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

 

Question 161

Which prompting technique asks a model to solve a task without providing examples?

1. Few-shot prompting
2. Zero-shot prompting
3. Retrieval augmentation
4. Fine-tuning

Correct Answer: 2

Explanation

Zero-shot prompting asks a foundation model to perform a task using instructions without providing task-specific examples. It is useful when the task can be clearly described through natural-language instructions and examples are unnecessary. For example, a developer can instruct a model to classify customer feedback as positive, negative, or neutral without including sample classifications. Few-shot prompting provides examples, while retrieval augmentation supplies external context. Fine-tuning changes model behavior through training rather than simply providing instructions during inference.

Question 162

A company wants to prevent an AI assistant from answering questions about topics that are outside the assistant’s approved business purpose. Which capability should the developer consider?

1. Topic restrictions through guardrails
2. Increasing the model context window
3. Increasing vector dimensions
4. Disabling logging

Correct Answer: 1

Explanation

Guardrails can help organizations restrict specific topics and control how generative AI applications respond to certain categories of requests. A developer can configure appropriate policies so that an assistant designed for a particular business purpose does not provide responses on prohibited or unrelated subjects. This helps establish boundaries around the application’s intended use. Increasing the context window or vector dimensions does not enforce behavioral restrictions, while disabling logging removes useful operational visibility. Guardrails should be combined with application-level authorization and validation where appropriate.

Question 163

A RAG system retrieves relevant documents, but users complain that answers are still based on outdated versions. What should the developer investigate first?

1. Source-data freshness and indexing
2. Model temperature
3. API response formatting
4. Output token limits

Correct Answer: 1

Explanation

If the retrieval results themselves are outdated, the developer should investigate the source-data ingestion and indexing pipeline. Updated documents may not have been detected, processed, embedded, or indexed correctly. The system should have a reliable mechanism for identifying document changes and refreshing the retrieval index. Metadata such as document version and modification date can also help track freshness. Changing temperature or output limits does not update retrieved knowledge. The first priority should therefore be ensuring that the retrieval layer reflects the latest authoritative source information.

Question 164

A developer needs to store embeddings and perform similarity searches for a large collection of enterprise documents. Which service is appropriate for this requirement?

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

Correct Answer: 2

Explanation

Amazon OpenSearch Service can support vector search and similarity-based retrieval for applications such as RAG. Document content can be processed into embeddings and stored in a vector-capable index, allowing queries to retrieve semantically similar content. This makes OpenSearch useful when a generative AI application requires scalable search and retrieval capabilities. CloudWatch focuses on monitoring, CloudTrail records API activity, and Secrets Manager securely stores credentials. Those services do not provide the primary vector-search functionality required by the scenario.

Question 165

A developer is creating a prompt for a model that must classify support tickets into predefined categories. Which instruction is most important?

1. Clearly define the allowed categories and expected output
2. Add unrelated business documents
3. Increase temperature significantly
4. Remove all formatting instructions

Correct Answer: 1

Explanation

Classification prompts should clearly define the available categories and explain the expected output format. This reduces ambiguity and helps the model consistently map incoming requests to the correct class. Developers can also provide examples when appropriate and validate the returned category against the approved list. Adding unrelated documents introduces unnecessary context, while a high temperature can increase output variability. Removing formatting instructions makes downstream processing more difficult. Clear task definitions, constrained outputs, and representative evaluation data can significantly improve classification reliability.

Question 166

A generative AI application must execute a series of dependent tasks, but one task can fail temporarily because of a downstream service timeout. Which workflow capability is especially useful?

1. Automatic retries
2. Larger embeddings
3. Higher temperature
4. Prompt duplication

Correct Answer: 1

Explanation

Automatic retries are useful when a workflow contains temporary failures such as network problems, service timeouts, or transient availability issues. AWS Step Functions supports retry and error-handling mechanisms that can allow failed tasks to be attempted again according to defined rules. Developers can configure retry counts, intervals, and backoff behavior appropriate for the workload. Larger embeddings and temperature do not address infrastructure failures. Prompt duplication also does not provide reliable recovery. Retry strategies should be combined with appropriate timeout and failure-handling policies.

Question 167

A company wants to ensure that generated responses do not contain sensitive customer identifiers before they are displayed to users. What should the application implement?

1. Sensitive-data detection and filtering
2. Higher retrieval depth
3. Increased model temperature
4. Larger output limits

Correct Answer: 1

Explanation

Sensitive-data detection and filtering can identify information such as customer identifiers and prevent it from being unnecessarily exposed in generated responses. Depending on the architecture, this can be supported through guardrails and application-level validation or redaction logic. The developer should also consider protecting sensitive information before it reaches the model when possible. Retrieval depth and output limits do not provide privacy protection, while increasing temperature can make output less predictable. Sensitive-data controls should therefore be part of the application’s broader security and privacy architecture.

Question 168

An application uses a foundation model to generate SQL queries from natural-language requests. What is the most important additional control before executing the generated SQL?

1. Validate and restrict the generated query
2. Increase temperature
3. Add more unrelated examples
4. Remove database authorization

Correct Answer: 3

Explanation

Generated SQL should be treated as untrusted input and validated before execution. The application can restrict permitted operations, validate syntax, limit accessible tables, enforce read-only behavior when appropriate, and apply database permissions independently of the model. This reduces the risk of destructive or unauthorized operations caused by incorrect model output or malicious prompts. Increasing temperature can make generated SQL less predictable. Additional examples do not guarantee safety, and removing database authorization would create significant security risks. Backend controls should never depend solely on model behavior.

Question 169

Which approach can reduce hallucinations when a model must answer questions using enterprise documents?

1. Retrieval-augmented generation
2. Increasing temperature
3. Removing source documents
4. Increasing unrelated prompt content

Correct Answer: 1

Explanation

Retrieval-augmented generation provides the model with relevant information retrieved from trusted external sources at inference time. This gives the model factual context that can be used to formulate answers instead of relying entirely on information encoded in its parameters. Developers should also instruct the model to remain grounded in retrieved content and indicate when sufficient information is unavailable. Increasing temperature generally increases response variability and may worsen unsupported claims. Removing source documents eliminates useful evidence, while unrelated context can increase confusion.

Question 170

A developer wants to determine whether a model’s responses are factually supported by the retrieved documents. Which evaluation dimension is most relevant?

1. Groundedness
2. Network throughput
3. Storage capacity
4. API availability

Correct Answer: 1

Explanation

Groundedness measures whether a generated response is supported by the information available to the model, such as retrieved documents in a RAG application. A grounded response should not introduce unsupported claims or information that cannot be traced to the provided context. Developers can create evaluation datasets containing expected evidence and assess model responses against that evidence. Network throughput, storage capacity, and API availability are important operational factors but do not directly measure whether generated statements are supported by the application’s source information.

Question 171

A company wants to reduce inference costs while maintaining acceptable answer quality. Which optimization should be evaluated?

1. Reduce unnecessary input and output tokens
2. Increase prompt length
3. Retrieve every available document
4. Increase model temperature

Correct Answer: 1

Explanation

Reducing unnecessary input and output tokens can lower inference costs because model usage is commonly influenced by the amount of processed text. Developers can remove redundant instructions, improve retrieval precision, limit irrelevant context, and set appropriate output limits. The goal is to preserve the information necessary for a high-quality response while eliminating unnecessary content. Retrieving every document or increasing prompt length increases token consumption. Temperature primarily influences response variability and is not a direct cost-optimization strategy.

Question 172

A developer needs to make a model invocation available only to a specific application role. Which AWS capability should be used?

1. IAM permissions
2. Prompt examples
3. Vector embeddings
4. Model temperature

Correct Answer: 1

Explanation

AWS Identity and Access Management provides permissions that control which identities can access AWS resources and perform specific actions. A developer can create an IAM policy that allows a particular application role to perform only the required model-related operations. This supports the principle of least privilege and helps prevent unauthorized access. Prompt examples and embeddings influence application behavior and retrieval, while temperature affects model generation. Access control should be enforced through trusted AWS and application security mechanisms rather than relying on instructions given to the model.

Question 173

A company wants to identify which version of a prompt produced a particular production response. What should the application maintain?

1. Prompt version information
2. Higher temperature
3. More vector dimensions
4. Larger response limits

Correct Answer: 1

Explanation

Prompt versioning allows developers to identify which prompt configuration was used for a particular model invocation. This is valuable for debugging, auditing, evaluation, and reproducing production behavior. When prompt templates change frequently, storing a version identifier alongside request metadata can help determine whether a change caused a quality or performance regression. Temperature and response limits describe model configuration but do not identify prompt revisions. Vector dimensions are related to embeddings and retrieval. Version control should therefore be included in the application’s prompt-management process.

Question 174

A developer wants to prevent an AI agent from repeatedly calling the same external tool because of an unexpected model decision. Which control can help?

1. Tool-call limits and execution safeguards
2. Higher temperature
3. More retrieved documents
4. Larger embeddings

Correct Answer: 3

Explanation

Tool-call limits and execution safeguards can restrict how many times an agent can invoke a particular tool or how long an agent workflow can continue. These controls help prevent runaway behavior, unexpected costs, excessive API usage, and repeated operations. Developers can combine limits with authorization, validation, timeouts, and monitoring. Increasing temperature may actually make behavior less predictable, while additional retrieval content does not control tool execution. Agentic applications should therefore include explicit boundaries around tool usage and workflow execution.

Question 175

A company wants to detect performance degradation after changing the retrieval configuration of a RAG application. What should the developer compare?

1. Retrieval and response quality metrics before and after the change
2. Only the application’s logo
3. The number of IAM users
4. S3 storage class names

Correct Answer: 1

Explanation

Retrieval configuration changes should be evaluated using consistent metrics before and after deployment. Useful measurements can include retrieval relevance, recall, groundedness, answer correctness, latency, and cost. Comparing results against a representative evaluation dataset can help determine whether a configuration change improves or harms the application. Looking only at unrelated infrastructure information does not provide evidence about retrieval quality. Controlled evaluation and monitoring allow developers to identify regressions and make informed decisions about retrieval parameters and ranking strategies.

Question 176

A generative AI application receives a large number of requests during unpredictable traffic spikes. Which architectural characteristic can help it scale automatically?

1. Serverless and event-driven components
2. Hard-coded request limits only
3. Larger prompts
4. Manual model selection for every request

Correct Answer: 1

Explanation

Serverless and event-driven architectures can automatically scale components based on workload demand. Services such as AWS Lambda can process requests without requiring developers to provision traditional servers, while event-driven systems can distribute work across scalable processing components. This architecture can help applications handle unpredictable traffic while reducing the need for constant capacity planning. Hard-coded limits may restrict availability rather than improve scalability. Prompt size and manual model selection do not provide infrastructure-level scaling. Appropriate throttling and queueing should also be considered.

Question 177

A developer wants to keep a generative AI application available when a temporary downstream dependency fails. Which design principle is most useful?

1. Graceful degradation and fallback handling
2. Increasing temperature
3. Removing error handling
4. Increasing vector dimensions

Correct Answer: 1

Explanation

Graceful degradation allows an application to continue providing useful functionality when one component is temporarily unavailable. For example, an AI assistant might return a controlled message, use a cached result, switch to an alternative workflow, or retry a transient operation instead of failing completely. This improves resilience and user experience. Error handling, timeouts, retries, and fallback paths should be designed according to the application’s requirements. Temperature and vector dimensions do not provide resilience against service failures, while removing error handling makes failures more disruptive.

Question 178

A company wants to make sure a generative AI application is tested against malicious prompts designed to bypass its safety controls. What type of testing should be included?

1. Adversarial testing
2. Storage testing only
3. UI color testing
4. Capacity testing only

Correct Answer: 1

Explanation

Adversarial testing intentionally uses challenging or malicious inputs to determine whether an AI application can be manipulated into violating its intended behavior or security controls. Examples can include prompt injection, attempts to bypass topic restrictions, malicious tool parameters, and requests designed to expose protected information. Testing should be performed before and after significant application changes and should include appropriate safety metrics. Storage and interface testing may still be necessary for other requirements, but they do not specifically assess an AI application’s resistance to adversarial prompts.

Question 179

A developer wants to monitor the number of failed model requests and receive an alert when failures exceed an acceptable threshold. Which approach is appropriate?

1. CloudWatch metrics and alarms
2. Increasing model temperature
3. Removing application logs
4. Increasing embedding size

Correct Answer: 1

Explanation

Amazon CloudWatch can collect operational metrics and support alarms based on defined thresholds. A generative AI application can publish or monitor relevant failure metrics and create an alarm when the failure rate exceeds an acceptable level. This enables the development or operations team to investigate problems before they significantly affect users. Increasing temperature and embedding size do not provide operational alerting. Removing logs would make troubleshooting more difficult. Metrics and alarms are therefore important components of an AI application’s observability strategy.

Question 180

A company is selecting between two foundation models. Model A provides slightly better quality but has much higher latency and cost than Model B, which meets the application’s quality requirement. Which decision is most appropriate?

1. Always select Model A because it is larger
2. Select Model B after validating it against representative requirements
3. Select Model A without testing
4. Use both models simultaneously for every request

Correct Answer: 2

Explanation

The best model is not necessarily the largest or most capable model. If Model B satisfies the application’s quality, safety, latency, and business requirements at significantly lower cost, it may be the better choice. The decision should be supported by representative evaluation data rather than assumptions about model size. Developers should compare response quality, groundedness, latency, throughput, cost, and safety for the intended workload. Using a more expensive model for every request may waste resources when a smaller model can adequately satisfy the application’s requirements.