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

 

Question 21

Which technique converts text into numerical vectors for semantic search?

1. Tokenization
2. Embedding
3. Encryption
4. Compression

Correct Answer: 2

Explanation

An embedding represents text as a numerical vector that captures aspects of its semantic meaning. Embeddings allow applications to compare the similarity between queries and stored content based on meaning rather than exact keyword matches. They are commonly used in retrieval-augmented generation applications, recommendation systems, and semantic search solutions. After documents are converted into embeddings, they can be stored in a vector database or supported search system. When a user submits a query, its embedding can be compared with stored vectors to retrieve relevant information.

Question 22

A developer is designing a customer-service chatbot that must answer questions using private company documents. What architecture is most suitable?

1. RAG
2. DNS
3. CDN
4. Batch processing only

Correct Answer: 1

Explanation

Retrieval-Augmented Generation (RAG) is suitable when a generative AI application needs to use private or frequently changing information. The system retrieves relevant information from a company knowledge base and provides it to the foundation model as context. This allows the model to generate responses based on organizational data without requiring the model itself to be retrained whenever documents change. RAG can also improve grounding and reduce unsupported answers when the retrieved content is relevant, current, and authoritative.

Question 23

What is the primary purpose of Amazon Bedrock model evaluation?

1. To create DNS records
2. To store application passwords
3. To compare model performance against defined criteria
4. To manage EC2 operating systems

Correct Answer: 3

Explanation

Model evaluation helps developers determine whether a foundation model is appropriate for a particular application. Evaluation can measure criteria such as accuracy, relevance, robustness, toxicity, latency, and other application-specific requirements. Teams can use representative datasets and consistent evaluation methods to compare different models or configurations. This is especially important before production deployment because the model that performs best in one scenario may not be the best choice for another. Evaluation provides measurable evidence for model selection and improvement.

Question 24

A developer wants to reduce hallucinations in an application that answers questions about company policies. Which approach is generally most effective?

1. Increase temperature
2. Remove the knowledge base
3. Increase unrelated context
4. Ground responses with trusted retrieved information

Correct Answer: 4

Explanation

Grounding a model with trusted, relevant information can reduce the likelihood of unsupported responses. In a company-policy application, a retrieval system can locate the appropriate policy documents and provide them as context to the foundation model. The prompt can also instruct the model to answer using the supplied sources and avoid inventing information when the answer is unavailable. Increasing temperature generally introduces more variation, while unrelated context can create confusion. Reliable retrieval and grounding are therefore important techniques for improving factual responses.

Question 25

What does temperature primarily influence in a generative AI model?

1. The randomness and variability of generated output
2. The number of IAM users
3. The size of an S3 bucket
4. The network bandwidth

Correct Answer: 1

Explanation

Temperature is a generation parameter that influences how deterministic or varied model responses are. Lower values generally encourage more predictable outputs, while higher values can produce greater variation and creativity. The appropriate setting depends on the application. A factual extraction task may benefit from lower variability, whereas creative content generation may tolerate or benefit from more variation. Temperature does not control AWS permissions, storage capacity, or network bandwidth. Developers should evaluate the parameter using representative prompts rather than assuming one value is ideal for every use case.

Question 26

Which AWS service can be used as a managed vector search capability for generative AI workloads?

1. Amazon Route 53
2. Amazon OpenSearch Service
3. AWS CloudTrail
4. Amazon SNS

Correct Answer: 2

Explanation

Amazon OpenSearch Service can support vector search workloads and is commonly used in retrieval architectures that require semantic similarity searches. Documents can be processed into embeddings and indexed so that applications can retrieve content based on vector similarity. This makes OpenSearch useful for RAG and other AI-powered search scenarios. Route 53 provides DNS services, CloudTrail records AWS API activity, and SNS provides messaging and notifications. The appropriate vector-search architecture depends on application requirements, scale, integration needs, and operational considerations.

Question 27

A team needs to provide a foundation model with only the most relevant portions of a large document. What should it use?

1. Chunking and retrieval
2. Full-document duplication
3. DNS caching
4. Image compression

Correct Answer: 1

Explanation

Chunking divides large documents into smaller sections, while retrieval identifies the sections most relevant to a user’s query. Together, these techniques help provide focused context to a foundation model. This can improve retrieval quality and reduce unnecessary token usage compared with sending an entire document for every request. Chunking strategies should consider document structure and semantic boundaries rather than simply splitting text at arbitrary lengths. Retrieval can then select appropriate chunks for the model’s context during each application interaction.

Question 28

Why should prompts be versioned in a production generative AI application?

1. To increase DNS resolution speed
2. To make prompts impossible to modify
3. To track changes and reproduce application behavior
4. To eliminate model evaluation

Correct Answer: 3

Explanation

Prompt versioning helps development teams track changes to instructions, templates, variables, and formatting requirements over time. When application behavior changes, teams can identify which prompt version was responsible and reproduce earlier results when necessary. Versioning also supports controlled testing, rollback, collaboration, and evaluation across prompt iterations. Without version control, small prompt changes can be difficult to identify and may make production issues harder to troubleshoot. It is therefore an important engineering practice for maintaining reliable generative AI applications.

Question 29

A developer needs to prevent a model from discussing certain prohibited topics. Which capability should be considered?

1. Amazon S3
2. Amazon Bedrock Guardrails
3. Amazon Route 53
4. AWS Direct Connect

Correct Answer: 2

Explanation

Amazon Bedrock Guardrails can help developers configure safeguards for generative AI applications. Organizations can define policies that restrict certain content categories or topics and help control model interactions according to application requirements. This can be useful for customer-facing applications where responses must comply with organizational policies. Guardrails should be considered alongside application-level validation, authorization, monitoring, and other security controls. S3, Route 53, and Direct Connect provide storage, DNS, and networking capabilities respectively and do not provide the same model-safety functionality.

Question 30

Which practice helps protect an application’s AWS credentials?

1. Hard-code credentials in prompts
2. Store credentials in source code
3. Share credentials among all developers
4. Use IAM roles and managed secret-storage mechanisms

Correct Answer: 4

Explanation

IAM roles and managed secret-storage mechanisms provide safer alternatives to embedding credentials directly in application code or prompts. IAM roles can provide temporary permissions to AWS resources, reducing the need for long-lived access keys. Services such as AWS Secrets Manager can securely store sensitive application secrets when they are required. Hard-coded credentials can accidentally be exposed through repositories, logs, or deployment packages. Applying least privilege and using secure credential-management practices helps reduce unauthorized access and supports stronger security for production AI applications.

Question 31

What is prompt injection?

1. A method of compressing prompts
2. An attack that attempts to manipulate an AI system through crafted instructions
3. A database backup technique
4. A method of increasing GPU memory

Correct Answer: 2

Explanation

Prompt injection occurs when an attacker or untrusted content attempts to influence a generative AI system by supplying instructions that conflict with the application’s intended behavior. This can happen through direct user prompts or indirectly through retrieved documents, web pages, or other external content. Developers should treat external data as potentially untrusted and use appropriate input controls, permissions, validation, guardrails, and output checks. Secure architecture is particularly important when an AI application can access confidential information or invoke external tools.

Question 32

A RAG system retrieves relevant documents, but the generated answer still contains incorrect information. What should the developer investigate first?

1. Retrieval quality and prompt grounding
2. DNS TTL values
3. S3 storage class
4. Network cable length

Correct Answer: 1

Explanation

When relevant documents are retrieved but answers remain incorrect, developers should examine whether the retrieved content is actually sufficient, whether the prompt clearly instructs the model to use that context, and whether the retrieved passages contain conflicting or outdated information. Retrieval ranking, chunking, metadata filters, and context ordering can also affect results. The model may still generate unsupported content if grounding instructions are weak or if retrieved information is ambiguous. Evaluating the complete retrieval-to-generation pipeline is therefore important.

Question 33

Which metric is especially useful when assessing how quickly a generative AI application responds to users?

1. Storage capacity
2. Latency
3. DNS record count
4. Number of IAM groups

Correct Answer: 2

Explanation

Latency measures the time required for an operation to produce a response. In generative AI applications, latency can affect user experience and may depend on model selection, prompt size, retrieved context, network conditions, workload, and output length. Developers should measure latency using realistic workloads rather than relying only on theoretical performance. Monitoring latency alongside quality and cost helps teams select an appropriate architecture. A highly capable model may not be suitable for an interactive application if its response time consistently exceeds business requirements.

Question 34

A developer wants an AI application to return only information supported by retrieved documents. Which instruction is useful?

1. Ask the model to invent missing details
2. Tell the model to ignore retrieved context
3. Instruct the model to use the provided context and state when information is unavailable
4. Remove all source documents

Correct Answer: 3

Explanation

A grounding instruction can tell the model to base its response on the supplied context and avoid inventing information that is not supported. It can also instruct the model to clearly state when the available sources do not contain an answer. This does not completely eliminate hallucinations, so applications should combine prompting with retrieval-quality improvements, validation, evaluation, and other safeguards. Ignoring retrieved information or encouraging invented details works against factual reliability. Clear instructions can therefore strengthen the behavior of knowledge-grounded applications.

Question 35

What is fine-tuning primarily used for?

1. Teaching a model application-specific patterns or behaviors using additional training data
2. Creating DNS records
3. Increasing S3 storage
4. Monitoring CloudWatch alarms

Correct Answer: 1

Explanation

Fine-tuning adapts a supported foundation model using additional training examples so that it can better perform particular tasks, follow desired patterns, or reflect specialized requirements. It differs from RAG, where external information is retrieved at inference time and supplied as context. Fine-tuning can be useful when consistent behavior or task-specific patterns are needed, but it should not automatically be treated as a replacement for retrieval of frequently changing factual information. Developers should evaluate whether prompting, RAG, fine-tuning, or a combination best meets the use case.

Question 36

A developer wants to test whether a new prompt improves answer quality without immediately replacing the production prompt. What should the developer do?

1. Delete the existing prompt
2. Test the new prompt against a representative evaluation dataset
3. Disable monitoring
4. Increase all model parameters randomly

Correct Answer: 2

Explanation

Testing a new prompt against a representative evaluation dataset allows developers to compare the new behavior with the existing implementation before production deployment. Evaluation can measure quality, factuality, relevance, formatting compliance, latency, and other application-specific criteria. This controlled process reduces the risk of introducing regressions that may not be obvious from a few manual examples. Production systems should use versioning and controlled rollout practices when appropriate. Prompt changes should therefore be evaluated systematically rather than adopted solely because a few responses appear better.

Question 37

Which AWS service is commonly used to store documents that may serve as source data for a generative AI knowledge base?

1. Amazon S3
2. Amazon Route 53
3. AWS CloudTrail
4. Amazon EventBridge Scheduler

Correct Answer: 1

Explanation

Amazon S3 provides highly durable object storage and can store documents such as PDFs, text files, and other source material used by generative AI applications. These objects can serve as input to data-processing or knowledge-base workflows before content is chunked, embedded, and indexed for retrieval. S3 can also integrate with many AWS services and application architectures. Route 53 handles DNS, CloudTrail provides auditing, and EventBridge Scheduler is designed for scheduled tasks. S3 is therefore a common choice for source-document storage.

Question 38

An AI agent can call a financial transaction API. What is the most important security principle when granting the agent access?

1. Give it unrestricted permissions
2. Use least privilege and restrict available actions
3. Disable authentication
4. Store API keys in the system prompt

Correct Answer: 2

Explanation

Agents that can perform consequential actions should receive only the permissions necessary to complete their intended tasks. Least privilege limits the potential impact of incorrect model decisions, compromised inputs, or unexpected tool calls. Developers should also consider authorization checks, input validation, transaction limits, human approval for high-impact operations, logging, and monitoring. Giving an agent unrestricted access creates unnecessary risk. API credentials should be securely managed rather than placed in prompts. Strong permission boundaries are essential for safe agentic applications.

Question 39

Which option best describes a foundation model?

1. A model trained broadly that can be adapted or prompted for many tasks
2. A DNS management system
3. A storage encryption key
4. A network routing table

Correct Answer: 1

Explanation

A foundation model is a broadly trained model that can support a wide range of tasks through prompting, adaptation, or other techniques. Generative foundation models can perform activities such as text generation, summarization, classification, question answering, and other supported tasks. Developers can select a model based on capabilities, cost, latency, context requirements, modality, and application needs. A foundation model is fundamentally different from infrastructure components such as routing tables, encryption keys, or DNS services because it provides learned capabilities for AI workloads.

Question 40

A production application needs to detect abnormal increases in model invocation errors. What should the team configure?

1. An S3 bucket policy only
2. A DNS alias
3. Monitoring metrics and alarms
4. A larger prompt

Correct Answer: 3

Explanation

Monitoring metrics and alarms can help teams detect abnormal increases in application errors and respond quickly to operational problems. For a generative AI workload, useful signals may include invocation failures, latency, throttling, request volume, and other service or application metrics. Alert thresholds should reflect normal workload behavior so that teams avoid excessive false alarms. Monitoring should be combined with logging and appropriate incident-response procedures. A larger prompt or DNS configuration does not provide the required operational visibility into model invocation failures.