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

 

Question 161

Which capability helps an AI application understand relationships between words based on meaning?

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

Correct Answer: 4

Explanation

Embeddings represent content as numerical vectors that capture semantic relationships. Text with similar meanings can produce vectors that are close together in vector space even when the wording is different. This makes embeddings useful for semantic search, recommendations, clustering, and RAG applications. For example, a search query and a document discussing the same topic can be matched through vector similarity even if they use different terms. Developers should select an appropriate embedding model and evaluate retrieval quality using representative application data.

Question 162

Which Azure service provides capabilities for speech recognition and speech synthesis?

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

Correct Answer: 1

Explanation

Azure AI Speech provides capabilities for working with spoken language and audio. Common scenarios include speech-to-text transcription, text-to-speech synthesis, speech translation, and other voice-related functionality. Developers can use these capabilities to build voice assistants, transcription systems, accessibility features, and multilingual applications. The service is different from Azure AI Language, which focuses primarily on text-based language understanding, and Azure AI Document Intelligence, which analyzes documents. Developers should evaluate speech solutions using realistic audio and language conditions.

Question 163

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

1. Increase model temperature
2. Replace authentication
3. Provide relevant information for generating responses
4. Convert images into audio

Correct Answer: 3

Explanation

A grounding source provides relevant information that a generative AI model can use when producing a response. This information can come from approved documents, databases, search indexes, or other external sources. Grounding is especially useful when the model needs to answer questions about current or organization-specific information. In a RAG architecture, the application retrieves relevant content and includes it as context for the model. Developers should also apply access controls and validate retrieved information to ensure that users receive appropriate and reliable content.

Question 164

Which technique can combine exact keyword matching with semantic similarity?

1. OCR
2. Hybrid search
3. Speech recognition
4. Text-to-speech

Correct Answer: 2

Explanation

Hybrid search combines keyword-based retrieval with semantic or vector-based retrieval. Keyword search can be effective when exact terms, product names, codes, or identifiers are important. Semantic or vector search can find conceptually related content even when the wording differs. Combining these approaches can provide broader and more useful retrieval for enterprise search and RAG applications. Azure AI Search supports hybrid retrieval scenarios. Developers should test different search configurations using realistic queries to determine which combination produces the most relevant results.

Question 165

Why is output validation important when using AI-generated data in an application?

1. It ensures the generated data meets expected requirements
2. It increases microphone volume
3. It creates new user accounts
4. It removes all model limitations

Correct Answer: 1

Explanation

Output validation checks whether AI-generated information meets the application’s expected format, values, and business rules before it is used. This is important because generative models can produce malformed, incomplete, or unexpected information. For example, an application receiving a generated customer record can verify that required fields exist and contain valid values. Validation is particularly important when model output is passed to APIs, databases, or automated workflows. It should be combined with error handling, authorization, and other application-level controls.

Question 166

Which approach is useful for processing a large number of documents overnight?

1. Interactive voice chat
2. Batch processing
3. Manual transcription
4. Real-time conversation

Correct Answer: 2

Explanation

Batch processing is suitable for workloads where many items need to be processed without requiring an immediate response for each individual item. A company could schedule document extraction, classification, indexing, or analysis jobs to run overnight. Batch workflows can improve resource utilization and simplify processing of large datasets. Developers should design these workflows with job tracking, retry handling, error reporting, and service quotas in mind. Results can be stored for later retrieval instead of requiring users to remain connected while processing occurs.

Question 167

Which capability identifies important entities such as people and organizations in text?

1. Named entity recognition
2. Object detection
3. Speech synthesis
4. OCR

Correct Answer: 1

Explanation

Named entity recognition identifies and classifies entities within text. Depending on the service and model, recognized entities can include people, organizations, locations, dates, products, and other categories. This capability can help applications extract structured information from emails, reports, customer messages, and documents. For example, a support system could identify an organization name from an incoming request and use it to route the case. NER is a natural language processing task and is separate from detecting physical objects in images.

Question 168

What should an AI application do when a user asks for information they are not authorized to access?

1. Return the information anyway
2. Enforce authorization and deny unauthorized access
3. Disable authentication
4. Ask another model to provide it

Correct Answer: 2

Explanation

The application should enforce authorization rules and prevent access to information the user is not permitted to view. This is especially important in enterprise AI systems where search indexes may contain documents with different security classifications. Authorization should be implemented using identity and access-control mechanisms rather than relying on the language model to decide whether information is sensitive. Retrieval can also use security filters to prevent unauthorized documents from becoming model context. These controls help reduce the risk of confidential information being disclosed.

Question 169

What is the purpose of a prebuilt AI model?

1. Provide a ready-made capability for a common task
2. Store user passwords
3. Replace network security
4. Increase document file size

Correct Answer: 1

Explanation

A prebuilt AI model provides an existing capability designed for a commonly encountered task. Examples can include document analysis, invoice processing, receipt extraction, language analysis, and other supported scenarios. Using a prebuilt model can reduce development time because the application does not need to create and train a specialized model from scratch for every use case. Developers should still evaluate the model using representative data because document layouts, language, terminology, and input quality can affect performance.

Question 170

Which Azure AI capability can analyze the visual content of an image?

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

Correct Answer: 2

Explanation

Azure AI Vision provides capabilities for analyzing visual content. Depending on the supported features and models, applications can use it for tasks such as image analysis, OCR, object detection, and other computer vision scenarios. These capabilities can help applications understand images without requiring developers to build every computer vision model themselves. Developers should select the specific vision capability based on the task, such as extracting text versus locating objects. Performance should be tested with images representative of the production environment.

Question 171

What does least-privilege access require?

1. Give every application administrator permissions
2. Allow unrestricted data access
3. Grant only the permissions required for the task
4. Remove all authentication

Correct Answer: 3

Explanation

Least privilege means granting a user, application, or service only the permissions necessary to perform its required tasks. This limits the potential impact of compromised credentials, configuration mistakes, or unintended actions. For example, an application that only needs to read documents should not automatically receive permission to delete them. Azure role-based access control and managed identities can help implement least-privilege access. Permissions should be reviewed periodically because applications and business requirements can change over time.

Question 172

Which process converts an audio recording into searchable text?

1. Speech-to-text
2. Text-to-speech
3. Image classification
4. Semantic ranking

Correct Answer: 1

Explanation

Speech-to-text converts spoken language in an audio source into written text. The resulting transcript can then be searched, summarized, analyzed, translated, or stored for later use. Common applications include meeting transcription, call-center analysis, voice commands, and accessibility solutions. Accuracy depends on factors such as audio quality, background noise, speaker characteristics, vocabulary, and supported language settings. Developers should test the speech recognition system with realistic recordings before deploying it in production, especially when accurate transcription is important.

Question 173

Why can a reranker improve a search-based AI solution?

1. It translates documents
2. It creates audio files
3. It places the most relevant retrieved results higher
4. It manages user passwords

Correct Answer: 3

Explanation

A reranker can improve search results by evaluating retrieved candidates and ordering them according to their relevance to the user’s query. Initial retrieval is often optimized for speed and broad candidate selection, so some results may be only partially relevant. Reranking can provide a more refined ordering before the results are passed to a generative model. This is particularly useful in RAG applications where the top retrieved documents become model context. Developers should measure retrieval and answer quality to determine whether reranking improves the overall solution.

Question 174

What is the primary purpose of prompt engineering?

1. Designing effective instructions and context for a model
2. Encrypting documents
3. Managing Azure subscriptions
4. Increasing storage capacity

Correct Answer: 1

Explanation

Prompt engineering involves designing instructions, context, examples, and constraints that guide a generative model toward the desired behavior. A well-designed prompt can specify the task, expected response format, relevant information, and limitations. Developers can test different prompts and compare their results using representative evaluation data. Prompt engineering can improve consistency and usefulness, but it does not replace security mechanisms. Authentication, authorization, validation, content filtering, and access controls are still required when building production AI applications.

Question 175

Which capability is most appropriate for extracting information from an invoice?

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

Correct Answer: 4

Explanation

Azure AI Document Intelligence provides document-analysis capabilities that can extract information from invoices and other supported business documents. Depending on the model, it can identify fields such as invoice numbers, dates, vendor information, totals, and line items. This can automate manual invoice-processing workflows and reduce the need for users to enter information by hand. Developers should validate extracted values because document quality, layouts, and formatting can vary. Testing should include representative invoices from the sources expected in production.

Question 176

What can a managed identity provide to an Azure AI application?

1. Password storage in source code
2. Identity-based authentication to supported resources
3. Public access to every service
4. Automatic model training

Correct Answer: 2

Explanation

A managed identity provides an Azure resource with an identity that can be used to authenticate to supported services. This allows applications to access resources without storing long-term credentials directly in source code or configuration. Developers can combine managed identities with role-based access control to grant the required permissions. This approach can simplify credential management and reduce exposure of secrets. The identity should still be granted only the permissions necessary for the application’s workload according to least-privilege principles.

Question 177

Which search approach is useful when users ask questions using different wording from the source documents?

1. Vector or semantic search
2. Exact filename matching only
3. OCR only
4. Audio transcription

Correct Answer: 1

Explanation

Vector and semantic search can identify relevant information even when the user’s wording differs from the wording in the source documents. Vector search compares semantic representations, while semantic search uses language understanding to improve relevance. For example, a user asking about cancelling a service may retrieve documents that discuss termination even if the word “cancel” is not present. This capability is valuable for natural-language enterprise search and RAG applications. Developers should evaluate results using real user queries and representative content.

Question 178

What is the purpose of monitoring token usage in a generative AI application?

1. To detect image quality problems
2. To understand consumption and manage costs
3. To create user passwords
4. To replace authorization

Correct Answer: 2

Explanation

Monitoring token usage helps developers understand how much model input and output the application consumes. Token consumption can affect operating costs, latency, and service quotas. Tracking usage can reveal unexpectedly large prompts, excessive retrieved context, or unusually long responses. Developers can then optimize chunk sizes, retrieval limits, prompts, and model selection when appropriate. Token monitoring should be combined with broader telemetry covering errors, latency, throughput, and quality so that cost optimization does not unintentionally reduce application performance.

Question 179

Which practice can help protect an AI application from excessive traffic?

1. Removing all request limits
2. Sharing API credentials
3. Rate limiting
4. Disabling monitoring

Correct Answer: 3

Explanation

Rate limiting controls the number or frequency of requests that an application accepts from a user, client, or service. It can help protect AI workloads from accidental traffic spikes, excessive consumption, and certain forms of abuse. Rate limits can also help applications operate within service quotas and control costs. When a limit is reached, the application should provide an appropriate response and may use controlled retry behavior where suitable. Limits should be based on expected workload, service capabilities, and application requirements.

Question 180

Why should an AI solution be evaluated before production deployment?

1. To identify quality, reliability, and safety issues
2. To remove authentication
3. To guarantee that every response is correct
4. To avoid collecting any metrics

Correct Answer: 1

Explanation

Evaluation helps developers identify quality, reliability, safety, and performance issues before an AI solution is exposed to production users. Testing should use representative data and scenarios that reflect the application’s expected workload. Depending on the solution, evaluation can examine accuracy, relevance, groundedness, response format, latency, and safety behavior. Evaluation cannot guarantee that every future response will be correct, but it provides evidence about how the system performs against defined requirements. Developers should continue monitoring and evaluating important changes after deployment.