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

 

Question 241

What is the purpose of an embedding?

1. Store user passwords
2. Convert speech into audio
3. Represent content numerically for similarity comparison
4. Detect document tables

Correct Answer: 3

Explanation

An embedding represents content as a numerical vector that captures aspects of its meaning. Text, queries, or other supported content can be converted into embeddings and compared using similarity measures. This makes embeddings useful for semantic search, recommendation systems, clustering, and RAG applications. When documents are indexed, their embeddings can be stored in a vector index. A user’s query can then be embedded and compared with stored vectors to identify semantically relevant content, even when the wording is not identical.

Question 242

Which service is commonly used for speech-to-text conversion?

1. Azure AI Speech
2. Azure AI Vision
3. Azure AI Search
4. Azure AI Language

Correct Answer: 1

Explanation

Azure AI Speech provides speech-to-text capabilities that convert spoken audio into written text. This can support applications such as meeting transcription, voice assistants, call analysis, and accessibility features. Recognition quality can vary depending on factors such as background noise, speaker accents, microphone quality, and specialized terminology. Developers should test the service using audio representative of the production environment. Once speech has been converted to text, the resulting transcript can also be processed by other AI services for summarization, translation, sentiment analysis, or information extraction.

Question 243

Why should RAG applications use approved knowledge sources?

1. To increase screen brightness
2. To reduce microphone volume
3. To provide trusted and relevant information
4. To disable vector search

Correct Answer: 3

Explanation

Approved knowledge sources help ensure that a RAG application retrieves information from sources that the organization considers relevant and trustworthy. If unapproved or unreliable content is indexed, the model may use that information when generating responses. Developers should identify appropriate sources, establish update processes, and apply access controls where necessary. Source quality directly affects grounding quality because the model can only use the information provided to it. Applications should also monitor retrieval results and periodically review whether indexed sources remain accurate and appropriate.

Question 244

Which search method can combine exact keyword matches with semantic similarity?

1. OCR
2. Hybrid search
3. Speech recognition
4. Sentiment analysis

Correct Answer: 2

Explanation

Hybrid search combines keyword-based retrieval with semantic or vector-based retrieval. Keyword matching can identify exact terms, identifiers, or phrases, while vector search can identify content with related meaning even when the wording differs. Combining these approaches can improve retrieval coverage in enterprise knowledge systems. The retrieved results can then be ranked or reranked before being supplied to a generative model. Developers should evaluate the search configuration using representative queries because the optimal balance between lexical and semantic signals depends on the application’s content and requirements.

Question 245

A developer needs to extract text from a scanned invoice. Which capability is relevant?

1. OCR
2. Text-to-speech
3. Semantic ranking
4. Rate limiting

Correct Answer: 1

Explanation

OCR, or optical character recognition, extracts machine-readable text from images and scanned documents. For an invoice, OCR can identify the visible text that may then be processed to extract fields such as invoice numbers, dates, vendor names, and amounts. Document Intelligence can provide more specialized document-analysis capabilities when structure and fields are required. Developers should consider scan quality, document orientation, handwriting, and layout complexity because these factors can affect extraction accuracy. Testing with representative invoices helps determine whether additional processing or validation is required.

Question 246

What should a tool description clearly explain?

1. The application’s billing address
2. The tool’s purpose and expected inputs
3. The user’s password
4. The model’s training data

Correct Answer: 2

Explanation

A tool description should clearly explain what the tool does and what information it expects as input. Good descriptions help a model determine when a particular tool is appropriate and how to construct a request. Developers should specify important parameters, expected formats, and relevant limitations where appropriate. However, descriptions should not be treated as a security mechanism. The application must still validate tool arguments and enforce authentication and authorization. Clear tool descriptions can improve agent reliability while strong application controls protect the underlying operations.

Question 247

Which Azure AI Language capability identifies important terms in a document?

1. Key phrase extraction
2. Object detection
3. Speech synthesis
4. Image segmentation

Correct Answer: 1

Explanation

Key phrase extraction identifies important concepts or terms within a body of text. It can help applications quickly identify major topics in customer feedback, reports, reviews, emails, or other documents. The extracted phrases can be used for categorization, search, indexing, analytics, or summarization workflows. Key phrase extraction is different from named entity recognition, which focuses on identifying entities such as people, organizations, locations, or dates. Developers should evaluate results with representative content because important terms can vary significantly between domains.

Question 248

What should be applied when users can access different document collections?

1. Access control
2. Higher temperature
3. Larger image resolution
4. Speech synthesis

Correct Answer: 1

Explanation

Access control should be applied when different users have different permissions for document collections. A RAG application should ensure that retrieval results respect the permissions associated with the requesting identity. This can involve document-level permissions, metadata filters, role-based access control, or other supported authorization mechanisms. Simply hiding restricted documents from the user interface is not sufficient if the backend can still retrieve them. Developers should test authorization boundaries carefully to ensure users cannot obtain protected information through alternate queries or agent workflows.

Question 249

Which computer vision capability can identify objects within an image?

1. Object detection
2. Language detection
3. Key phrase extraction
4. Speech translation

Correct Answer: 1

Explanation

Object detection identifies objects within an image and can provide information about their locations, commonly through bounding regions. This is useful in scenarios such as inventory analysis, safety monitoring, retail applications, and visual inspection. Object detection differs from simple image classification because it can identify individual objects and their positions rather than assigning only an overall category to the image. Developers should evaluate detection accuracy using representative images because object size, lighting, image quality, occlusion, and background conditions can influence results.

Question 250

Why might a developer choose a managed Azure AI service instead of building a model from scratch?

1. To avoid all testing
2. To access ready-to-use AI capabilities
3. To eliminate authentication
4. To guarantee perfect results

Correct Answer: 2

Explanation

Managed Azure AI services provide ready-to-use AI capabilities that can reduce the development effort required to build common AI features from scratch. Depending on the service, developers can use capabilities for language analysis, speech processing, computer vision, document analysis, search, or generative AI. Managed services can simplify infrastructure management and integration, but they still require appropriate configuration, security, evaluation, and monitoring. Developers should select a service based on the application’s requirements rather than assuming a managed service automatically provides perfect accuracy or suitability.

Question 251

What is metadata in a search index?

1. Additional information describing indexed content
2. A replacement for authentication
3. A speech-recognition model
4. A type of microphone

Correct Answer: 1

Explanation

Metadata is additional information associated with indexed content that can help applications organize, filter, or interpret search results. Examples include document dates, categories, departments, language, product identifiers, or access classifications. Metadata can be especially useful in RAG systems because retrieval can be restricted according to specific attributes. Developers should design metadata fields carefully and keep them accurate as source documents change. Metadata alone does not provide security unless it is correctly integrated with authorization and filtering logic enforced by the application or search system.

Question 252

Which process converts spoken audio into written words?

1. Text-to-speech
2. Speech-to-text
3. Semantic search
4. OCR

Correct Answer: 2

Explanation

Speech-to-text converts spoken audio into written text. It is commonly used for meeting transcription, voice interfaces, call-center analysis, accessibility applications, and other audio-processing scenarios. The resulting transcript can be passed to additional services for summarization, translation, sentiment analysis, or information extraction. Recognition accuracy depends on factors such as audio quality, background noise, speaker characteristics, and language. Developers should evaluate the service using recordings that resemble expected production conditions and should handle cases where recognition confidence or transcript quality is insufficient.

Question 253

What should be measured when evaluating retrieval quality?

1. Whether relevant information is returned
2. Only the application’s screen size
3. Only the number of users
4. Microphone battery level

Correct Answer: 1

Explanation

Retrieval quality should be evaluated by determining whether the search system returns information that is relevant to the user’s query. Developers can create representative test questions and identify which documents or passages should be retrieved for each one. Metrics and human review can then help assess relevance, ranking, coverage, and other retrieval characteristics. Poor retrieval can lead to weak RAG responses even when the language model itself performs well. Improving chunking, embeddings, filtering, query rewriting, or ranking may help address retrieval weaknesses.

Question 254

What is a retry policy used for in an AI application?

1. Handling appropriate temporary failures
2. Removing all security controls
3. Increasing image brightness
4. Changing document ownership

Correct Answer: 1

Explanation

A retry policy allows an application to attempt a failed operation again when the failure may be temporary. Examples can include transient service errors, temporary network problems, or certain capacity-related conditions. Retries should be controlled rather than performed continuously because excessive retries can increase traffic and costs. Developers can use techniques such as delays and exponential backoff where appropriate. Not every error should be retried; invalid requests, authorization failures, or permanent configuration problems generally require a different handling strategy.

Question 255

Why can human approval be useful for high-impact agent actions?

1. It allows a person to review the proposed action
2. It increases vector dimensions
3. It replaces model evaluation
4. It automatically corrects all prompts

Correct Answer: 1

Explanation

Human approval provides an opportunity for a person to review an AI agent’s proposed action before an important operation is completed. This can be useful for actions involving financial transactions, account changes, sensitive data, or other consequential operations. The approval interface should provide enough context for the reviewer to understand what will happen. Human approval does not replace authentication, authorization, validation, or monitoring. Instead, it adds an additional control for situations where automated execution could have significant consequences.

Question 256

Which factor should influence model selection?

1. Required quality, latency, capability, and cost
2. Only the model’s display name
3. Only the user’s keyboard
4. Only the document color

Correct Answer: 1

Explanation

Model selection should consider the application’s required quality, supported capabilities, latency, cost, context requirements, and workload characteristics. A model that performs well on complex reasoning may be unnecessary for a simple classification or extraction task. Conversely, a smaller model may not meet the quality requirements of a complex workload. Developers should evaluate candidate models using representative application data rather than selecting solely by reputation or size. Operational factors such as throughput, availability, token usage, and deployment requirements can also influence the final configuration.

Question 257

Which approach can retrieve information using both meaning and exact terms?

1. Hybrid search
2. OCR
3. Text-to-speech
4. Sentiment analysis

Correct Answer: 1

Explanation

Hybrid search combines semantic or vector-based retrieval with keyword-based retrieval. Semantic retrieval helps identify content that is conceptually related to a query, while keyword retrieval is useful when exact words, codes, names, or identifiers matter. Using both signals can improve search coverage in many enterprise applications. The results can then be ranked or reranked before being supplied to a generative model. Developers should evaluate hybrid search using actual queries and documents because different workloads may require different ranking configurations and weighting between search methods.

Question 258

What can help protect an AI agent’s tools from malicious instructions?

1. Tool restrictions and input validation
2. Removing authentication
3. Granting unrestricted permissions
4. Disabling all monitoring

Correct Answer: 1

Explanation

Tool restrictions and input validation can reduce the risk that malicious or unexpected instructions cause an AI agent to perform unsafe operations. Each tool should receive only the permissions necessary for its purpose, and generated arguments should be checked against expected schemas and business rules. Authentication and authorization should also be enforced independently of the model. For sensitive operations, human approval can provide an additional safeguard. These controls are particularly important because agent workflows may process untrusted user input or retrieved external content.

Question 259

Why is data residency important for some AI applications?

1. Organizations may have requirements about where data is stored or processed
2. It determines microphone volume
3. It automatically improves model reasoning
4. It replaces data encryption

Correct Answer: 1

Explanation

Data residency refers to requirements or policies concerning the geographic location where data is stored or processed. Some organizations may have legal, regulatory, contractual, or internal requirements governing the locations in which certain information can be handled. When designing an AI solution, developers should understand the applicable requirements and verify that selected services and configurations support them. Data residency does not replace other security controls. Applications may still require encryption, access control, retention policies, data minimization, and appropriate monitoring.

Question 260

Why should an AI application be evaluated continuously after deployment?

1. Model, data, and usage changes can affect performance
2. Evaluation is only useful before coding
3. Production systems never change
4. Monitoring automatically fixes every problem

Correct Answer: 1

Explanation

Continuous evaluation helps identify changes in AI application quality after deployment. Model updates, prompt changes, new documents, changing user behavior, retrieval configuration changes, and new edge cases can all affect performance over time. Developers can maintain evaluation datasets and monitor metrics such as accuracy, relevance, groundedness, safety, latency, and error rates. Periodic evaluation can reveal regressions that may not appear during initial testing. Combining continuous evaluation with monitoring and user feedback helps teams identify problems and make controlled improvements to production AI systems.