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

 

Question 321

A company is building a customer-support assistant that must answer questions using frequently changing product documentation. The company does not want to retrain the foundation model whenever documentation changes. Which architecture is most appropriate?

1. Retrieval-Augmented Generation (RAG)
2. Full model retraining for every document update
3. Increasing the model temperature
4. Removing the knowledge source

Correct Answer: 1

Explanation

Retrieval-Augmented Generation allows an application to retrieve current information from an external knowledge source and provide that information to a foundation model during inference. This is well suited for frequently changing documentation because the underlying model does not need to be retrained whenever source content changes. A typical RAG pipeline includes document ingestion, chunking, embedding generation, vector storage, retrieval, and grounded generation. This approach can improve freshness while avoiding the operational cost and delay associated with repeatedly retraining or customizing a model.

Question 322

An AI agent must create a support ticket only after receiving confirmation from a human operator. Which design best enforces this requirement?

1. Allow the model to create tickets automatically
2. Add a human approval step before the action is executed
3. Increase the agent’s permissions
4. Store the approval instruction in the system prompt only

Correct Answer: 2

Explanation

A human approval step provides an explicit control boundary before a potentially consequential action is performed. The agent can prepare the proposed ticket and present its details to an authorized human, who can approve or reject the operation. This approach is stronger than relying only on model instructions because the approval requirement is enforced by application workflow logic. Permissions should still follow least-privilege principles. Human-in-the-loop controls are particularly useful for actions involving financial transactions, account changes, sensitive communications, or other operations where incorrect autonomous execution could cause significant impact.

Question 323

A developer is designing an AI agent tool that updates customer records. Which tool characteristic is most important for preventing accidental duplicate updates when requests are retried?

1. Idempotent operation design
2. Higher temperature
3. Larger embedding vectors
4. Longer system prompts

Correct Answer: 1

Explanation

An idempotent operation produces the same intended result when the same request is processed multiple times. This is especially important for agent tools because network failures, timeouts, or orchestration retries can cause an operation to be submitted again. An update API can use an idempotency key or another mechanism to detect duplicate requests. This prevents retries from unintentionally creating duplicate records or repeating side effects. Model temperature and embedding dimensions do not provide protection against duplicate tool execution. Idempotency should be implemented in trusted application infrastructure.

Question 324

A developer needs to ensure that confidential information is not unnecessarily included in prompts sent to a foundation model. Which practice is most appropriate?

1. Increase the number of retrieved documents
2. Apply data minimization before inference
3. Disable authentication
4. Include the complete customer database

Correct Answer: 2

Explanation

Data minimization means providing only the information required for the model to complete the requested task. Before inference, applications should remove unnecessary personal, confidential, or sensitive information and restrict retrieved context to relevant content. This reduces privacy exposure, lowers token consumption, and can improve response quality by eliminating irrelevant context. Sending entire databases into prompts creates unnecessary risk and cost. Data minimization should complement access controls, encryption, logging policies, and other security mechanisms rather than replacing them.

Question 325

A RAG application retrieves relevant documents, but users complain that answers contain information from outdated versions of those documents. What should the developer investigate first?

1. Model temperature
2. Index freshness and document synchronization
3. UI font size
4. API response formatting

Correct Answer: 2

Explanation

When outdated information is retrieved, the developer should investigate whether the knowledge base or vector index is synchronized with the current source documents. Problems may occur when updated files are not ingested, old vectors remain active, deletion events are not propagated, or synchronization jobs fail. The retrieval pipeline should track document versions or timestamps where appropriate and verify that obsolete content is removed or superseded. Changing temperature does not fix stale retrieval. Maintaining index freshness is essential for RAG applications whose source information changes regularly.

Question 326

A developer wants to allow an AI agent to access an internal service without exposing the service’s credentials to the model. Which design is most appropriate?

1. Put the credentials inside the prompt
2. Ask the model to memorize the credentials
3. Use a trusted backend tool that securely obtains credentials
4. Include credentials in tool descriptions

Correct Answer: 3

Explanation

The model should never receive sensitive credentials merely to invoke an authorized service. Instead, the application can expose a controlled tool or backend service that authenticates securely using IAM roles, Secrets Manager, or another appropriate mechanism. The model provides structured parameters, while trusted infrastructure performs authentication and authorization. This separates model-generated instructions from security-sensitive operations. Putting credentials in prompts, tool descriptions, or model context can expose secrets through logs, outputs, or unintended model behavior and violates secure credential-management practices.

Question 327

An organization wants to compare two foundation models using the same set of business questions before selecting one for production. Which evaluation method is most appropriate?

1. Benchmark both models with the same representative evaluation dataset
2. Test each model with different questions
3. Select the model with the largest parameter count
4. Compare only model names

Correct Answer: 1

Explanation

A controlled model evaluation should use the same representative dataset and consistent scoring criteria for each candidate model. This makes it possible to compare factors such as accuracy, relevance, groundedness, safety, latency, and cost under similar conditions. Using different questions can introduce dataset bias and make the comparison unreliable. Parameter count alone does not establish that a model will perform better for a specific business workload. Evaluation results should reflect actual application requirements and include important edge cases before a production decision is made.

Question 328

A company wants to prevent an AI application from returning personally identifiable information found in retrieved documents unless the user is authorized to see it. Which control should be prioritized?

1. Authorization-aware retrieval and access controls
2. Higher model temperature
3. Larger output limits
4. Random document selection

Correct Answer: 1

Explanation

Authorization should be enforced before sensitive information reaches the model whenever possible. Retrieval systems can apply user or role-based access controls so that users retrieve only documents they are permitted to access. This reduces the possibility that the model will expose unauthorized information through its response. Prompt instructions and output filtering can provide additional protection but should not replace authoritative access controls. Developers should also consider document-level metadata, identity propagation, audit logging, and data classification when implementing retrieval systems that handle sensitive information.

Question 329

A developer is creating an AI application that accepts images along with user questions. Which consideration is especially important when designing the inference workflow?

1. Ensure the selected model supports the required multimodal input
2. Disable input validation
3. Convert every image into an unrelated text prompt
4. Use only a text-only model

Correct Answer: 1

Explanation

Multimodal applications require a model and inference interface that support the required input types, such as images and text. Developers should verify supported formats, image size limits, token or input constraints, preprocessing requirements, and the expected request structure. Input validation is also important because uploaded content may be malformed or unsafe. A text-only model cannot directly provide the same multimodal capabilities. The application should also monitor latency and cost because image processing can affect inference resources differently from ordinary text requests.

Question 330

A team notices that an AI agent repeatedly calls the same tool without reaching a conclusion. What should the developer add to the agent workflow?

1. An unlimited execution loop
2. A termination condition and maximum iteration limit
3. More tool permissions
4. A larger temperature value

Correct Answer: 2

Explanation

Agent workflows should include explicit termination conditions to prevent uncontrolled planning or tool-call loops. A maximum iteration or tool-call limit provides a safety boundary when an agent cannot reach a valid conclusion. Additional conditions can detect successful completion, repeated failures, invalid tool responses, or lack of progress. Unlimited loops can increase latency, cost, and the risk of unintended actions. Temperature does not reliably solve orchestration problems. Agent workflows should also provide appropriate error handling and monitoring so repeated execution patterns can be diagnosed.

Question 331

A developer needs to maintain an audit trail showing who changed an AWS resource configuration used by an AI application. Which service is most appropriate?

1. Amazon S3
2. AWS CloudTrail
3. Amazon DynamoDB
4. Amazon Bedrock

Correct Answer: 2

Explanation

AWS CloudTrail records AWS API activity and can provide information about actions performed against supported AWS resources. This makes it useful for auditing configuration changes, identifying the identity associated with an operation, and investigating security events. CloudTrail records can support governance and compliance requirements when configured appropriately. S3 is an object storage service, DynamoDB is a NoSQL database, and Bedrock provides managed generative AI capabilities. Organizations should also establish suitable log retention, access controls, and monitoring processes around audit data.

Question 332

A RAG application needs to restrict search results to documents belonging to the user’s department before semantic ranking occurs. Which technique is most appropriate?

1. Metadata filtering
2. Increasing temperature
3. Increasing output tokens
4. Random sampling

Correct Answer: 1

Explanation

Metadata filtering allows a retrieval system to restrict candidate documents according to attributes such as department, region, document type, tenant, or access classification. Applying appropriate filters before or during retrieval can improve both security and relevance by preventing unrelated or unauthorized documents from entering the candidate set. Semantic similarity alone may retrieve content that is conceptually relevant but not permitted for the user. Metadata should therefore be maintained accurately during ingestion and associated with each indexed document or chunk.

Question 333

A company deploys a new prompt version to only a small percentage of users while keeping the previous version active for everyone else. What deployment strategy does this represent?

1. Full replacement
2. Batch deletion
3. Canary deployment
4. Database replication

Correct Answer: 3

Explanation

Canary deployment introduces a new version to a small subset of traffic while the existing version continues serving the majority of users. Developers can compare metrics such as response quality, latency, error rates, safety incidents, and cost before expanding the rollout. This reduces the impact of unexpected problems and provides an opportunity to stop or roll back the release. Canary strategies can be applied to prompts, models, application code, retrieval logic, or other AI components where controlled production validation is beneficial.

Question 334

A developer wants to reduce repeated identical requests to an AI application and avoid unnecessary model invocations. Which optimization can help?

1. Response caching
2. Increasing temperature
3. Adding more retrieved documents
4. Disabling request validation

Correct Answer: 1

Explanation

Response caching can avoid repeated model invocations when requests are identical or sufficiently equivalent and the cached result remains valid. This can reduce latency, inference costs, and backend workload. Cache design should consider expiration, invalidation, personalization, authorization, and whether the underlying knowledge source has changed. A cached response must not be returned to a user who is not authorized to see its contents. Increasing temperature does not reduce repeated work, and adding unnecessary retrieval context generally increases processing requirements.

Question 335

A developer is troubleshooting an AI application where the model frequently produces malformed JSON despite being instructed to return structured output. What should be investigated?

1. Output constraints and structured-response support
2. S3 lifecycle configuration only
3. CloudTrail retention only
4. Database backup frequency

Correct Answer: 1

Explanation

Malformed structured output can result from insufficient output constraints, ambiguous prompts, unsupported response-format features, or model behavior. Developers should determine whether the selected model and inference interface support structured responses or schema-based output constraints. The application should also validate the response before parsing it and implement safe handling for invalid output. Strongly defining the expected schema can improve reliability. Storage lifecycle policies, audit-log retention, and database backups do not directly address malformed model-generated JSON.

Question 336

An organization needs an event-driven process that starts an AI workflow whenever a specific application event occurs. Which AWS capability is appropriate?

1. Amazon EventBridge
2. AWS KMS
3. Amazon Macie only
4. Amazon Route 53

Correct Answer: 1

Explanation

Amazon EventBridge can route events from AWS services and applications to appropriate targets, enabling event-driven architectures. An AI workflow can be triggered when an event such as a document upload, customer action, or processing completion occurs. Event-driven designs can reduce unnecessary polling and allow components to operate independently. KMS provides encryption key management, Macie focuses on data discovery and security-related analysis, and Route 53 provides DNS services. Event routing should include appropriate filtering, retries, permissions, and failure-handling mechanisms.

Question 337

A developer is deciding between fine-tuning a model and using RAG for a frequently updated knowledge base. Which option is generally more appropriate for keeping factual knowledge current?

1. RAG
2. Fine-tune after every document change
3. Increase temperature
4. Remove external knowledge

Correct Answer: 1

Explanation

RAG is generally more suitable when an application needs access to frequently changing factual information. Instead of modifying model parameters whenever information changes, the application retrieves current content from an external knowledge source at inference time. Fine-tuning can be useful for adapting behavior, style, or task performance, but repeatedly fine-tuning for every knowledge update can create unnecessary operational overhead. RAG also allows source documents to be updated independently of the foundation model, provided that ingestion and indexing pipelines keep the retrieval system current.

Question 338

A production AI service receives many requests at the same time and begins returning throttling errors. Which response is most appropriate?

1. Ignore the errors
2. Increase request concurrency without limits
3. Analyze service quotas and implement controlled concurrency
4. Increase model temperature

Correct Answer: 3

Explanation

Throttling indicates that a service or resource has reached an applicable request or concurrency limit. Developers should examine service quotas, current request rates, concurrency, and workload patterns before deciding on remediation. Controlled concurrency, queueing, backoff, and appropriate retry policies can prevent clients from overwhelming downstream services. Where appropriate, quota increases can also be requested. Unlimited concurrency can make throttling worse and create cascading failures. Temperature affects generation behavior and does not address service capacity or request-rate limitations.

Question 339

A company wants to ensure that an AI application continues functioning when its primary model temporarily becomes unavailable. Which architecture is most appropriate?

1. A single model with no error handling
2. A fallback model or controlled degradation path
3. Unlimited retries against the same failed endpoint
4. Removing monitoring

Correct Answer: 2

Explanation

A fallback model or controlled degradation path can maintain useful application behavior when the primary model experiences an outage, quota issue, or other temporary failure. The fallback should be selected according to acceptable quality, cost, latency, and capability requirements. The application should distinguish retryable failures from permanent errors and avoid unlimited retries that can increase load during an outage. Monitoring should track fallback usage because frequent fallback activation may indicate an underlying reliability problem that requires investigation.

Question 340

A company wants to detect whether an AI application’s answers are relevant, grounded, and safe before releasing a major model update. Which approach is strongest?

1. Evaluate the new version against a representative test suite using defined quality and safety metrics
2. Test only one question
3. Release immediately and wait for complaints
4. Select the model based only on cost

Correct Answer: 1

Explanation

A representative evaluation suite provides a controlled way to assess an AI system before production release. The test suite should contain realistic user requests, difficult cases, safety scenarios, and expected edge conditions. Metrics can include relevance, groundedness, factuality, toxicity, refusal behavior, latency, and cost depending on application requirements. Comparing results against an existing baseline helps identify regressions. Testing only one question cannot represent production behavior, while cost alone does not establish whether a model meets quality or safety requirements.