View Full Microsoft AI-103 Exam Dumps and Practice Test Dumps.

 

Question 281

Which capability helps an AI agent remember information during a conversation?

1. OCR
2. Image classification
3. State or memory management
4. Rate limiting

Correct Answer: 4

Explanation

State or memory management allows an AI agent to maintain relevant information across multiple interactions or steps in a workflow. For example, an agent may need to remember a user’s selected product, previous request, or current task status. Developers should carefully determine what information needs to be retained and for how long. Sensitive information should not be stored unnecessarily. State management can improve conversational continuity, but applications should also consider data protection, access controls, expiration policies, and consistency when maintaining agent state.

Question 282

What is the main purpose of a model catalog?

1. Discover available AI models and their capabilities
2. Store application passwords
3. Convert documents into audio
4. Monitor network cables

Correct Answer: 1

Explanation

A model catalog provides a way to discover available AI models and review information about their capabilities, supported scenarios, and other relevant characteristics. Developers can use this information when selecting a model for a particular workload. Different models may vary in language support, multimodal capabilities, context size, performance, cost, and other properties. Model selection should ultimately be validated using representative application data. A catalog helps developers understand available choices, but testing is still necessary to determine whether a model meets production requirements.

Question 283

Why is prompt engineering useful in generative AI applications?

1. It increases storage capacity
2. It improves the instructions given to the model
3. It replaces authentication
4. It creates vector indexes automatically

Correct Answer: 2

Explanation

Prompt engineering involves designing and refining instructions and input context to guide a generative model toward the desired behavior. Clear prompts can specify the task, required format, relevant constraints, and other useful information. Developers can use examples, explicit instructions, and structured context when appropriate. Prompt engineering alone does not guarantee factual accuracy or security, so applications should still use retrieval controls, validation, authorization, and evaluation. Testing different prompts against representative examples can help determine whether a prompt produces consistent and useful results.

Question 284

Which component can provide external information to an AI agent during execution?

1. Tool
2. Temperature
3. Token counter
4. Content filter

Correct Answer: 1

Explanation

A tool allows an AI agent to interact with external systems or obtain information that is not directly available in the model’s current context. Examples include APIs for order status, calendars, databases, or business applications. Tools extend an agent’s capabilities beyond text generation. Because tools can perform real operations, developers should restrict permissions, authenticate requests, validate arguments, and handle errors appropriately. Tool descriptions should also clearly explain their purpose and expected inputs so the agent can determine when a particular tool is appropriate.

Question 285

Which feature is useful for retrieving documents based on meaning?

1. Speech recognition
2. OCR
3. Semantic or vector search
4. Text-to-speech

Correct Answer: 3

Explanation

Semantic and vector search can retrieve documents based on meaning rather than relying only on exact keyword matches. Vector search uses embeddings to compare the semantic representation of a query with representations of indexed content. Semantic search can also use language-aware ranking to identify relevant results. These approaches are useful for knowledge bases, enterprise search, and RAG systems. Developers should evaluate retrieval quality using representative queries and documents because the effectiveness of semantic retrieval depends on content quality, configuration, language, and the selected search capabilities.

Question 286

What is the purpose of an AI agent’s planning capability?

1. Break a complex task into appropriate steps
2. Store encryption keys
3. Convert images to text
4. Increase microphone volume

Correct Answer: 1

Explanation

Planning allows an AI agent to organize a complex task into multiple steps that may involve reasoning, retrieval, and tool use. For example, an agent might first identify a user’s request, retrieve relevant information, call an external service, and then prepare a response. Developers should define appropriate boundaries around agent planning because unrestricted actions can introduce security and reliability risks. Tool permissions, validation, authorization, and error handling remain necessary even when an agent can plan tasks independently.

Question 287

Which service can extract text and structure from supported documents?

1. Azure AI Document Intelligence
2. Azure AI Speech
3. Azure AI Search only
4. Azure AI Language only

Correct Answer: 1

Explanation

Azure AI Document Intelligence can analyze supported documents and extract text, fields, tables, and other structural information depending on the selected capability. It is useful for automating workflows involving invoices, receipts, forms, applications, and similar documents. Extracted information can be passed to downstream applications after validation. Developers should test different document layouts and quality levels because real-world files can vary significantly. Proper error handling is also important when documents are incomplete, corrupted, unsupported, or significantly different from the examples used during development.

Question 288

Why should retrieved content be separated from trusted system instructions?

1. Retrieved content may contain untrusted instructions
2. It increases image resolution
3. It removes the need for search
4. It guarantees model accuracy

Correct Answer: 1

Explanation

Retrieved content may contain text that attempts to influence the model’s behavior, especially when documents originate from external or user-controlled sources. Separating untrusted retrieved content from trusted system instructions helps reduce the risk that embedded instructions will be interpreted as higher-priority application rules. Developers should also restrict tool permissions, validate inputs, and enforce authorization outside the model. Prompt-injection defenses are particularly important in RAG and agent systems because retrieved documents can contain unexpected or malicious content.

Question 289

What does a content safety mechanism help detect or control?

1. Potentially harmful or inappropriate content
2. Vector dimensions
3. Database indexes
4. Document page numbers only

Correct Answer: 1

Explanation

Content safety mechanisms help identify or control supported categories of potentially harmful or inappropriate content in AI inputs and outputs. They can provide an additional protection layer around generative AI applications. Depending on the service and configuration, content safety controls may help detect categories that require blocking, filtering, review, or other handling. Developers should not rely on content safety alone. Authentication, authorization, input validation, monitoring, and application-specific policies are still necessary. Testing should include representative scenarios and known edge cases to understand the system’s behavior.

Question 290

What is the main benefit of using an indexer with a supported data source?

1. Automating content ingestion into a search index
2. Increasing model temperature
3. Converting speech to images
4. Removing document permissions

Correct Answer: 1

Explanation

An indexer can automate the process of retrieving content from a supported data source and populating a search index. This is useful when an application needs to keep searchable content synchronized with an external repository. Developers can configure the data source, target index, field mappings, and execution schedule according to the workload. Indexing should be monitored for failures and data freshness. Applications with sensitive content should also ensure that appropriate access-control information is preserved and applied during retrieval so users receive only authorized information.

Question 291

Which metric is useful when monitoring an AI application’s responsiveness?

1. Latency
2. Image color
3. Document font
4. File extension

Correct Answer: 1

Explanation

Latency measures the time required for an operation or request to complete and is an important operational metric for AI applications. High latency can negatively affect interactive experiences, while unexpectedly low or high values can also reveal changes in application behavior. Developers can monitor request latency alongside token usage, error rates, throughput, and service availability. Different workloads may have different acceptable latency requirements. Monitoring these metrics over time helps teams identify performance changes and investigate whether model, retrieval, infrastructure, or application changes are affecting responsiveness.

Question 292

What is the purpose of role-based access control?

1. Assign permissions according to defined roles
2. Generate embeddings
3. Convert audio to text
4. Increase model creativity

Correct Answer: 1

Explanation

Role-based access control, or RBAC, assigns permissions according to defined roles rather than granting every user unrestricted access. In AI applications, RBAC can help control access to models, search resources, storage, tools, and other Azure resources. Developers should assign only the roles required for each identity and regularly review them. RBAC supports the principle of least privilege but does not eliminate the need for application-level authorization in every scenario. Proper role design can reduce unnecessary access and help protect AI resources and data.

Question 293

Why might an application use a smaller model for classification?

1. Simple classification may not require a large model
2. Smaller models always provide perfect accuracy
3. Classification requires no evaluation
4. Smaller models automatically provide administrator access

Correct Answer: 1

Explanation

A simple classification task may not require the capabilities of a larger generative model. A smaller model can potentially provide sufficient quality while using fewer resources and reducing latency or cost. The decision should be based on evaluation rather than model size alone. Developers should test candidate models with representative examples and measure accuracy, latency, cost, and other requirements. If a smaller model does not meet the required quality, a more capable model may be necessary. Model selection should always reflect the application’s actual workload.

Question 294

What can source citations provide in a RAG response?

1. References to supporting retrieved information
2. User administrator privileges
3. Automatic model retraining
4. Image enhancement

Correct Answer: 1

Explanation

Source citations can identify the documents or passages that support information included in a generated response. They can make a RAG application more transparent by helping users understand where an answer came from and allowing them to review the supporting material when appropriate. Citations should correspond to actually retrieved and relevant sources rather than being generated without evidence. Developers should evaluate citation accuracy along with response quality and grounding. Access controls must also be applied so users do not receive citations or links to information they are not authorized to access.

Question 295

Which technique can help reduce hallucinations in a knowledge-based application?

1. Grounding responses in retrieved information
2. Removing all source documents
3. Increasing unrelated context
4. Disabling evaluation

Correct Answer: 1

Explanation

Grounding a model’s response in relevant information retrieved from approved knowledge sources can help reduce unsupported claims. Instead of asking the model to answer only from its internal knowledge, the application supplies context from documents or other trusted sources. However, grounding does not guarantee that every response will be correct. Retrieval quality, source freshness, prompt design, and model behavior all affect the result. Developers should evaluate whether generated statements are actually supported by retrieved evidence and should implement appropriate response validation for important workflows.

Question 296

What should happen when an AI service temporarily returns a retryable error?

1. Use controlled retry logic when appropriate
2. Retry continuously without limits
3. Disable authentication
4. Delete the application

Correct Answer: 1

Explanation

A retryable error may result from a temporary network issue, transient service problem, or temporary capacity condition. In such cases, controlled retry logic can allow the application to try the operation again after an appropriate delay. Developers should avoid unlimited retries because they can increase traffic, costs, and system load. Exponential backoff and maximum retry limits are commonly used approaches where supported. Permanent errors, invalid requests, and authorization failures generally should not be handled through repeated retries and require different corrective actions.

Question 297

Which capability identifies the language of a text input?

1. Language detection
2. Object detection
3. Speech synthesis
4. Document indexing

Correct Answer: 1

Explanation

Language detection identifies the language used in a written text input. This can help applications route content to appropriate processing workflows, such as translation, language-specific analysis, or multilingual customer support. It is different from speech recognition, which converts spoken audio into text. Developers should confirm that the languages required by the application are supported and evaluate the service using realistic inputs. Short text, mixed-language content, unusual spelling, and ambiguous phrases can sometimes make language identification more challenging.

Question 298

Why is data minimization useful in an AI application?

1. It reduces unnecessary collection and exposure of sensitive information
2. It increases model temperature
3. It guarantees perfect retrieval
4. It eliminates authorization requirements

Correct Answer: 1

Explanation

Data minimization means collecting, processing, and retaining only the information necessary for the intended purpose. In AI applications, reducing unnecessary data can lower privacy and security risks and may also reduce processing costs. Developers should determine what information the model actually needs and avoid sending unrelated sensitive data in prompts or retrieved context. Retention policies, access controls, encryption, and secure logging should also be considered. Data minimization is particularly important when applications process personal, confidential, or business-sensitive information.

Question 299

Which component is responsible for converting text into embeddings for vector retrieval?

1. Embedding model
2. Speech synthesizer
3. Content filter
4. OCR engine

Correct Answer: 1

Explanation

An embedding model converts supported content, such as text, into numerical vectors that capture semantic characteristics of the input. These vectors can then be stored in a vector index and compared with query embeddings during retrieval. The embedding model used for documents should be compatible with the model used to embed search queries. Developers should evaluate embedding quality using representative data because different models can perform differently across languages, domains, and content types. Embeddings are a key component of many semantic and RAG retrieval architectures.

Question 300

What should developers verify before deploying an AI agent with business tools?

1. Permissions, tool inputs, security, and error handling
2. Only the agent’s name
3. Only the user interface color
4. Only the prompt length

Correct Answer: 1

Explanation

Before deploying an AI agent with business tools, developers should verify permissions, authentication, authorization, tool argument validation, error handling, and appropriate security controls. Agents may interact with external systems and can therefore create real-world effects if a tool is misused or receives incorrect arguments. Sensitive operations may require human approval. Developers should also evaluate the agent with representative and adversarial scenarios and monitor it after deployment. Proper preparation helps ensure that the agent operates within defined business rules rather than relying solely on model-generated instructions.