View Full Microsoft AB-100 Exam Dumps and Practice Test Dumps.

 

Question 361

What is the main purpose of a model selection matrix?

1. To store conversation history
2. To define user permissions
3. To manage document chunks
4. To compare models against defined business and technical requirements

Correct Answer: 4

Explanation

A model selection matrix provides a structured way to compare candidate models against requirements such as quality, latency, cost, context capacity, availability, safety, and supported capabilities. Instead of choosing a model based on a single characteristic, architects can evaluate several relevant dimensions for the intended workload. The matrix can also document tradeoffs and support consistent decision-making across projects. Requirements should be weighted according to the specific business scenario, because a model suitable for one workload may not be appropriate for another.

Question 362

Which identity option allows an Azure-hosted application to access supported resources without storing credentials in application code?

1. Managed identity
2. Shared password
3. Hard-coded API key
4. Anonymous access

Correct Answer: 1

Explanation

Managed identity provides an Azure workload with an identity that can be authorized to access supported resources without requiring developers to store credentials directly in application code. This reduces the risk associated with exposed passwords, keys, or connection secrets. Permissions should still follow least privilege, and the target resource must recognize and authorize the identity. Managed identities are particularly useful for production AI applications that access storage, databases, search services, or other Azure resources.

Question 363

An agent repeatedly receives the same request within a short period. Which technique can reduce unnecessary processing?

1. Fine-tuning
2. Caching
3. Data classification
4. Human escalation

Correct Answer: 2

Explanation

Caching can reduce unnecessary processing when identical or safely reusable results are requested repeatedly. Instead of invoking a model or backend service for every identical request, the system can reuse an appropriate cached result for a defined period. Cache design must consider freshness, privacy, authorization, and whether the result is still valid for the current user. Sensitive or user-specific information should not be returned from an improperly shared cache. Caching is therefore both a performance and architecture-design consideration.

Question 364

Which control helps prevent an AI application from sending sensitive information to an unauthorized service?

1. Larger context window
2. Model temperature
3. Data loss prevention policy
4. Prompt summarization

Correct Answer: 3

Explanation

Data loss prevention policies can help identify and restrict the movement or use of sensitive information according to organizational rules. In an AI architecture, DLP controls can help prevent sensitive data from being sent to unauthorized applications, connectors, or external services. DLP should work alongside identity, access controls, data classification, and application-level validation. The exact capabilities depend on the platform and configuration. Organizations should test policies with representative scenarios to ensure legitimate business workflows continue to function.

Question 365

Why should AI-generated citations point to the actual supporting source?

1. To increase token limits
2. To improve traceability and help users verify information
3. To remove access controls
4. To eliminate retrieval

Correct Answer: 2

Explanation

Grounded citations provide traceability by showing users which source information supports an AI-generated response. This is particularly valuable for enterprise applications where users need to verify policies, procedures, financial information, or other business content. Citations do not guarantee that an answer is correct, so the underlying retrieval and source-quality processes remain important. Access controls should also be applied before content is retrieved. A well-designed system should avoid presenting citations that expose documents the requesting user is not authorized to access.

Question 366

Which metric is especially useful when evaluating how many retrieved documents are relevant to a query?

1. Retrieval precision
2. API availability
3. Token quota
4. Model throughput

Correct Answer: 1

Explanation

Retrieval precision measures the proportion of retrieved results that are relevant to the query. A higher precision generally means the retrieval system is returning fewer irrelevant documents among its selected results. This can be particularly important in RAG systems because irrelevant context may distract the model and consume valuable context space. Precision should be considered alongside other retrieval measures, such as recall, because retrieving only a small number of highly relevant documents may still miss important evidence needed to answer the query.

Question 367

A business wants an AI system to classify support tickets into predefined categories. Which AI approach is most directly suited to this task?

1. Predictive or classification AI
2. Autonomous planning only
3. Long-term memory
4. Document archiving

Correct Answer: 1

Explanation

Classification AI is designed to assign inputs to predefined categories based on learned patterns or configured rules. For support tickets, categories might include billing, technical support, account access, or service requests. Generative AI could also participate in a broader workflow, such as summarizing the ticket or generating a response, but classification is the direct capability for assigning categories. The selected approach should be evaluated using representative tickets and business-defined accuracy requirements before being deployed into production.

Question 368

What is a major reason to use a custom model instead of a general-purpose model?

1. To remove all governance requirements
2. To avoid collecting evaluation data
3. To address a specialized workload when existing models are insufficient
4. To guarantee zero latency

Correct Answer: 3

Explanation

A custom model may be appropriate when a specialized workload has requirements that available general-purpose models cannot adequately satisfy. Examples can include domain-specific classification or specialized prediction tasks where representative training data is available. However, developing and maintaining a custom model introduces additional costs, infrastructure, evaluation requirements, and lifecycle responsibilities. Teams should first determine whether prompting, retrieval, configuration, or an existing specialized model can satisfy the requirement. Custom development should be justified by measurable business or technical needs.

Question 369

Which practice helps protect an agent from jailbreak attempts?

1. Disabling authentication
2. Content safety and prompt protection controls
3. Increasing token limits
4. Removing system instructions

Correct Answer: 2

Explanation

Content safety controls and prompt protection mechanisms can help identify or reduce attempts to manipulate an agent into violating its intended behavior. Security measures may include prompt shields, input filtering, system instructions, output validation, and adversarial testing. These controls should be layered rather than relying on a single defense. Sensitive actions should also require technical authorization and, when appropriate, human approval. Security testing should include realistic jailbreak attempts so weaknesses can be identified before and after deployment.

Question 370

Which factor is important when deciding whether an AI workload can meet a strict response-time requirement?

1. Model and system latency
2. Number of document owners
3. Retention policy name
4. User interface font

Correct Answer: 1

Explanation

Model and system latency directly affects whether an AI application can meet a required response time. Total latency can include model inference, retrieval, tool calls, network communication, orchestration, and post-processing. Architects should measure end-to-end latency under realistic workloads rather than considering model response time alone. Techniques such as caching, parallel execution, smaller models, targeted retrieval, and optimized infrastructure may improve performance. The acceptable latency should be defined as part of the business requirements and validated during testing.

Question 371

What is the purpose of a circuit breaker in an AI application?

1. Permanently delete failed requests
2. Prevent repeated calls to an unhealthy service
3. Increase model temperature
4. Expand the context window

Correct Answer: 2

Explanation

A circuit breaker helps prevent an application from repeatedly calling a service that is currently failing or unavailable. After failures exceed a defined threshold, the circuit can temporarily stop requests and allow the system to use a fallback, return a controlled error, or wait before retrying. This protects dependent services and prevents cascading failures. Circuit breakers are especially useful when AI agents depend on external APIs or backend services. They should be combined with appropriate timeout, retry, and monitoring strategies.

Question 372

An external API limits an agent to a fixed number of requests per minute. What should the architecture implement?

1. API throttling and rate-limit handling
2. Unlimited retries
3. Larger prompts
4. Permanent request storage

Correct Answer: 1

Explanation

API throttling and rate-limit handling help ensure that an agent operates within the limits imposed by an external service. The application can monitor usage, limit request rates, queue work, and use controlled retry behavior when limits are reached. Exponential backoff can reduce repeated pressure on the API. Architects should also consider concurrency and quotas because multiple agent instances may share the same service limit. Proper rate-limit handling improves reliability and helps prevent avoidable service failures during periods of increased demand.

Question 373

Which information is most important when defining an agent’s scope?

1. The tasks, boundaries, and actions the agent is responsible for
2. The number of colors in the interface
3. The employee’s screen resolution
4. The database backup schedule

Correct Answer: 1

Explanation

An agent’s scope should clearly define the tasks it is responsible for, the information it may access, and the actions it may perform. Clear boundaries reduce unintended behavior and make security, testing, monitoring, and governance easier. Scope should also identify situations that require escalation or human intervention. A narrowly defined agent can be easier to control than an agent given broad responsibilities without clear limits. Scope should be reviewed whenever business requirements, integrations, or permissions change.

Question 374

A RAG system retrieves relevant documents but often misses the best matching passage. What should the team investigate first?

1. Retrieval configuration and search strategy
2. User interface colors
3. Password expiration
4. Database backup frequency

Correct Answer: 1

Explanation

When relevant information exists but the retrieval system frequently misses the best passage, the team should investigate retrieval configuration and search strategy. Factors can include query formulation, embeddings, chunking, metadata filters, hybrid search, reranking, and indexing quality. Evaluation should use representative queries with known relevant sources so changes can be measured objectively. Improving the language model alone may not solve a retrieval problem. In a RAG architecture, retrieval quality is a separate component that requires its own testing and monitoring.

Question 375

What is the main purpose of a knowledge ingestion pipeline?

1. Prepare and update source information for retrieval
2. Increase user permissions
3. Replace identity management
4. Disable model evaluation

Correct Answer: 2

Explanation

The primary purpose of a knowledge ingestion pipeline is to collect, process, and prepare source information so it can be used effectively by a retrieval system. Processing may include document extraction, cleaning, chunking, metadata creation, embedding generation, indexing, and scheduled updates. The pipeline should also account for deleted or changed content so outdated information does not remain available indefinitely. Access controls and data governance should be considered throughout ingestion because source documents may contain sensitive or restricted information.

Question 376

Which approach can improve resilience when a primary model becomes temporarily unavailable?

1. Model fallback
2. Removing monitoring
3. Disabling authentication
4. Increasing prompt length

Correct Answer: 1

Explanation

Model fallback allows an application to use an alternative model or workflow when the primary model is unavailable or fails defined requirements. The fallback should be selected based on compatibility, quality, latency, cost, and business impact. Not every fallback needs to provide identical capabilities; some applications may use a simpler model or controlled response when the primary service is unavailable. Fallback behavior should be tested regularly so failures do not create unexpected behavior or expose sensitive operations.

Question 377

Which practice helps ensure that changes to prompts are traceable?

1. Prompt versioning
2. Removing deployment records
3. Sharing one editable prompt with everyone
4. Disabling testing

Correct Answer: 1

Explanation

Prompt versioning maintains identifiable versions of prompts and records the changes made between them. This allows teams to determine which prompt produced a particular behavior and makes it easier to compare versions during evaluation. Versioning is useful for troubleshooting, rollback, controlled deployment, and governance. Prompt changes should ideally be tested against representative evaluation datasets before reaching production. Combining version control with documented ownership and release processes provides better traceability across the AI application’s lifecycle.

Question 378

Why should AI applications use separate development, test, and production environments?

1. To reduce uncontrolled changes reaching production
2. To eliminate all testing
3. To provide every developer with production access
4. To avoid version control

Correct Answer: 1

Explanation

Separate development, test, and production environments provide controlled stages for building, validating, and deploying AI solutions. Developers can experiment without directly affecting production users or business data. Changes can then be evaluated in a test environment before deployment. This separation also supports access control, versioning, approval processes, and troubleshooting. Production environments should contain only validated configurations and components. The exact environment structure can vary by organization, but uncontrolled direct changes to production increase operational and security risk.

Question 379

A company wants to determine whether an AI pilot created measurable business value. What should it define?

1. Business success metrics and acceptance criteria
2. More model parameters
3. Additional unrelated data
4. Unlimited user access

Correct Answer: 3

Explanation

A pilot should have measurable business success metrics and acceptance criteria established before results are assessed. Depending on the use case, metrics might include processing time, resolution rate, cost per task, user satisfaction, accuracy, or reduction in manual work. Technical metrics alone may not demonstrate business value. Baseline measurements from the existing process are also useful for comparison. Clear success criteria help organizations determine whether the pilot should be improved, expanded, redesigned, or discontinued based on evidence.

Question 380

Which activity should be included in an AI solution’s production readiness review?

1. Disabling logs
2. Removing access controls
3. Verifying security, reliability, monitoring, and operational requirements
4. Allowing unrestricted tool access

Correct Answer: 3

Explanation

A production readiness review should verify that the AI solution meets required security, reliability, monitoring, performance, governance, and operational standards. The review can examine authentication, authorization, data protection, failure handling, telemetry, evaluation results, deployment processes, and incident procedures. Business acceptance criteria should also be satisfied before release. Production readiness is broader than confirming that the model generates useful responses. A solution should be capable of operating safely and reliably within its intended business environment.