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

 

Question 221

What is the primary purpose of an agent’s short-term memory?

1. Store permanent business records
2. Replace the knowledge base
3. Manage information relevant to the current task or conversation
4. Define user permissions

Correct Answer: 3

Explanation

Short-term memory allows an AI agent to retain information that is relevant during an active conversation or task. This can include recent user messages, intermediate results, decisions, and temporary task state. It helps the agent maintain continuity without repeatedly asking the user for information. Short-term memory differs from long-term storage because its contents are generally associated with the current interaction or workflow. Persistent business information should instead be stored in appropriate systems such as databases, knowledge repositories, or enterprise applications.

Question 222

An agent must remember a customer’s preferences across multiple conversations. Which capability is most appropriate?

1. Long-term memory
2. Token compression
3. Prompt formatting
4. Load balancing

Correct Answer: 1

Explanation

Long-term memory is appropriate when an agent needs to retain useful information across separate sessions or conversations. Customer preferences, previously confirmed settings, or other durable context can be stored in an approved persistent data source and retrieved when needed. The implementation should still follow privacy, retention, and access-control requirements. Short-term conversation context is not designed for durable storage across sessions. Properly designed long-term memory helps provide continuity while keeping persistent information under appropriate governance and security controls.

Question 223

A customer-service agent receives a request, checks an order system, and then creates a return request. Which capability coordinates these steps?

1. Data classification
2. Agent planning
3. Model fine-tuning
4. Content filtering

Correct Answer: 2

Explanation

Agent planning enables an agent to determine and coordinate multiple steps required to complete a goal. In this scenario, the agent needs to understand the customer’s request, retrieve order information, determine whether the request is valid, and then initiate the return process. Planning can involve selecting tools, maintaining task state, and determining the sequence of actions. This differs from simply generating a response because the agent must coordinate actions and intermediate results before completing the requested business process.

Question 224

Why should an AI agent validate tool outputs before using them?

1. To increase token limits
2. To make the model larger
3. To remove authentication requirements
4. To detect invalid, unexpected, or unsafe results

Correct Answer: 4

Explanation

Tool output validation helps prevent an agent from blindly trusting information returned by external systems or APIs. A tool may return incomplete, malformed, outdated, or unexpected data because of service failures or incorrect inputs. Validation can check required fields, data types, allowed values, and business rules before the information is used in later reasoning or actions. This improves reliability and reduces the chance that an incorrect tool response will cause an inappropriate answer or business operation.

Question 225

Which mechanism is commonly used to protect an API from excessive request volume?

1. Rate limiting
2. Prompt chaining
3. Embedding generation
4. Context expansion

Correct Answer: 1

Explanation

Rate limiting controls how many requests a client, user, or application can make to an API within a defined period. It helps protect services from excessive traffic, accidental request loops, and resource exhaustion. AI agents that call APIs repeatedly can encounter rate limits, so architects should also consider retry policies and backoff strategies. Proper rate limiting supports service reliability and predictable capacity. It should be combined with authentication and authorization rather than being treated as a replacement for access control.

Question 226

An external API occasionally returns temporary errors. What should an agent architecture typically use to handle these failures?

1. Permanent data deletion
2. Retry with appropriate backoff
3. Larger prompts
4. Manual model retraining

Correct Answer: 2

Explanation

Temporary service failures can often be handled through controlled retries with an appropriate backoff strategy. Instead of immediately sending repeated requests, the system waits progressively longer between attempts, reducing pressure on the unavailable service. Retry limits should also be defined to prevent endless loops. For persistent failures, the agent can use a fallback service or escalate to a human. The retry strategy should consider the API’s documented behavior, especially whether repeated operations are safe and whether requests are idempotent.

Question 227

What is the main benefit of an idempotent operation in an agent workflow?

1. It guarantees zero latency
2. It removes authentication
3. Repeating the same operation does not create unintended additional effects
4. It increases model context automatically

Correct Answer: 3

Explanation

An idempotent operation can be safely repeated without causing additional unintended effects after the first successful execution. This is especially important when agents retry API calls because network failures may make it unclear whether the original request succeeded. For example, a carefully designed transaction operation can use a unique request identifier to prevent duplicate records. Idempotency therefore improves reliability in automated workflows and reduces the risk of duplicate purchases, tickets, records, or other business actions.

Question 228

A company wants an agent to access an internal application without storing a user’s password in the agent configuration. Which approach should be considered?

1. Managed identity or another secure workload identity
2. Hard-coded password
3. Password in the system prompt
4. Plain-text credential file

Correct Answer: 1

Explanation

Managed identities and other secure workload identity mechanisms allow applications and services to authenticate without embedding long-lived passwords directly into configuration files or prompts. Access can then be controlled through appropriate permissions and role assignments. This reduces credential exposure and simplifies credential management. The exact identity mechanism depends on the platform and architecture, but secrets should not be placed in prompts, source code, or plain-text files. Agents should receive only the permissions required for their approved business operations.

Question 229

Which search approach combines keyword matching with semantic similarity?

1. Exact-match search
2. Hybrid search
3. File compression
4. Static prompting

Correct Answer: 2

Explanation

Hybrid search combines traditional keyword-based retrieval with semantic search techniques such as vector similarity. Keyword matching is useful when exact terms, identifiers, product codes, or names matter, while semantic search can find conceptually related content even when wording differs. Combining both approaches can improve retrieval quality for enterprise knowledge scenarios. The final architecture may also use filtering or reranking to improve results. Hybrid retrieval is particularly useful when a knowledge base contains both structured terminology and natural-language documents.

Question 230

A RAG system retrieves many documents, but only a few are actually relevant. Which improvement can help?

1. Increase unrelated context
2. Remove all metadata
3. Improve retrieval and reranking
4. Disable the knowledge source

Correct Answer: 3

Explanation

When retrieval returns too many irrelevant documents, improving retrieval quality and reranking can help prioritize information that is most useful for the user’s query. Techniques may include better query formulation, metadata filtering, hybrid search, improved embeddings, and reranking retrieved results. Providing large amounts of irrelevant context can increase token usage and may reduce answer quality. The goal is to provide the model with a smaller set of high-quality evidence that directly supports the requested answer.

Question 231

What does a context window determine for a language model?

1. The maximum amount of information it can process as context for a request
2. The number of users in an organization
3. The number of API keys available
4. The database storage capacity

Correct Answer: 1

Explanation

A model’s context window defines how much input and relevant conversational or retrieved information can be processed within a request, subject to the model’s limits. This may include system instructions, user messages, conversation history, retrieved documents, and other context. If an application provides excessive information, it may need summarization, truncation, or selective retrieval. Context-window capacity is therefore an important architectural consideration when designing agents that handle long conversations or large knowledge sources.

Question 232

An agent must answer questions using only approved company documents. What should the architecture emphasize?

1. Random web results
2. Unrestricted user uploads
3. Uncontrolled model memory
4. Controlled knowledge sources and grounding

Correct Answer: 4

Explanation

Controlled knowledge sources and grounding help ensure that the agent bases its responses on approved company information. The architecture should define which repositories are trusted, how documents are indexed and retrieved, and which users can access particular content. Retrieval should respect existing permissions and data boundaries. Grounding can reduce unsupported responses by providing relevant evidence to the model. Governance should also address document freshness, ownership, retention, and procedures for removing outdated or unauthorized information from the knowledge system.

Question 233

Which technique can reduce the amount of irrelevant content placed into an agent’s context?

1. Query rewriting and targeted retrieval
2. Increasing all document sizes
3. Disabling retrieval
4. Adding every available document

Correct Answer: 1

Explanation

Query rewriting and targeted retrieval can improve the relevance of information supplied to an agent. A rewritten query can clarify the user’s intent and help the retrieval system locate more appropriate documents. Filters based on metadata, permissions, dates, or document types can further reduce unnecessary results. Sending every available document to the model increases context consumption and can introduce conflicting or irrelevant information. Targeted retrieval therefore supports better answer quality, lower token usage, and more efficient agent operation.

Question 234

A business needs to classify incoming support requests into predefined categories. Which AI approach is generally suitable?

1. Image generation
2. Classification
3. Speech synthesis
4. Vector storage

Correct Answer: 2

Explanation

Classification is designed to assign inputs to predefined categories based on their characteristics. For example, support requests could be classified as billing, technical support, account access, or shipping issues. A classification model or appropriately configured AI capability can perform this task efficiently. The architecture should define the allowed categories and establish evaluation criteria for accuracy. Generative AI may also be involved in broader workflows, but a straightforward categorization requirement does not necessarily require a complex autonomous agent.

Question 235

What is a key reason to use a model catalog when selecting an AI model?

1. It automatically approves every model
2. It replaces security testing
3. It provides information for comparing available models and capabilities
4. It eliminates deployment requirements

Correct Answer: 3

Explanation

A model catalog can provide information about available models, supported capabilities, deployment options, performance characteristics, and other relevant attributes. This helps architects compare models against business requirements such as accuracy, latency, cost, context capacity, and supported workloads. Model selection should still include security, compliance, licensing, and evaluation considerations. A catalog supports informed selection but does not automatically determine which model is appropriate. Organizations should validate candidate models against representative workloads before production adoption.

Question 236

An organization wants to reduce AI costs while maintaining acceptable response quality. Which strategy can help?

1. Route simple requests to smaller models
2. Send every request to the largest model
3. Increase prompt length unnecessarily
4. Disable monitoring

Correct Answer: 1

Explanation

Routing simple workloads to smaller or less expensive models can reduce AI costs while preserving appropriate quality. More capable models can remain available for complex requests that require advanced reasoning or larger context. A model-routing strategy should be based on tested criteria rather than assumptions. Organizations can evaluate accuracy, latency, token consumption, and business impact for representative requests. Cost monitoring should continue after deployment because workload patterns may change and model usage can increase over time.

Question 237

Which control helps prevent an agent from accessing data outside its assigned business function?

1. Larger context windows
2. Broader API permissions
3. Unrestricted connectors
4. Role-based access control

Correct Answer: 4

Explanation

Role-based access control, or RBAC, can restrict an agent or its associated identity to permissions appropriate for its assigned responsibilities. This supports the principle of least privilege by preventing unnecessary access to applications, records, or operations. RBAC should be combined with authentication, authorization checks, and appropriate data-level controls where required. Increasing an agent’s permissions simply to make integrations easier can create unnecessary security exposure. Access should therefore be explicitly designed around the agent’s legitimate business tasks.

Question 238

What should an AI team establish before moving an agent from pilot to production?

1. Only a larger prompt
2. Production readiness criteria
3. Unlimited user access
4. Removal of monitoring

Correct Answer: 2

Explanation

Production readiness criteria provide measurable requirements that an agent should satisfy before deployment to real users. These can include accuracy, groundedness, latency, reliability, security, privacy, cost, monitoring, failure handling, and business acceptance criteria. Testing should use representative scenarios rather than only successful demonstrations. Clear readiness gates reduce the chance of deploying an immature solution. They also provide stakeholders with a consistent basis for deciding whether the system has met the technical and business requirements established for production use.

Question 239

An agent produces incorrect answers after a major change to its retrieval pipeline. What should the team investigate first?

1. Retrieval and evaluation results before and after the change
2. User interface colors
3. Office network printers
4. Employee vacation schedules

Correct Answer: 1

Explanation

A significant retrieval change can alter which evidence is provided to the model, directly affecting answer quality. The team should compare evaluation results before and after the change and inspect retrieval relevance, document coverage, ranking, filters, and grounding behavior. Representative test cases can help identify whether the problem affects specific query types or the overall system. Monitoring and evaluation data provide evidence for diagnosing the regression. Changes should be validated systematically before being promoted to production environments.

Question 240

Why is versioning important for AI prompts and agent configurations?

1. It prevents all model errors
2. It removes the need for testing
3. It allows teams to track and reproduce changes
4. It guarantees lower operating costs

Correct Answer: 3

Explanation

Versioning allows teams to track changes to prompts, agent instructions, configurations, tools, and related assets over time. When behavior changes, teams can identify what was modified and reproduce earlier configurations for comparison or rollback. Versioning also supports controlled testing and collaboration across development environments. It does not guarantee that an AI system will be accurate or inexpensive. Combining version control with evaluation, deployment gates, monitoring, and documented change management creates a more reliable lifecycle for AI solutions.