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

 

Question 261

What is the main purpose of a vector database or vector index in an AI application?

1. Store passwords
2. Store and search embeddings
3. Convert speech to text
4. Generate invoices

Correct Answer: 2

Explanation

A vector database or vector index stores embeddings and enables similarity-based searches over those numerical representations. When content is converted into embeddings, the resulting vectors can be indexed and later compared with a query embedding. This allows applications to retrieve content based on semantic similarity rather than exact keyword matches. Vector search is commonly used in RAG, recommendation, and knowledge-retrieval scenarios. Developers should ensure that the embedding model, vector dimensions, indexing configuration, and similarity method are compatible with the application’s requirements.

Question 262

Which Azure service is appropriate for analyzing the structure of invoices?

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

Correct Answer: 3

Explanation

Azure AI Document Intelligence is designed to analyze documents and extract information from their content and structure. For invoices, it can help identify fields and other relevant elements from supported document formats. This makes it useful for automating business processes that previously required manual data entry. Developers should validate extracted information before sending it to financial or business systems. Testing should include different invoice layouts, scan qualities, and document variations because extraction results can differ depending on the input.

Question 263

Which method helps a model use information from an external knowledge base?

1. Retrieval-augmented generation
2. Image segmentation
3. Speech synthesis
4. Language detection

Correct Answer: 1

Explanation

Retrieval-augmented generation allows a language model to use information retrieved from an external knowledge base when generating a response. Instead of relying only on information encoded during model training, the application searches approved sources and provides relevant results as context. This is useful for company policies, product documentation, support articles, and other information that may change over time. The effectiveness of RAG depends on retrieval quality, source freshness, chunking, ranking, access controls, and the model’s ability to use the supplied context appropriately.

Question 264

What does least privilege mean in an AI application?

1. Giving users every available permission
2. Giving identities only the permissions they require
3. Disabling all authorization
4. Sharing one administrator account

Correct Answer: 2

Explanation

Least privilege means granting users, applications, agents, and services only the permissions necessary to perform their intended tasks. This reduces the potential impact of mistakes, compromised credentials, or malicious input. For example, an AI agent that only needs to read customer records should not automatically receive permission to delete them. Developers should regularly review assigned permissions and remove unnecessary access. Least privilege should be combined with authentication, authorization, monitoring, and appropriate validation to provide stronger protection for AI workloads and connected business systems.

Question 265

Which capability allows an AI application to process both a photograph and a written question?

1. Multimodal AI
2. Rate limiting
3. Key phrase extraction
4. Batch indexing

Correct Answer: 1

Explanation

Multimodal AI allows supported models to process more than one type of input, such as text and images. A user might provide a photograph of equipment and ask a question about what appears in the image. The model can use information from both inputs when generating its response. Multimodal capabilities are useful for visual question answering, document understanding, product analysis, and similar scenarios. Developers should verify the model’s supported input types, image limitations, token or size constraints, and expected performance before deploying a multimodal workflow.

Question 266

What is the purpose of chunk overlap in document processing?

1. Preserve context that crosses chunk boundaries
2. Remove all document metadata
3. Convert text into speech
4. Increase user permissions

Correct Answer: 1

Explanation

Chunk overlap repeats a small portion of content between adjacent chunks so that information near a boundary is less likely to lose important context. Without overlap, a sentence or related concept may be divided between two chunks, making retrieval less useful. The appropriate overlap depends on document structure and chunk size. Excessive overlap can increase the amount of indexed content and token usage, so developers should balance context preservation against storage and processing costs. Retrieval evaluation can help determine whether the selected overlap improves results.

Question 267

Which service can identify sentiment in customer comments?

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

Correct Answer: 2

Explanation

Azure AI Language provides sentiment analysis capabilities for supported text inputs. An application can use sentiment analysis to examine customer comments, reviews, survey responses, or support messages and identify sentiment-related information. The results can support analytics, routing, or prioritization workflows. Sentiment analysis should not be treated as a perfect interpretation of human emotion because context, sarcasm, ambiguity, and specialized language can affect results. Developers should evaluate the service using representative examples from the target application before relying on the results operationally.

Question 268

Why should an AI agent validate tool inputs before execution?

1. To ensure arguments meet expected formats and rules
2. To increase model temperature
3. To remove authentication
4. To create larger embeddings

Correct Answer: 1

Explanation

Tool input validation ensures that arguments generated by an AI model meet expected formats, types, ranges, and business rules before an external operation is performed. For example, an application can verify that an account identifier exists and that a requested transaction amount is within allowed limits. This protects downstream systems from malformed or unexpected requests. Validation should be combined with authorization because a valid argument does not necessarily mean the requesting user is permitted to perform the action. Sensitive operations may also require human approval.

Question 269

Which search approach is most useful when exact product codes must be matched?

1. Keyword search
2. Speech recognition
3. Sentiment analysis
4. Image classification

Correct Answer: 1

Explanation

Keyword search is useful when exact terms, identifiers, product codes, names, or other precise strings need to be matched. Semantic or vector search may retrieve conceptually similar content, but exact identifiers often require lexical matching to ensure the intended value is found. In many enterprise applications, keyword search can be combined with semantic search through hybrid search. Developers should select the retrieval approach according to the data and query patterns. Testing with real product codes and representative queries can reveal which configuration provides reliable results.

Question 270

What is a major purpose of Azure Key Vault in an AI solution?

1. Securely manage supported secrets and keys
2. Generate speech transcripts
3. Analyze photographs
4. Create search embeddings

Correct Answer: 1

Explanation

Azure Key Vault can help applications securely manage supported secrets, keys, and certificates instead of placing sensitive values directly in source code or application files. This can reduce the risk of accidental credential exposure and supports centralized access management. Applications can use appropriate identities and permissions to access required secrets. Developers should also avoid writing sensitive values to logs and should establish appropriate rotation and access-review practices. Key Vault is one component of a broader security strategy that should also include authentication, authorization, monitoring, and data protection.

Question 271

What should an AI application do if a retrieved document is outside the user’s permissions?

1. Return the document anyway
2. Exclude the unauthorized content
3. Disable authentication
4. Store the document in the prompt

Correct Answer: 2

Explanation

Unauthorized content should be excluded from retrieval results and should not be supplied to the model for generation. Access controls must be enforced before sensitive information reaches the response-generation stage. Depending on the architecture, permissions can be applied through document-level security, identity-aware filtering, metadata filters, or other supported mechanisms. Simply instructing the model not to mention private information is not sufficient because the model may still receive the restricted content. Developers should test retrieval with users having different permissions to verify isolation.

Question 272

Which technique can reduce irrelevant retrieved documents?

1. Better filtering and ranking
2. Removing all search indexes
3. Increasing every chunk size indefinitely
4. Disabling evaluation

Correct Answer: 1

Explanation

Filtering and ranking can reduce the amount of irrelevant information returned by a retrieval system. Metadata filters can restrict results based on attributes such as date, department, language, or access level, while ranking and reranking can place more relevant documents higher in the result set. Developers can also improve chunking, embeddings, and query formulation. The goal is not simply to retrieve more documents but to provide useful context to the model. Retrieval performance should be measured using representative queries and expected relevant results.

Question 273

Which capability can identify a person’s name or organization in text?

1. Named entity recognition
2. Text-to-speech
3. Object detection
4. Vector indexing

Correct Answer: 1

Explanation

Named entity recognition, or NER, identifies entities within text and categorizes them according to supported entity types. Examples can include people, organizations, locations, dates, and other recognized categories. NER can help applications extract structured information from emails, reports, customer messages, and documents. The extracted entities can then support search, classification, analytics, or downstream workflows. Developers should test NER with representative domain-specific content because unusual names, specialized terminology, formatting, and language differences can affect recognition quality.

Question 274

What is the purpose of grounding in a generative AI application?

1. Connect responses to relevant supporting information
2. Increase microphone volume
3. Remove authorization checks
4. Create user accounts

Correct Answer: 1

Explanation

Grounding connects a model’s response to information provided by trusted sources, such as retrieved documents or approved knowledge bases. This can help reduce unsupported responses and provide the model with information that may not be available in its training data. Grounding is commonly implemented through retrieval-based architectures. Developers should evaluate whether the retrieved information actually supports the generated response rather than assuming that retrieval alone guarantees accuracy. Source citations, retrieval evaluation, and response validation can provide additional ways to assess grounding quality.

Question 275

Which processing approach is suitable for thousands of independent documents that do not require immediate results?

1. Batch processing
2. Real-time speech recognition
3. Interactive chat only
4. Manual processing only

Correct Answer: 1

Explanation

Batch processing is suitable when many independent items can be processed without requiring an immediate response for each one. For example, an organization may need to analyze thousands of archived documents overnight. Batch workflows can improve operational efficiency by processing large workloads systematically and can sometimes reduce the overhead associated with individual real-time requests. Developers should design appropriate error handling, progress tracking, retry behavior, and result storage. Batch processing is less appropriate when users require an immediate response to every request.

Question 276

What can query rewriting improve in a RAG application?

1. The form of a query sent to the retrieval system
2. The user’s monitor resolution
3. The microphone battery
4. The document’s original author

Correct Answer: 1

Explanation

Query rewriting can modify a user’s original question into a form that may be more effective for retrieval. For example, it can clarify terminology, expand ambiguous references, or create a search-friendly formulation. This can help when the wording used by users differs from the language found in indexed documents. However, rewriting must preserve the original intent. Developers should evaluate rewritten queries because an incorrect transformation can retrieve unrelated information. Query rewriting can be combined with semantic, vector, keyword, or hybrid retrieval depending on the application.

Question 277

Which capability converts text into an audio response?

1. Text-to-speech
2. Speech-to-text
3. Named entity recognition
4. Vector search

Correct Answer: 1

Explanation

Text-to-speech converts written text into synthesized spoken audio. It can be used to create voice responses for assistants, accessibility applications, automated announcements, and other voice-enabled experiences. Developers can select an appropriate supported language and voice and then generate audio from the application’s text output. The application should consider latency, pronunciation, voice quality, and the user’s environment. Text-to-speech is different from speech-to-text, which converts spoken audio into written text and is commonly used for transcription.

Question 278

Which practice helps protect an AI application from excessive model usage?

1. Usage monitoring and rate limiting
2. Removing request controls
3. Granting unlimited access
4. Disabling telemetry

Correct Answer: 1

Explanation

Usage monitoring and rate limiting can help control excessive model usage. Monitoring provides visibility into request volume, token consumption, latency, and other operational patterns, while rate limits restrict how frequently requests can be submitted. Together, these controls can help manage costs, protect service capacity, and identify unusual activity. Developers should configure limits according to legitimate workload requirements and handle limit responses appropriately. Monitoring should also be reviewed regularly so that unexpected increases in usage can be investigated before they create operational or financial problems.

Question 279

What should be done before an AI application accesses a protected API?

1. Verify authentication and authorization
2. Remove all permissions
3. Trust the model automatically
4. Disable logging

Correct Answer: 1

Explanation

Authentication and authorization should be verified before an AI application accesses a protected API. Authentication establishes the identity making the request, while authorization determines whether that identity has permission to perform the requested action. The AI model itself should not be treated as an authorization mechanism. Developers should validate tool arguments and enforce permissions at the application or service boundary. Sensitive API operations may also require additional business-rule validation or human approval. These controls help prevent agents from performing unauthorized actions even when they generate technically valid requests.

Question 280

Why should an AI application maintain monitoring after deployment?

1. To identify errors, performance changes, and unusual behavior
2. To eliminate all testing
3. To guarantee perfect responses
4. To replace access controls

Correct Answer: 1

Explanation

Monitoring provides visibility into how an AI application behaves after deployment. Developers can track errors, latency, request volume, token consumption, service availability, and other operational signals. Depending on the application, monitoring can also help identify changes in retrieval quality, safety issues, or unusual usage patterns. Continuous monitoring allows teams to investigate problems that may not have appeared during initial testing. It does not guarantee perfect responses or replace security controls. Monitoring works best when combined with evaluation, logging policies, alerting, and appropriate operational procedures.