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

 

Question 181

Which component converts text into numerical representations that can be used for semantic similarity searches?

1. IAM policy
2. Embedding model
3. API Gateway
4. CloudTrail

Correct Answer: 2

Explanation

An embedding model converts text or other supported data into numerical vectors that represent semantic meaning. These vectors can then be stored in a vector database or search service and compared with query embeddings to identify semantically similar content. Embeddings are a fundamental component of many RAG architectures because they allow applications to retrieve relevant information even when the wording between a query and source document is different. IAM manages access, API Gateway manages APIs, and CloudTrail records activity rather than generating semantic representations.

Question 182

A developer is designing a RAG application for technical documentation. Some documents contain important product codes that must match exactly, while other queries depend on semantic meaning. Which retrieval approach is most appropriate?

1. Keyword search only
2. Semantic search only
3. Hybrid search
4. Random document selection

Correct Answer: 3

Explanation

Hybrid search combines keyword-based retrieval with semantic or vector-based retrieval. This is useful when an application needs both exact matching and meaning-based matching. Product codes, model numbers, and identifiers may require precise keyword matching, while natural-language questions can benefit from semantic similarity. Combining the two approaches can provide broader and more accurate retrieval than relying on either method alone. Random selection provides no relevance guarantee, while using only one retrieval method may miss important results depending on how users phrase their questions.

Question 183

What is a primary purpose of chunking documents in a RAG pipeline?

1. To divide large content into retrievable sections
2. To encrypt documents
3. To authenticate users
4. To increase API request limits

Correct Answer: 1

Explanation

Chunking divides large documents into smaller sections that can be independently indexed and retrieved. Well-designed chunks allow the retrieval system to identify specific portions of a document that are relevant to a user’s query instead of returning an entire large document. Chunk size and overlap should be selected according to the structure and meaning of the source material. Poor chunking can separate related information or create excessive fragments. Encryption, authentication, and API limits are separate concerns and are not the primary purpose of document chunking.

Question 184

A company wants to prevent unauthorized users from invoking a generative AI API. Which control should be implemented at the application or API layer?

1. Authentication and authorization
2. Higher model temperature
3. Larger document chunks
4. More prompt examples

Correct Answer: 1

Explanation

Authentication verifies the identity of an API caller, while authorization determines what that authenticated identity is allowed to do. These controls should be enforced before sensitive AI operations are performed. Depending on the architecture, Amazon API Gateway and AWS identity services can help implement appropriate access controls. Prompt instructions cannot reliably prevent unauthorized users from reaching an API because the model operates after the request has already entered the application. Proper authentication, authorization, and least-privilege permissions should therefore form the foundation of API security.

Question 185

A developer wants to make an AI assistant answer only when sufficient information exists in the retrieved context. What should the prompt instruct the model to do?

1. Always provide an answer
2. State when the available context is insufficient
3. Invent missing information
4. Ignore retrieved documents

Correct Answer: 2

Explanation

Instructing the model to acknowledge when the retrieved context is insufficient can reduce unsupported responses. The prompt should clearly state that the model should use the provided evidence, avoid inventing facts, and indicate when the available information does not support a reliable answer. This approach is especially useful in RAG applications where factual grounding is important. Requiring the model to always answer can encourage hallucinations when evidence is missing. Ignoring retrieval removes valuable information, while inventing information directly conflicts with trustworthy AI practices.

Question 186

A generative AI application stores customer documents in Amazon S3. The application should automatically process a document whenever it is uploaded. Which architecture is appropriate?

1. Event-driven processing triggered by the upload
2. Manual processing once per year
3. Increasing model temperature
4. Storing the document inside a prompt

Correct Answer: 1

Explanation

An event-driven architecture can automatically initiate processing when a new object is uploaded to Amazon S3. An S3 event can trigger downstream processing such as an AWS Lambda function or another workflow component. The processing pipeline could extract text, validate the document, generate embeddings, and update a retrieval index. This design reduces manual intervention and allows new information to become available to the AI application automatically. Temperature does not affect document processing, and embedding a complete document directly into a prompt is inefficient.

Question 187

A company wants to evaluate whether an AI application’s responses are both relevant to the question and factually supported. Which combination of evaluation criteria is most useful?

1. Relevance and groundedness
2. Storage capacity and disk size
3. IAM users and roles
4. API endpoint names

Correct Answer: 1

Explanation

Relevance measures whether the response appropriately addresses the user’s question, while groundedness evaluates whether the response is supported by the information provided to the model. Both are important for RAG and enterprise generative AI applications. A response can be relevant but still contain unsupported claims, so evaluating only one dimension may not reveal all quality problems. Storage capacity, IAM configuration, and endpoint naming are important infrastructure concerns but do not directly measure the quality or factual support of generated responses.

Question 188

A developer wants to reduce hallucinations in an enterprise RAG application. Which combination is most effective?

1. Relevant retrieval, grounded prompts, and evaluation
2. Higher temperature and larger prompts
3. Fewer security controls
4. Unrestricted tool access

Correct Answer: 1

Explanation

Reducing hallucinations generally requires improvements across retrieval, generation, and evaluation. The application should retrieve authoritative and relevant information, instruct the model to remain grounded in that information, and evaluate outputs for unsupported claims. Improving chunking, filtering, reranking, and source quality can further strengthen retrieval. Increasing temperature may make responses more variable and does not inherently improve factuality. Unrestricted tool access and weaker security controls can introduce additional risks. A layered approach provides stronger reliability than relying on a single prompt instruction.

Question 189

Which AWS service is primarily responsible for securely storing application secrets such as API keys and passwords?

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 store and manage sensitive information such as passwords, API keys, and other credentials. Applications can retrieve secrets securely at runtime instead of embedding them directly in source code or configuration files. Access can be controlled through IAM policies, and secret rotation can be implemented where supported. CloudWatch focuses on monitoring, OpenSearch provides search capabilities, and Step Functions orchestrates workflows. Separating credentials from application code is an important security practice for generative AI applications and their supporting services.

Question 190

A company wants to encrypt sensitive data stored by its generative AI application and maintain control over the encryption keys. Which AWS service should it consider?

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

Correct Answer: 1

Explanation

AWS Key Management Service provides centralized management of cryptographic keys used to protect data. AWS services such as Amazon S3 can integrate with KMS to encrypt stored information while allowing organizations to control key access and monitor key usage. This is particularly important for applications processing confidential or regulated data. Lambda provides compute, API Gateway manages APIs, and CloudWatch provides monitoring. KMS should be combined with IAM permissions and appropriate encryption configurations to ensure that only authorized users and services can access protected information.

Question 191

A developer notices that model responses are too random for a classification task. Which inference parameter should the developer review first?

1. Temperature
2. Vector dimension
3. Chunk overlap
4. API timeout

Correct Answer: 1

Explanation

Temperature controls the degree of randomness or variability in many foundation model generation processes. For tasks such as classification, structured extraction, or deterministic responses, a lower temperature may produce more consistent outputs, depending on the model’s supported behavior. However, developers should still evaluate the results using representative test cases because parameter effects can vary between models. Vector dimensions affect embeddings, chunk overlap affects document retrieval, and API timeout controls request duration. None of these directly controls generation randomness in the same way as temperature.

Question 192

A developer wants to ensure that a model does not exceed a specified response size. Which setting should be considered?

1. Maximum output tokens
2. Vector dimensions
3. S3 lifecycle policy
4. IAM policy

Correct Answer: 1

Explanation

Maximum output tokens can limit how much content a foundation model generates in a response. Setting an appropriate output limit helps control response length, latency, and potentially inference costs. The value should be large enough to support the application’s expected responses while preventing unnecessarily long outputs. Vector dimensions affect embeddings, S3 lifecycle policies manage stored objects, and IAM policies control access permissions. Output limits should be evaluated together with prompt design and application requirements to ensure that useful responses are not unintentionally truncated.

Question 193

A company wants to make sure an AI agent cannot access customer records belonging to another business unit. Which control should be enforced independently of the model’s instructions?

1. Authorization controls
2. Higher temperature
3. More few-shot examples
4. Larger context windows

Correct Answer: 1

Explanation

Authorization must be enforced through trusted application and infrastructure controls rather than relying solely on model instructions. The application should verify the user’s identity and permissions before allowing the agent to retrieve specific customer records. Backend services should also enforce access restrictions so that even an incorrectly generated tool request cannot bypass authorization. Prompt instructions can guide the model but should never be considered a security boundary. Least-privilege IAM permissions, application authorization, and resource-level controls can work together to protect sensitive records.

Question 194

A developer is testing a prompt change and wants to determine whether response quality improved without changing the model or evaluation data. What should remain constant?

1. The evaluation conditions
2. The user’s identity
3. The application logo
4. The S3 storage class

Correct Answer: 1

Explanation

Controlled evaluation requires keeping relevant testing conditions constant so that changes in results can be attributed to the prompt modification. The same model, evaluation dataset, inference settings, and scoring criteria should generally be used when comparing prompt versions. Changing multiple variables simultaneously makes it difficult to determine which change affected the results. The goal is to isolate the prompt as the primary experimental variable. Application branding and S3 storage classes are unrelated to prompt-quality evaluation and should not influence the comparison.

Question 195

Which approach can help protect an AI application from prompt injection attempts against an agent that can access tools?

1. Validate tool inputs and enforce authorization
2. Give the agent administrator permissions
3. Disable authentication
4. Allow every tool without restrictions

Correct Answer: 1

Explanation

Prompt injection can attempt to manipulate an AI system into performing actions that were not intended by the application. Tool inputs should therefore be validated, and authorization should be enforced independently by trusted application components and backend services. Agents should receive only the permissions required for their approved tasks. Giving administrator permissions or unrestricted tool access increases the potential impact of a successful attack. Developers should also consider guardrails, monitoring, confirmation requirements for sensitive actions, and other defense-in-depth measures.

Question 196

A RAG application returns relevant documents but still produces poor answers because each chunk contains incomplete information. What should the developer investigate?

1. Chunk size and overlap
2. IAM password length
3. CloudTrail retention
4. API Gateway domain name

Correct Answer: 1

Explanation

Chunk size and overlap directly influence how much contextual information is available to the retrieval system. If chunks are too small, important relationships or sentences may be separated across different chunks, causing the model to receive incomplete evidence. Appropriate overlap can preserve context between neighboring chunks, while larger meaningful chunks may keep related information together. The optimal configuration depends on document structure and the retrieval workload. IAM password settings, CloudTrail retention, and API domain names do not directly address incomplete retrieved context.

Question 197

A company wants to use different foundation models depending on request complexity. Simple requests should use a lower-cost model, while difficult requests should use a more capable model. What architecture can support this strategy?

1. Model routing
2. Static prompt duplication
3. Larger vector dimensions
4. Disabling evaluation

Correct Answer: 1

Explanation

Model routing allows an application to select different foundation models according to request characteristics or business requirements. Simple tasks can potentially be handled by lower-cost models, while complex tasks can be directed to more capable models. Routing decisions may consider task type, complexity, latency requirements, cost constraints, or previous evaluation results. This strategy can improve overall efficiency while maintaining required quality. Static prompts and vector dimensions do not provide model-selection logic, and disabling evaluation would make it difficult to validate routing decisions.

Question 198

A developer wants to detect whether a new AI application version causes more errors than the previous version. Which operational practice is most useful?

1. Compare monitored error metrics before and after deployment
2. Remove all application logs
3. Increase prompt temperature
4. Delete the previous version immediately

Correct Answer: 1

Explanation

Comparing monitored error metrics before and after deployment helps identify whether a new version introduces regressions. Developers can monitor error rates, latency, model invocation failures, retrieval failures, and other application-specific indicators. A baseline from the previous version provides useful context for determining whether changes are significant. Removing logs eliminates diagnostic evidence, while temperature is unrelated to most deployment-level error comparisons. Keeping the previous version available also supports rollback if the new release demonstrates unacceptable behavior.

Question 199

A company wants to improve the reliability of an AI application that depends on several external services. Which design practice is most appropriate?

1. Use timeouts, retries, and failure handling
2. Remove monitoring
3. Allow unlimited retries
4. Ignore downstream failures

Correct Answer: 1

Explanation

Timeouts, retries, and failure-handling mechanisms help applications respond predictably when external dependencies experience temporary failures or slow responses. Timeouts prevent requests from waiting indefinitely, while controlled retries can recover from transient errors. Failure-handling logic can provide fallback behavior or return an appropriate response when recovery is not possible. Unlimited retries should be avoided because they can increase costs and overload downstream services. Monitoring is also important for identifying recurring dependency problems and measuring application reliability over time.

Question 200

A developer is preparing a generative AI application for production. Testing shows acceptable response quality, but the application has high latency, inconsistent outputs, and excessive cost. What should the developer do before deployment?

1. Deploy immediately because quality is acceptable
2. Optimize and re-evaluate performance, consistency, and cost
3. Increase temperature
4. Remove monitoring to reduce overhead

Correct Answer: 2

Explanation

Production readiness requires more than acceptable response quality. The developer should investigate latency, output consistency, token usage, model selection, retrieval efficiency, and infrastructure performance before deployment. Changes should then be tested against representative workloads to confirm that improvements do not negatively affect quality or safety. Increasing temperature may worsen consistency, while removing monitoring reduces visibility into production behavior. A balanced evaluation of quality, performance, reliability, security, and cost helps ensure that the application is ready to operate effectively at production scale.