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

 

Question 281

Which approach is most effective for ensuring that an AI application uses the latest approved documents in its RAG responses?

1. Maintain an automated ingestion and index-refresh pipeline
2. Increase model temperature
3. Add more unrelated examples to the prompt
4. Disable document versioning

Correct Answer: 1

Explanation

An automated ingestion and index-refresh pipeline helps ensure that newly approved documents become available to the retrieval system. The pipeline can detect new or modified source files, extract their content, create embeddings, and update the relevant search index. Monitoring should also verify that processing completed successfully and that index freshness meets application requirements. Increasing temperature does not update source information, while unrelated prompt examples cannot compensate for stale retrieval data. Maintaining clear ingestion and indexing workflows is essential for reliable enterprise RAG applications.

Question 282

A developer is designing a prompt for an AI system that extracts invoice information. The output must contain invoice number, date, vendor, and total amount. Which prompt strategy is most appropriate?

1. Ask the model for unrestricted prose
2. Define the required fields and output structure explicitly
3. Remove all output constraints
4. Ask the model to invent missing fields

Correct Answer: 2

Explanation

Explicitly defining the required fields and output structure helps a model consistently perform information extraction. The prompt should identify the expected fields, describe how values should be represented, and specify what to do when a field is unavailable. Structured output capabilities and programmatic validation can provide additional reliability. Unrestricted prose makes downstream processing more difficult, while asking the model to invent missing values can introduce inaccurate information. Extraction systems should also be tested with diverse invoices to identify formatting and edge-case failures.

Question 283

A RAG application needs to search documents using both semantic similarity and exact product identifiers. Which retrieval architecture is most appropriate?

1. Vector search only
2. Keyword search only
3. Hybrid retrieval combining semantic and lexical search
4. Random document retrieval

Correct Answer: 3

Explanation

Hybrid retrieval combines semantic vector search with lexical or keyword search. This is particularly useful when an application handles both natural-language questions and exact identifiers such as product codes, part numbers, or policy references. Semantic search can identify conceptually related content, while keyword search can provide strong exact-match behavior. Combining the results can improve retrieval coverage and relevance. Developers should tune ranking, weighting, metadata filtering, and top-k values using representative queries to determine the configuration that best meets the application’s retrieval requirements.

Question 284

A production AI service must continue operating if a noncritical external enrichment API becomes unavailable. Which design is most appropriate?

1. Make the enrichment API mandatory for every request
2. Retry the API indefinitely
3. Disable all error logging
4. Provide a fallback path that allows reduced functionality

Correct Answer: 4

Explanation

A fallback path allows an application to continue operating with reduced functionality when a noncritical dependency becomes unavailable. For example, an AI application might omit optional enrichment information and still provide its core response. This approach is an example of graceful degradation and improves resilience. Indefinite retries can increase latency and create additional load, while disabling logging removes valuable diagnostic information. Developers should define which dependencies are critical, establish appropriate timeouts and retry limits, and test fallback behavior before relying on it in production.

Question 285

A developer wants to prevent a model from receiving sensitive fields that are not required for completing a user’s request. Which principle should be applied?

1. Data minimization
2. Maximum data exposure
3. Prompt duplication
4. Unrestricted retrieval

Correct Answer: 1

Explanation

Data minimization means providing only the information required to complete a specific task. Sending unnecessary sensitive fields to a foundation model can increase privacy, security, compliance, and cost risks. Developers should determine which data elements are necessary and exclude unrelated information before constructing prompts or retrieval context. This principle can be combined with authorization, filtering, redaction, and appropriate retention policies. Providing maximum available data is generally unnecessary and can make both security and model performance more difficult to manage.

Question 286

A developer wants to ensure an AI agent cannot execute more than five tool calls during a single workflow. Which control is most appropriate?

1. Increase the context window
2. Implement an explicit tool-call limit
3. Increase the model temperature
4. Remove tool descriptions

Correct Answer: 2

Explanation

An explicit tool-call limit can prevent an agent from repeatedly invoking tools and entering an uncontrolled loop. The application can maintain a counter for the current workflow and stop or redirect execution after the configured maximum is reached. This can reduce unnecessary costs, latency, and potential misuse of downstream services. Tool-call limits should complement authorization, input validation, and state tracking. Context size and temperature affect different aspects of model behavior, while removing tool descriptions can make tool selection less reliable.

Question 287

A company wants to compare a customized model with a foundation model for a specialized classification task. What should the evaluation use?

1. Different evaluation criteria for each model
2. Only the models’ names
3. The same representative dataset and consistent metrics
4. A single manually selected example

Correct Answer: 3

Explanation

A fair model comparison requires the candidates to be evaluated under comparable conditions. Using the same representative dataset and consistent metrics allows developers to determine whether customization provides measurable benefits. Depending on the task, metrics may include classification accuracy, precision, recall, latency, cost, and safety-related criteria. Testing only one example cannot reliably characterize model performance. Different datasets or scoring methods can also introduce bias into the comparison. Evaluation results should be interpreted against the actual requirements of the production workload.

Question 288

A developer needs to coordinate document extraction, validation, embedding generation, and index updates with retries for failed steps. Which AWS service is well suited for this workflow?

1. Amazon S3
2. AWS Secrets Manager
3. Amazon CloudWatch
4. AWS Step Functions

Correct Answer: 4

Explanation

AWS Step Functions can orchestrate multi-step workflows and provide capabilities such as sequential execution, parallel branches, retries, error handling, and conditional logic. A document-processing pipeline can use Step Functions to coordinate extraction, validation, embedding generation, and index updates while responding appropriately to failures. S3 can store documents, Secrets Manager can manage credentials, and CloudWatch can monitor operations, but they do not provide the same workflow-orchestration functionality. Explicit workflow states also make complex AI pipelines easier to understand and troubleshoot.

Question 289

Which configuration is most useful for preventing a generated response from becoming unnecessarily long?

1. Increasing the number of retrieved documents
2. Setting an appropriate maximum output-token limit
3. Increasing temperature
4. Removing response-format instructions

Correct Answer: 2

Explanation

A maximum output-token limit restricts how much content the model can generate. It can help control response length, latency, and inference cost. Prompt instructions should also specify the desired level of detail so the model understands the expected response style. Developers need to choose a limit carefully because setting it too low may truncate useful answers. Increasing temperature does not directly control response length, and adding more retrieval context can increase the opportunity for longer responses. Output limits should be validated using representative workloads.

Question 290

A generative AI application uses an external service that sometimes returns temporary 429 errors. Which response strategy is most appropriate?

1. Immediately retry continuously without delay
2. Ignore the error
3. Use controlled retries with exponential backoff and jitter
4. Increase model temperature

Correct Answer: 3

Explanation

HTTP 429 responses commonly indicate that a client is sending requests too quickly or has encountered a rate limit. Controlled retries using exponential backoff and jitter can reduce repeated pressure on the external service and give it time to recover. A maximum retry count should be configured to prevent endless attempts. The application should also monitor these errors and consider request-rate management or quota adjustments when appropriate. Immediate continuous retries can worsen throttling and increase latency. Temperature is unrelated to HTTP request-rate handling.

Question 291

A developer wants to protect a model invocation API from unauthorized callers. Which control should be implemented first?

1. Authentication and authorization
2. Larger document chunks
3. Higher temperature
4. Additional few-shot examples

Correct Answer: 1

Explanation

Authentication and authorization establish who is allowed to access an API and what actions that identity may perform. These controls should be enforced outside the foundation model because prompts cannot reliably prevent unauthorized users from reaching backend services. Depending on the architecture, API Gateway, IAM, application identity systems, and backend authorization can work together to enforce access policies. Document chunking and prompt examples affect model behavior rather than API security. Proper access controls should also follow least-privilege principles and be supported by audit logging.

Question 292

A RAG application retrieves the correct documents, but the model receives too much redundant context. Which optimization is most appropriate?

1. Increase top-k substantially
2. Remove all retrieval
3. Add unrelated documents
4. Deduplicate and rerank the retrieved context

Correct Answer: 4

Explanation

Deduplicating and reranking retrieved context can reduce redundant information while preserving the most relevant evidence. Duplicate or highly similar passages consume input tokens and may make it harder for the model to focus on the information that matters. A retrieval pipeline can combine similarity search with metadata filtering, deduplication, reranking, and appropriate top-k selection. Increasing the number of results can worsen context bloat, while removing retrieval eliminates the application’s source evidence. Context optimization should be evaluated for both answer quality and token efficiency.

Question 293

A developer needs to identify whether a new application release increases model invocation errors. Which monitoring approach is most useful?

1. Compare error metrics against a pre-release baseline
2. Delete previous monitoring data
3. Disable CloudWatch alarms
4. Increase prompt length

Correct Answer: 1

Explanation

A baseline allows developers to compare production behavior before and after a release. Error rates, invocation failures, throttling, latency, and other relevant metrics can be monitored to determine whether the new version introduces regressions. CloudWatch can support dashboards and alarms for many operational metrics. Removing previous monitoring data eliminates the comparison point, while disabling alarms reduces the ability to detect problems quickly. Prompt length is unrelated to establishing a reliable deployment baseline. Monitoring should remain active throughout rollout and stabilization.

Question 294

A company wants to ensure that an AI-generated response does not reveal confidential information from retrieved documents that the requesting user is not authorized to access. What is the strongest architectural approach?

1. Tell the model not to reveal confidential information
2. Add more examples to the system prompt
3. Enforce authorization before retrieval and response generation
4. Increase model temperature

Correct Answer: 3

Explanation

Authorization should be enforced by trusted application components before sensitive information is retrieved and supplied to the model. The retrieval system should filter documents according to the user’s permissions, and backend services should independently enforce access controls. Relying only on a prompt instruction to hide confidential information is unsafe because models are not security boundaries. Authorization should therefore occur before the model receives protected data whenever possible. This approach reduces the chance that unauthorized information becomes part of the model’s context in the first place.

Question 295

Which practice helps ensure that changes to prompts can be reproduced and rolled back?

1. Prompt versioning
2. Random prompt modification
3. Deleting previous prompts
4. Editing production prompts without records

Correct Answer: 1

Explanation

Prompt versioning provides a history of prompt changes and allows developers to reproduce or roll back specific configurations. This is important because seemingly small changes to system instructions can affect model behavior, safety, retrieval usage, and output formatting. Version control can also associate prompts with evaluation results and deployment versions. Random or undocumented changes make troubleshooting difficult and prevent teams from determining which configuration caused a behavior change. Prompt versioning should be integrated with testing and deployment workflows for production AI applications.

Question 296

A developer wants to identify whether an AI application is spending too many tokens on retrieved context. Which metric should be examined?

1. IAM policy count
2. Input token usage
3. S3 bucket count
4. CloudTrail event names

Correct Answer: 2

Explanation

Input token usage indicates how much text is being sent to the foundation model and can help identify excessive retrieval context. High input-token consumption may result from large chunks, excessive top-k values, redundant passages, unnecessary prompt instructions, or inefficient document selection. Developers can optimize context through filtering, reranking, deduplication, and better chunking. Monitoring token usage also helps with cost analysis and latency optimization. IAM policies, S3 bucket counts, and CloudTrail event names do not directly measure the amount of model input being processed.

Question 297

A developer wants to protect an AI agent from executing a destructive action based solely on an untrusted user request. Which additional control is appropriate?

1. Human approval or explicit authorization for sensitive actions
2. Higher model temperature
3. More retrieved documents
4. Larger output limits

Correct Answer: 1

Explanation

Sensitive or destructive actions should require an independent authorization boundary rather than being executed solely because a user or model requested them. Depending on the risk, the application can require explicit authorization, human approval, confirmation, or additional policy checks before executing the tool call. The agent should also use least-privilege permissions and validate tool parameters. Model temperature, retrieval size, and output limits do not provide authorization. Sensitive actions should be designed with defense in depth because incorrect model behavior can have real-world consequences.

Question 298

A RAG system’s retrieval quality decreases after the organization changes its embedding model. What should the developer consider?

1. Keep the old vector representations indefinitely
2. Re-embed the indexed documents using the compatible embedding configuration
3. Increase API Gateway timeout
4. Disable retrieval evaluation

Correct Answer: 2

Explanation

Changing an embedding model can make existing vector representations incompatible or less comparable with newly generated query embeddings. Depending on the model and architecture, the developer may need to re-embed the source documents using the new embedding configuration and rebuild or update the vector index. Retrieval quality should then be evaluated against a representative benchmark. Keeping mixed representations without validating compatibility can produce poor similarity results. API Gateway timeout settings do not correct embedding-space inconsistencies, and disabling evaluation removes an important way to detect the problem.

Question 299

A developer is optimizing an AI application and finds that a smaller foundation model provides comparable quality for simple requests at a lower cost. What should the application consider?

1. Route suitable simple requests to the lower-cost model
2. Always use the most expensive model
3. Increase all prompts to compensate
4. Disable quality evaluation

Correct Answer: 1

Explanation

Model routing can direct requests to different models based on task complexity, quality requirements, latency, or cost. If a smaller model provides acceptable results for simple requests, using it for those workloads can reduce inference costs while preserving quality where it matters. More complex requests can continue using a more capable model when necessary. The routing strategy should be validated with representative evaluation data to ensure that quality remains within acceptable thresholds. Automatically selecting the largest model for every request can create unnecessary expense.

Question 300

A production generative AI application requires continuous visibility into latency, errors, and operational health. Which combination is most appropriate?

1. Prompt examples and larger context windows
2. S3 lifecycle rules and document tags
3. CloudWatch metrics, logs, and alarms
4. Increased model temperature and output tokens

Correct Answer: 3

Explanation

CloudWatch metrics, logs, and alarms provide a foundation for monitoring operational health in AWS workloads. Developers can track latency, invocation errors, throttling, resource utilization, and application-specific indicators. Dashboards help visualize trends, while alarms can notify teams when important thresholds are exceeded. Logs provide additional detail for troubleshooting individual failures. Prompt examples and context size influence model behavior but do not provide comprehensive operational observability. Continuous monitoring is essential for identifying performance regressions, reliability issues, and unexpected production behavior.