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

 

Question 101

Which technique helps a model perform a task by showing examples within the prompt?

1. Zero-shot prompting
2. Few-shot prompting
3. Vector indexing
4. Model monitoring

Correct Answer: 2

Explanation

Few-shot prompting provides examples of expected inputs and outputs directly in the prompt. These examples help the model recognize the desired pattern without requiring changes to the model’s underlying parameters. It can be useful for classification, extraction, formatting, and other specialized tasks. Developers should select examples that accurately represent real application inputs. Too many examples can increase token usage, so prompts should be evaluated for both quality and efficiency. Few-shot prompting is particularly useful when simple instructions alone do not produce consistent results.

Question 102

Which AWS service can be used to store source documents for a generative AI knowledge base?

1. Amazon S3
2. Amazon Route 53
3. AWS CloudTrail
4. Amazon SNS

Correct Answer: 1

Explanation

Amazon S3 provides durable object storage and is commonly used to store documents that serve as source material for generative AI applications. These documents can include PDFs, text files, manuals, policies, and other supported content. A downstream process can read the documents, divide them into chunks, generate embeddings, and index them for retrieval. S3 can also integrate with other AWS services used in AI workflows. Route 53, CloudTrail, and SNS serve DNS, auditing, and messaging purposes respectively.

Question 103

What is the primary purpose of an embedding model?

1. Generate IAM policies
2. Create DNS records
3. Convert data into numerical representations
4. Monitor API requests

Correct Answer: 3

Explanation

An embedding model converts input such as text into a numerical vector representation that captures semantic characteristics. These vectors can be stored in a vector database and compared with other vectors to identify similar content. In RAG applications, document embeddings are created during indexing, while a user’s query is converted into another embedding during retrieval. Similarity between the vectors helps identify relevant content. Embedding models therefore support semantic retrieval rather than directly generating natural-language responses.

Question 104

A company wants to restrict an AI assistant from generating content about specific prohibited topics. Which capability should it consider?

1. Amazon CloudFront
2. Amazon S3
3. AWS Lambda
4. Amazon Bedrock Guardrails

Correct Answer: 4

Explanation

Amazon Bedrock Guardrails can help developers apply configurable safety policies to generative AI applications. Organizations can define denied topics and other content-related controls to reduce undesirable interactions. Guardrails can be used as part of a broader responsible AI architecture that also includes authentication, authorization, monitoring, validation, and evaluation. They do not replace all application security controls, but they provide an important layer for managing model interactions. CloudFront, S3, and Lambda provide content delivery, storage, and compute capabilities rather than equivalent model-safety controls.

Question 105

A RAG system returns relevant documents, but the model ignores important information in them. What should the developer review first?

1. Prompt instructions and context placement
2. DNS configuration
3. IAM group names
4. S3 storage class only

Correct Answer: 1

Explanation

When retrieval returns useful information but the generated answer does not use it properly, the developer should review how the retrieved context is presented to the model. Prompt instructions should clearly explain that the supplied context should be used as the primary source for the answer. Developers should also examine context ordering, formatting, length, conflicting information, and whether important passages are being truncated. Improving retrieval alone may not solve the problem if the generation prompt does not effectively guide the model to use the retrieved information.

Question 106

Which practice best supports least-privilege access for an AI application?

1. Giving every component administrator access
2. Using narrowly scoped IAM permissions
3. Sharing one access key across all services
4. Storing credentials in prompts

Correct Answer: 2

Explanation

Least privilege means granting an identity only the permissions required to perform its intended tasks. For AI applications, narrowly scoped IAM roles and policies can limit access to specific services, resources, and actions. This reduces the potential impact of compromised credentials, malicious inputs, or unexpected model behavior. Shared administrator credentials create unnecessary risk and make auditing more difficult. Credentials should also be managed through appropriate security mechanisms rather than embedded in prompts. Least privilege is particularly important for agents capable of calling external AWS services.

Question 107

What does retrieval-augmented generation combine?

1. DNS and networking
2. Storage and encryption
3. Information retrieval and generation
4. Monitoring and billing

Correct Answer: 3

Explanation

Retrieval-Augmented Generation combines a retrieval process with a generative model. First, the application searches an external knowledge source for information relevant to the user’s request. The retrieved content is then supplied to the foundation model as context, allowing the model to generate a response grounded in that information. This approach is useful for private, domain-specific, or frequently changing knowledge. RAG can improve factual grounding without requiring the foundation model to be retrained whenever source documents are updated.

Question 108

A developer needs an application to return valid machine-readable data. Which approach is most appropriate?

1. Ask for a completely free-form response
2. Remove formatting instructions
3. Provide structured output requirements
4. Increase temperature significantly

Correct Answer: 3

Explanation

Structured output requirements help a model produce responses in a predictable format that downstream software can process. A developer can specify required fields, acceptable values, data types, and formatting rules. Depending on the supported model and API capabilities, structured-output mechanisms may provide additional control. The resulting data should still be validated because a correctly formatted response can contain incorrect information. Free-form output can be harder to parse reliably, while higher temperature generally increases variability rather than improving structural consistency.

Question 109

Which metric is most directly associated with the time taken to receive a model response?

1. Latency
2. Token vocabulary
3. Storage capacity
4. Document count

Correct Answer: 1

Explanation

Latency measures the time required for a request to receive a response. In generative AI applications, latency can be influenced by model selection, prompt size, retrieved context, output length, service conditions, and application architecture. Interactive applications often have stricter latency requirements than asynchronous workloads. Developers should monitor latency using realistic production-like requests and consider it alongside quality and cost. A model that produces excellent responses may still be unsuitable for a real-time application if its response time is consistently too high.

Question 110

A developer wants to reduce hallucinations in a question-answering system. Which approach is most useful?

1. Increase randomness
2. Remove all source documents
3. Ground responses using trusted retrieved information
4. Increase output length

Correct Answer: 3

Explanation

Grounding responses in trusted retrieved information can reduce unsupported statements in knowledge-based applications. The system can retrieve relevant documents and instruct the model to use that context when generating the response. Developers should also configure the application to acknowledge when the available information is insufficient rather than inventing an answer. Retrieval quality, prompt design, output validation, and evaluation all contribute to reliability. Increasing randomness or output length does not directly improve factual grounding and may make unsupported content more likely.

Question 111

Which technique divides a large document into smaller pieces for retrieval?

1. Reranking
2. Chunking
3. Fine-tuning
4. Token billing

Correct Answer: 2

Explanation

Chunking divides large documents into smaller sections that can be independently embedded, indexed, retrieved, and supplied as model context. Proper chunking is important because the retrieved section should contain enough information to answer a query without including excessive unrelated content. Developers may use fixed-size, semantic, or document-structure-based approaches depending on the data. Chunking should be evaluated because very small chunks can lose context while very large chunks can reduce retrieval precision and increase token usage.

Question 112

What is one major advantage of RAG over relying only on a model’s pretrained knowledge?

1. It can provide access to external and updated information
2. It eliminates all model errors
3. It requires no data processing
4. It guarantees perfect responses

Correct Answer: 1

Explanation

RAG allows a generative AI application to retrieve information from an external knowledge source during inference. This is particularly valuable when information is private, domain-specific, or frequently updated. Instead of relying entirely on knowledge learned during model training, the application can provide current source material as context. RAG does not guarantee perfect answers, because retrieval and generation can still fail. However, with good document processing, retrieval, ranking, and grounding instructions, it can significantly improve the usefulness of knowledge-based applications.

Question 113

A developer wants to compare two foundation models using the same test dataset. What should be measured?

1. Only model name
2. Quality, latency, cost, and relevant application metrics
3. Only context-window size
4. Only parameter count

Correct Answer: 2

Explanation

Foundation models should be evaluated using criteria that reflect the application’s actual requirements. Developers can compare response quality, factuality, relevance, latency, token usage, cost, safety, and other appropriate metrics using the same representative dataset. Parameter count or context-window size alone does not determine whether a model is the best choice. A smaller model may provide sufficient quality at lower cost and latency. A controlled evaluation provides objective evidence for selecting a model rather than relying on assumptions based on model size or popularity.

Question 114

Which service helps record AWS API activity for auditing?

1. Amazon Polly
2. Amazon CloudWatch
3. AWS CloudTrail
4. Amazon S3

Correct Answer: 3

Explanation

AWS CloudTrail records AWS API activity and provides information that can help organizations understand which actions were performed, when they occurred, and which identity or service made the request. This makes CloudTrail useful for auditing, security investigations, governance, and compliance activities. CloudWatch focuses primarily on monitoring and observability, S3 provides object storage, and Polly provides text-to-speech capabilities. Generative AI applications that interact with multiple AWS resources can benefit from CloudTrail records when investigating unexpected or unauthorized API activity.

Question 115

A model must classify support requests into “Billing,” “Technical,” or “Account.” What can improve classification consistency?

1. Clear category definitions and examples
2. Random category names
3. Unrelated retrieval results
4. No instructions

Correct Answer: 1

Explanation

Clear category definitions help a model understand the distinctions between possible classifications. Developers can strengthen the prompt by describing each category, providing representative examples, and specifying the exact expected output. This is especially useful when categories have overlapping concepts. The classification system should then be tested against a representative dataset containing difficult and ambiguous requests. Developers may also use few-shot prompting when examples improve results. Clear instructions and consistent evaluation generally provide better results than relying on the model to infer category definitions.

Question 116

What should an application do when an AI-generated tool call contains an invalid parameter?

1. Execute it anyway
2. Validate and reject or correct the request
3. Give the model administrator access
4. Disable authentication

Correct Answer: 2

Explanation

Tool inputs generated by an AI model should be validated before execution. If a parameter is invalid, outside an allowed range, or inconsistent with authorization requirements, the application should reject or safely handle the request rather than executing it automatically. Validation is especially important for tools that can modify data, send messages, make transactions, or access sensitive systems. Developers should combine input validation with authentication, authorization, least privilege, logging, and monitoring to create safer tool-using AI applications.

Question 117

Which prompt technique uses instructions without providing examples?

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

Correct Answer: 3

Explanation

Zero-shot prompting asks the model to perform a task based on instructions without including task-specific examples. For instance, a prompt may define a classification task and provide category descriptions without showing previous classified examples. This approach is simple and does not require additional example data in every request. If the results are inconsistent, developers can improve the instructions or consider few-shot prompting, retrieval, or model customization. Zero-shot prompting is often a useful starting point for testing whether a model can perform a task from instructions alone.

Question 118

A production application suddenly experiences increased model invocation failures. Which action should be taken first?

1. Ignore the failures
2. Disable all monitoring
3. Investigate logs, metrics, quotas, and recent changes
4. Increase prompt temperature

Correct Answer: 3

Explanation

A sudden increase in invocation failures should be investigated using available logs, metrics, service information, quotas, and recent application changes. The problem could be related to throttling, configuration, permissions, invalid requests, service availability, deployment changes, or other dependencies. Monitoring tools can help identify when the problem started and whether it affects all requests or only specific workloads. Developers should gather evidence before making unrelated changes. A structured investigation helps isolate the root cause and restore reliable application behavior more efficiently.

Question 119

Why should prompt versions be tracked in a production AI application?

1. To identify and reproduce behavioral changes
2. To increase storage automatically
3. To eliminate model evaluation
4. To disable user input

Correct Answer: 1

Explanation

Prompt versioning allows developers to identify exactly which instructions were used for a particular application behavior. This is valuable when comparing experiments, investigating regressions, reproducing results, and rolling back unsuccessful changes. Prompt modifications can significantly affect model behavior even when the underlying model remains unchanged. By keeping versions and evaluation results, teams can make controlled improvements rather than changing prompts without traceability. Prompt versioning is therefore an important software-engineering practice for maintaining reliable generative AI systems.

Question 120

A company is deploying an AI agent that can access sensitive business systems. Which design provides the strongest protection?

1. Give the agent unrestricted administrator access
2. Store credentials inside the prompt
3. Allow every user to invoke every tool
4. Use least privilege, authorization, validation, monitoring, and human approval for high-risk actions

Correct Answer: 4

Explanation

Agents that can access sensitive systems require multiple layers of protection. Least-privilege permissions restrict what the agent can access, while authorization determines which users and workflows can request specific actions. Tool inputs and outputs should be validated, and activity should be logged and monitored. High-risk actions may require human approval before execution. Credentials should be securely managed rather than placed in prompts. Combining these controls limits the impact of malicious inputs, compromised accounts, incorrect model decisions, and unexpected agent behavior in production environments.