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

 

Question 381

Which capability allows an AI application to process both text and images?

1. Keyword search
2. Speech-to-text
3. Text summarization
4. Multimodal AI

Correct Answer: 4

Explanation

Multimodal AI allows a model or application to work with more than one type of input, such as text and images. This can support scenarios like asking questions about an uploaded image, analyzing a document image together with written instructions, or combining visual and textual information. Developers should verify that the selected model supports the required input types and size limits. They should also evaluate multimodal responses using representative examples because image quality, visual complexity, and ambiguous instructions can affect the generated result.

Question 382

What is the main purpose of Azure AI Search in a RAG solution?

1. Retrieve relevant information
2. Convert text to speech
3. Detect objects
4. Translate audio

Correct Answer: 1

Explanation

Azure AI Search can provide the retrieval layer in a RAG solution by searching indexed content and returning information relevant to a user’s query. Depending on the configuration, applications can use keyword, vector, semantic, or hybrid search. Retrieved content can then be supplied to a generative model as context. Developers should configure indexing, chunking, ranking, filtering, and access controls appropriately. Retrieval quality should be evaluated independently because poor search results can lead to inaccurate or poorly grounded generated responses.

Question 383

What does a content safety system help detect or reduce?

1. Network latency
2. Harmful or inappropriate content
3. Database duplication
4. Search index size

Correct Answer: 2

Explanation

Content safety capabilities help detect or reduce certain categories of harmful or inappropriate content in AI applications. They can be applied to inputs, outputs, or other relevant content depending on the selected service and configuration. Content safety is one part of a broader responsible AI design that may also include access controls, prompt injection defenses, monitoring, human review, and application-level validation. Developers should test safety controls using realistic and adversarial examples and understand that filtering does not replace authorization or other security mechanisms.

Question 384

Why should retrieved documents include useful metadata?

1. To increase audio quality
2. To generate passwords
3. To support filtering and organization
4. To change model temperature

Correct Answer: 3

Explanation

Metadata provides additional information about indexed content, such as document type, department, date, language, region, or access-related attributes. This information can help applications filter and organize search results before supplying content to a generative model. Metadata can therefore improve both relevance and access control when designed correctly. Developers should ensure metadata remains accurate when source documents change. Incorrect or outdated metadata can cause useful documents to be excluded or inappropriate documents to be retrieved, so metadata quality should be included in evaluation and monitoring.

Question 385

Which approach is useful for protecting secrets used by an AI application?

1. Store them in source code
2. Put them in prompts
3. Use a secure secret-management service
4. Log them for debugging

Correct Answer: 3

Explanation

A secure secret-management service can protect credentials, keys, and other sensitive values used by an AI application. Storing secrets directly in source code, prompts, or ordinary log files can increase the risk of accidental exposure. Azure Key Vault is an example of a service designed for secure management of secrets and keys. Applications should also use appropriate identities and least-privilege permissions when accessing secrets. Developers should review logs and configuration files to ensure credentials are not accidentally exposed during development, deployment, or troubleshooting.

Question 386

Which technique can improve retrieval when users ask vague questions?

1. Query rewriting
2. Image classification
3. Text-to-speech
4. OCR

Correct Answer: 1

Explanation

Query rewriting can transform a vague or conversational question into a clearer search query before retrieval occurs. For example, a short follow-up question may need information from previous conversation context before it can be searched effectively. Rewriting can make important terms, entities, or constraints more explicit. Developers should evaluate rewritten queries because an incorrect rewrite can change the user’s intended meaning and reduce retrieval quality. Query rewriting can be combined with semantic, vector, or hybrid search to improve the overall retrieval process.

Question 387

What should happen if a user is not authorized to access a retrieved document?

1. Return the document anyway
2. Exclude or deny access to the document
3. Send the document to another user
4. Remove authentication

Correct Answer: 2

Explanation

A document that the requesting user is not authorized to access should not be returned to the user or supplied to the model for generation. Authorization should be enforced before restricted information reaches the response-generation stage. This is particularly important in enterprise RAG systems where an index may contain documents belonging to different users, teams, or security groups. Developers can use identity information, metadata filters, and resource-level permissions as part of the access-control design. Security testing should verify that unauthorized users cannot retrieve restricted content.

Question 388

Which capability can extract tables and fields from supported documents?

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

Correct Answer: 3

Explanation

Azure AI Document Intelligence can analyze supported documents and extract information such as text, tables, layout elements, and fields depending on the selected model. This makes it useful for processing forms, invoices, receipts, and other structured or semi-structured documents. Developers should select an appropriate prebuilt or custom model and validate extracted values before using them in important workflows. Testing should include different layouts and document qualities because extraction accuracy can vary. Sensitive documents should also be protected with appropriate access controls and data-handling practices.

Question 389

Why is human review useful for high-impact AI actions?

1. It provides an opportunity to verify important actions
2. It increases embedding dimensions
3. It removes the need for authentication
4. It guarantees model accuracy

Correct Answer: 1

Explanation

Human review can provide an additional control before an AI system performs a sensitive or high-impact action. A reviewer can inspect the proposed action and relevant information before approving or rejecting it. This can be useful for financial transactions, account changes, sensitive communications, or other operations where an incorrect action could have significant consequences. Human review should complement technical controls such as authorization, validation, monitoring, and tool restrictions. Developers should clearly define which actions require approval and ensure unapproved actions cannot execute automatically.

Question 390

Which parameter primarily affects response variability?

1. Token limit
2. Temperature
3. Search filter
4. Indexer schedule

Correct Answer: 2

Explanation

Temperature primarily influences the degree of randomness or variability in generated responses. Lower temperature settings generally produce more predictable outputs, while higher settings can allow greater variation. The appropriate value depends on the application’s purpose. Tasks requiring consistent formatting or predictable responses may benefit from lower variability, while creative applications may allow more variation. Temperature does not guarantee factual correctness and should not be used as a replacement for grounding, retrieval, evaluation, or output validation. Developers should test configurations against representative prompts.

Question 391

Which capability identifies positive or negative opinions in text?

1. Sentiment analysis
2. OCR
3. Vector search
4. Object detection

Correct Answer: 1

Explanation

Sentiment analysis evaluates text to determine the expressed sentiment, such as positive, negative, or neutral, depending on the capabilities of the selected service. It can be useful for analyzing customer reviews, survey responses, support conversations, and other feedback. Developers should test sentiment analysis with representative domain-specific text because sarcasm, mixed opinions, context, and unusual terminology can affect results. Sentiment output should not automatically be treated as perfect ground truth. For important workflows, applications should use evaluation and appropriate validation before acting on the result.

Question 392

What does least-privilege access require?

1. Giving every user administrator permissions
2. Giving only necessary permissions
3. Disabling role-based access control
4. Sharing one account with all users

Correct Answer: 2

Explanation

Least privilege requires users, applications, agents, and services to receive only the permissions necessary to perform their intended tasks. This limits the potential impact of compromised credentials, accidental actions, or malicious input. For example, an agent that only retrieves order information should not automatically receive permissions to delete customer records. Developers should regularly review roles, tool permissions, and resource access. Least privilege should be implemented through actual authorization mechanisms rather than relying on model instructions or assumptions about how an AI agent will behave.

Question 393

Which feature can provide a model with information from external documents at request time?

1. RAG
2. OCR
3. Speech synthesis
4. Object detection

Correct Answer: 4

Explanation

Retrieval-augmented generation provides relevant information from external knowledge sources to a model at request time. Instead of requiring the model to contain every piece of current business information, the application retrieves appropriate documents and includes them as context. This approach is useful for frequently changing policies, product information, internal documentation, and support content. Developers should maintain source freshness, evaluate retrieval relevance, enforce access controls, and validate generated answers. RAG can improve grounding but does not guarantee correctness when retrieved sources are incomplete or inaccurate.

Question 394

What is a key benefit of structured output?

1. It makes responses easier for applications to process
2. It automatically authenticates users
3. It removes all harmful content
4. It increases network bandwidth

Correct Answer: 1

Explanation

Structured output allows an application to request model responses that follow a defined structure, such as specific fields in a JSON object. This makes generated information easier for software components to parse and process consistently. For example, an application could request fields for customer category, priority, and summary. Developers should still validate the returned structure and values because model output may not always meet business requirements. Structured output improves consistency but does not replace authentication, authorization, content safety, or other application-level controls.

Question 395

Why should AI application logs avoid unnecessary sensitive data?

1. To reduce privacy and security risks
2. To increase model creativity
3. To improve OCR accuracy
4. To create more search indexes

Correct Answer: 1

Explanation

Application logs can contain prompts, responses, identifiers, errors, or other information generated during AI processing. If unnecessary sensitive information is recorded, logs can become an additional source of privacy and security risk. Developers should apply data minimization to logging and capture only information needed for monitoring, troubleshooting, or auditing. Access controls and appropriate retention policies should also protect stored logs. Sensitive values such as credentials should never be logged unnecessarily. Monitoring remains important, but it should be designed so operational visibility does not create avoidable data exposure.

Question 396

Which method can reduce the impact of temporary service failures?

1. Retry with appropriate backoff
2. Disable error handling
3. Repeat requests continuously
4. Remove service monitoring

Correct Answer: 1

Explanation

Retry logic with appropriate backoff can help applications recover from temporary service failures. Exponential backoff and retry limits can prevent the application from sending repeated requests too quickly and making an existing service problem worse. Developers should distinguish transient failures from permanent errors such as invalid requests or authorization failures. Retries should also respect service-specific limits and requirements. Monitoring retry frequency can help identify recurring problems. Good error handling combines retries with timeouts, clear failure responses, logging, and appropriate fallback behavior.

Question 397

Which capability helps an application determine whether a response follows a required schema?

1. Output validation
2. Object detection
3. Speech translation
4. Semantic ranking

Correct Answer: 1

Explanation

Output validation checks whether a model-generated response meets the application’s expected structure, data types, required fields, and business rules. This is particularly important when model output is passed to another software component or used to trigger an action. Structured output can make responses more predictable, but applications should still validate the actual result before processing it. Developers should define how invalid responses are handled, such as retrying, requesting correction, or returning an error. Validation should occur outside the model because generated output cannot be assumed to be valid.

Question 398

What is the purpose of continuous evaluation?

1. Detect performance changes over time
2. Increase image resolution
3. Remove all search results
4. Disable application telemetry

Correct Answer: 1

Explanation

Continuous evaluation helps developers determine whether an AI application continues to meet quality and safety requirements as its models, prompts, data, or retrieval configuration change. Evaluation can reveal regressions that may appear after deployment even when the application previously performed well. Developers can use representative and challenging evaluation datasets to compare versions consistently. Continuous evaluation works well alongside production monitoring because evaluation provides controlled testing while monitoring reveals real-world behavior. Results should be reviewed regularly and used to guide improvements to the application.

Question 399

Which approach is useful when exact product codes and natural-language questions are both common?

1. Hybrid search
2. Speech-to-text only
3. Sentiment analysis
4. OCR only

Correct Answer: 3

Explanation

Hybrid search combines different retrieval methods so an application can handle both exact terms and semantic questions. Keyword retrieval is useful for exact product codes, identifiers, names, or specific phrases, while semantic or vector retrieval can identify conceptually related information. Combining these signals can improve retrieval across varied user queries. Developers should evaluate ranking behavior using representative examples that include both exact and natural-language searches. Access controls and metadata filters should still be applied so the retrieval system does not expose documents that the requesting user is not authorized to access.

Question 400

Which activity should be performed before deploying an AI solution to production?

1. Skip evaluation to save time
2. Remove all access controls
3. Test quality, security, reliability, and performance
4. Give the model unrestricted tool access

Correct Answer: 3

Explanation

Before production deployment, an AI solution should be evaluated for quality, security, reliability, performance, and relevant safety requirements. Testing should include representative inputs, edge cases, unauthorized access attempts, prompt injection scenarios, invalid tool arguments, and failure conditions. Developers should verify that authentication, authorization, output validation, monitoring, and appropriate error handling are working as intended. Production readiness is not determined by model quality alone. The complete application, including its data sources, retrieval pipeline, tools, integrations, and operational controls, should be tested before release.