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

 

Question 81

Which AWS service is designed to provide access to foundation models for generative AI applications?

1. Amazon Bedrock
2. Amazon Route 53
3. AWS Shield
4. Amazon EFS

Correct Answer: 1

Explanation

Amazon Bedrock is a managed AWS service that provides access to supported foundation models through APIs. Developers can use these models to create applications for content generation, summarization, question answering, classification, and other generative AI workloads. Bedrock also provides capabilities that support retrieval, agents, guardrails, model customization, and evaluation. Because AWS manages the underlying model infrastructure, developers can focus more on application logic and integration. Route 53, Shield, and EFS serve DNS, security, and file-storage purposes rather than providing foundation model access.

Question 82

Which AWS service can provide a managed search capability for vector embeddings?

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

Correct Answer: 2

Explanation

Amazon OpenSearch Service can support vector search and is commonly used in generative AI architectures that require semantic retrieval. Embeddings can be indexed and searched based on vector similarity, allowing applications to retrieve documents that are conceptually related to a user’s query. This makes OpenSearch useful for RAG and semantic-search workloads. The service can also support other search patterns depending on the architecture. SNS is a messaging service, Lambda provides serverless compute, and Route 53 handles DNS management.

Question 83

A developer is creating a chatbot that must remember information from the current conversation. What capability is needed?

1. Conversation state or memory management
2. DNS routing
3. Image compression
4. Database encryption only

Correct Answer: 1

Explanation

A conversational application needs a way to maintain relevant state so that later requests can use information from earlier interactions. Depending on the architecture, this may involve storing conversation history, summarizing previous exchanges, or maintaining application-specific memory. Developers should carefully manage how much information is retained because excessive history can increase token usage and potentially expose unnecessary sensitive data. Conversation state should also be separated from long-term knowledge retrieval when appropriate. Effective state management helps create more coherent multi-turn interactions.

Question 84

What is one reason to use metadata filters during RAG retrieval?

1. To restrict results to relevant attributes such as department or document date
2. To increase model temperature
3. To create IAM passwords
4. To remove embeddings

Correct Answer: 1

Explanation

Metadata filters allow a retrieval system to restrict search results according to attributes associated with documents. For example, an application might retrieve only documents belonging to a particular department, product, region, language, or date range. Filtering can reduce irrelevant results and improve the quality of context provided to the model. Metadata filtering can be combined with vector or keyword search for stronger retrieval. It is particularly useful when a knowledge base contains many documents but only a specific subset is relevant to a user’s request.

Question 85

A developer wants to compare generated answers against known expected results. What is this activity called?

1. Evaluation
2. Tokenization
3. Chunking
4. Indexing

Correct Answer: 1

Explanation

Evaluation is the process of measuring how well a generative AI system performs against defined criteria or expected outcomes. Developers can compare model outputs with reference answers, human judgments, automated metrics, or application-specific requirements. Evaluation can measure dimensions such as accuracy, relevance, factuality, safety, consistency, and formatting. Because generative outputs may vary, evaluation should use representative datasets rather than relying on a few examples. Regular evaluation helps teams identify regressions and select appropriate prompts, models, and retrieval configurations.

Question 86

Which approach is most appropriate when a model needs current information that changes frequently?

1. RAG
2. Repeated manual retraining for every change
3. DNS caching
4. Static prompts without external data

Correct Answer: 1

Explanation

RAG allows an application to retrieve current information from an external knowledge source during inference. This is useful when data changes frequently because the knowledge base can be updated independently of the foundation model. For example, product catalogs, company policies, or support documentation can be updated and reindexed without retraining the model for every factual change. Developers should still monitor retrieval quality and data freshness. RAG is therefore often preferable to repeatedly modifying the model when the primary requirement is access to changing information.

Question 87

What does a higher temperature generally do to model output?

1. Makes output more varied
2. Guarantees accuracy
3. Reduces token count to zero
4. Disables safety controls

Correct Answer: 1

Explanation

A higher temperature generally increases variability in generated responses. This can be useful for creative applications where different wording or ideas are desirable. However, higher variability may be less appropriate for tasks requiring highly consistent outputs, such as structured extraction or deterministic classification. Temperature does not guarantee accuracy, remove token usage, or disable safety mechanisms. Developers should test temperature settings against the application’s actual requirements and evaluate whether increased variation improves or harms the desired user experience.

Question 88

A developer needs to make a foundation model follow a specific response format. What should the prompt include?

1. Clear output requirements
2. Unrelated documents
3. Random instructions
4. No constraints

Correct Answer: 1

Explanation

A prompt should clearly describe the expected response format when an application requires predictable output. Developers can specify required fields, ordering, data types, length constraints, allowed values, or formatting rules. Examples can also be included when they improve understanding. Clear output requirements make model responses easier to validate and process by downstream applications. However, generated output should still be validated before being used in sensitive workflows. Explicit instructions are generally more effective than assuming the model will infer the application’s required structure.

Question 89

Which technique allows a model to receive several examples before performing a task?

1. Few-shot prompting
2. Fine-grained IAM
3. Vector compression
4. Data replication

Correct Answer: 1

Explanation

Few-shot prompting provides examples of inputs and desired outputs within the prompt. These examples can help the model infer task patterns, classification rules, formatting requirements, or desired response styles. Developers should choose examples that accurately represent the expected task and avoid unnecessary examples that consume context. Few-shot prompting does not modify the model’s underlying parameters. If examples do not provide sufficient improvement, developers may consider better instructions, retrieval, model selection, or customization depending on the application’s needs.

Question 90

A team wants to reduce the risk that users can make an AI agent perform unauthorized actions. What should be implemented?

1. Authorization and least-privilege permissions
2. Unlimited tool access
3. Shared administrator credentials
4. No authentication

Correct Answer: 1

Explanation

AI agents that can call external tools should operate under carefully controlled authorization and least-privilege permissions. The application should determine which users can request particular actions and which tools the agent is allowed to invoke. Sensitive operations may require additional validation or human approval. Shared administrator credentials and unrestricted access increase the potential impact of malicious prompts, compromised accounts, or incorrect model decisions. Authentication, authorization, tool restrictions, logging, and monitoring should work together to protect agentic workflows.

Question 91

Which component converts a user’s text query into a vector for semantic search?

1. Embedding model
2. Load balancer
3. IAM role
4. DNS resolver

Correct Answer: 1

Explanation

An embedding model converts text into a numerical representation that captures semantic characteristics. In a semantic-search workflow, the user’s query is transformed into an embedding and compared with embeddings stored for documents or other content. The system can then identify vectors that are similar to the query and retrieve the corresponding information. This process is fundamental to many RAG architectures. An IAM role manages permissions, a DNS resolver handles domain-name resolution, and a load balancer distributes traffic rather than generating semantic representations.

Question 92

A RAG system retrieves outdated documents even though newer versions exist. Which improvement is most appropriate?

1. Add document version or freshness metadata to retrieval
2. Increase temperature
3. Remove all metadata
4. Increase the number of old documents retrieved

Correct Answer: 1

Explanation

Document freshness and version information can help a retrieval system prioritize current content. Metadata such as publication date, revision date, version number, status, or effective date can be used to filter or rank documents. This is especially important for policies, product information, procedures, and other content that changes over time. Increasing temperature affects generation rather than retrieval, while removing metadata can make freshness filtering more difficult. Retrieval evaluation should confirm that the revised strategy consistently selects the correct and current source material.

Question 93

What is the purpose of a system prompt?

1. To provide high-level instructions that guide model behavior
2. To store AWS access keys
3. To create a vector index
4. To configure DNS

Correct Answer: 1

Explanation

A system prompt can provide high-level instructions, behavioral requirements, role definitions, constraints, and other guidance for how a model should respond within an application. Developers can use it to establish consistent behavior across user requests. System instructions should be carefully designed because user input and retrieved content may introduce conflicting or untrusted instructions. Sensitive credentials should never be placed in prompts. Effective system prompts are part of a broader application architecture that should also include security controls, validation, monitoring, and evaluation.

Question 94

A developer wants to reduce unnecessary information sent to a foundation model from a knowledge base. What should be optimized?

1. Retrieval relevance
2. DNS records
3. IAM usernames
4. Network cables

Correct Answer: 1

Explanation

Retrieval relevance determines how effectively a RAG system selects information that actually supports the user’s request. Developers can improve relevance through better chunking, embeddings, search configuration, metadata filtering, query transformation, reranking, and appropriate retrieval limits. Reducing irrelevant context can lower token usage and may improve answer quality by reducing distractions. Retrieval should be evaluated using representative queries and known relevant documents. Optimizing relevance is therefore an important part of improving both the efficiency and reliability of knowledge-grounded AI applications.

Question 95

Which AWS service is commonly used to securely store application secrets?

1. AWS Secrets Manager
2. Amazon CloudFront
3. Amazon Route 53
4. Amazon Athena

Correct Answer: 1

Explanation

AWS Secrets Manager provides secure storage and management for sensitive information such as passwords, API keys, and other application secrets. Applications can retrieve secrets programmatically instead of embedding them directly in source code or configuration files. This helps reduce accidental exposure and supports secure credential-management practices. Access to secrets should still be controlled using appropriate IAM permissions. CloudFront provides content delivery, Route 53 handles DNS, and Athena supports querying data. Secrets Manager is therefore appropriate for protecting credentials required by generative AI applications.

Question 96

A developer is building an AI application that must process images as well as text. What should be considered when selecting a model?

1. Whether the model supports the required modalities
2. Only the model’s name
3. DNS performance
4. IAM group size

Correct Answer: 1

Explanation

Model selection should account for the modalities required by the application. If an application needs to process images and text, the selected foundation model must support the necessary inputs and outputs. Developers should also evaluate quality, latency, cost, context limits, supported features, and other requirements. Choosing a model solely based on its name or popularity can result in incompatibility with the workload. Multimodal applications should be tested using representative content to ensure that the selected model performs adequately for the intended use cases.

Question 97

What is hallucination in a generative AI application?

1. A confident but unsupported or incorrect generated response
2. A successful database backup
3. A DNS failure
4. A model deployment process

Correct Answer: 1

Explanation

A hallucination occurs when a generative AI model produces information that is unsupported, inaccurate, or fabricated while presenting it as though it were valid. Hallucinations can result from insufficient context, ambiguous prompts, model limitations, or retrieval problems. Developers can reduce the risk through grounding, high-quality retrieval, clear instructions, output validation, evaluation, and appropriate guardrails. No single technique guarantees that hallucinations will never occur. Applications should therefore be designed to identify and manage incorrect model behavior, particularly in high-impact use cases.

Question 98

A developer wants to determine whether an AI application is financially practical at scale. Which metric should be monitored?

1. Cost per request or workload
2. DNS record count
3. Number of screen pixels
4. IAM policy length

Correct Answer: 1

Explanation

Cost per request or workload helps developers understand how generative AI expenses may change as usage grows. Costs can be affected by model choice, input and output tokens, retrieval operations, storage, infrastructure, and request volume. Teams should evaluate cost using realistic production workloads rather than relying only on small proof-of-concept usage. Cost should be considered alongside quality and latency because the cheapest solution may not meet business requirements. Monitoring unit economics helps organizations determine whether an AI application can operate sustainably at scale.

Question 99

Which practice helps detect regressions after changing a model or prompt?

1. Re-running a representative evaluation dataset
2. Deleting previous results
3. Disabling monitoring
4. Avoiding testing

Correct Answer: 1

Explanation

Re-running a representative evaluation dataset allows developers to compare the new version with previous results and identify changes in application behavior. Metrics can include factuality, relevance, classification accuracy, safety, formatting compliance, latency, and cost. This is especially important because changing a model or prompt can improve one type of response while unintentionally degrading another. Maintaining versioned evaluation datasets makes comparisons easier and supports controlled development. Continuous evaluation is therefore an important practice for maintaining reliable generative AI applications.

Question 100

A company is preparing a generative AI application for production. Which approach provides the strongest overall design?

1. Focus only on model quality
2. Ignore cost because AI systems are unpredictable
3. Combine model evaluation, security, monitoring, retrieval quality, cost controls, and testing
4. Deploy immediately after a successful demonstration

Correct Answer: 3

Explanation

A production generative AI application requires more than a capable foundation model. Teams should evaluate model quality, retrieval performance, security, privacy, latency, scalability, cost, monitoring, and failure handling. Testing should cover representative, edge-case, and potentially adversarial inputs. Access controls and guardrails should protect users and resources, while evaluation should detect regressions after changes. A successful demonstration only proves that a concept can work under limited conditions. A comprehensive engineering approach provides stronger evidence that the application can operate reliably and safely in production.