and consider other factors such as reranking, metadata filtering, and top-k configuration. Encryption, authentication, and storage durability are unrelated to the retrieval threshold.

Question 258

A company wants to reduce the impact of a failed model dependency by providing users with a simpler alternative response. Which reliability technique is this?

1. Prompt expansion
2. Vector normalization
3. Graceful degradation
4. Dataset duplication

Correct Answer: 3

Explanation

Graceful degradation allows an application to continue providing useful functionality when a dependency becomes unavailable or performs poorly. For example, an AI application might provide a predefined response, use a simpler model, return previously validated information, or disable a nonessential AI feature temporarily. This approach can improve user experience and system resilience during partial failures. Prompt expansion and vector normalization address different concerns, while dataset duplication does not itself provide a runtime fallback strategy. Fallback behavior should be tested before production deployment.

Question 259

A developer wants to test whether an AI application can be manipulated by malicious instructions embedded in user input. What type of testing should be performed?

1. Adversarial testing
2. Storage testing only
3. UI color testing
4. File compression testing

Correct Answer: 1

Explanation

Adversarial testing intentionally exposes an AI application to malicious, unexpected, or manipulative inputs to identify weaknesses. For generative AI systems, tests can include prompt injection, attempts to bypass restrictions, sensitive-data extraction, tool misuse, and other abuse scenarios. Results can be used to improve prompts, guardrails, authorization controls, validation, and monitoring. Traditional functional testing remains important, but it may not reveal security weaknesses specific to AI interactions. Adversarial testing should be repeated as models, prompts, tools, and application workflows change.

Question 260

A developer needs to expose an AI backend through an API while limiting excessive request traffic from clients. Which capability should be considered?

1. API throttling
2. Increasing model temperature
3. Removing authentication
4. Increasing document chunk size

Correct Answer: 1

Explanation

API throttling limits the rate at which clients can send requests and helps protect backend services from excessive traffic. It can improve stability, prevent uncontrolled resource consumption, and help manage service quotas. Amazon API Gateway provides capabilities that can support controlled API traffic and request management. Throttling should be configured according to expected workload patterns and combined with authentication, authorization, monitoring, and appropriate retry behavior. Temperature affects model generation, while document chunk size affects retrieval rather than API traffic management.