Microsoft Azure AI Engineer Associate AI-102 AI-102 Questions 226–300 | Page 4/11

226

MCQeasy

Refer to the exhibit. You are reviewing the configuration of an Azure OpenAI Service resource. The resource is configured with customer-managed keys for encryption. What is the primary benefit of this configuration?

A.Enhanced control over data encryption keys

B.Simplified deployment process

C.Improved model performance

D.Reduced operational costs

AnswerA

Customer-managed keys let you supply and rotate your own Key Vault key, so Microsoft cannot decrypt the data. This satisfies the requirement for control over the encryption key lifecycle, rather than relying on Microsoft-managed keys.

Why this answer

Customer-managed keys (CMK) allow you to control and manage the encryption keys used to protect your data at rest in Azure OpenAI Service. This provides enhanced control over who can access the keys, when they are rotated, and how they are stored, which is critical for meeting compliance and security requirements. The primary benefit is not performance, cost, or deployment simplicity, but rather the ability to enforce your own key lifecycle and access policies.

Exam trap

The trap here is that candidates often confuse customer-managed keys with platform-managed keys, assuming the primary benefit is cost savings or performance gains, when in reality the core advantage is granular control over encryption key governance and compliance.

How to eliminate wrong answers

Option B is wrong because customer-managed keys add complexity to the deployment process (you must create and manage a Key Vault, set permissions, and configure key rotation), not simplify it. Option C is wrong because encryption keys have no impact on model inference speed or accuracy; performance is determined by model size, token limits, and compute resources. Option D is wrong because CMK typically increases operational costs due to the need for additional Key Vault resources, key management overhead, and potential charges for key operations.

Full explanation →

227

MCQmedium

You are building a solution to automatically tag images uploaded to an Azure Storage blob container using Azure AI Vision. The solution must process images as soon as they are uploaded. Which service should you use to trigger the image analysis?

A.Azure Functions with a timer trigger

B.Azure Event Grid with an Azure Function trigger

C.Azure Batch with a job schedule

D.Azure Logic Apps with a recurrence trigger

AnswerB

Event Grid delivers blob-created events the instant an upload completes, and its Azure Function trigger invokes analysis immediately. This satisfies the requirement to process images as soon as they are uploaded, unlike polling or scheduled batch approaches.

Why this answer

Azure Event Grid is the correct choice because it provides a serverless event-driven architecture that can react to blob storage events (e.g., BlobCreated) in near real-time. By configuring an Event Grid subscription on the storage account, you can trigger an Azure Function that uses Azure AI Vision to analyze the image as soon as it is uploaded, without polling or scheduled checks.

Exam trap

The trap here is that candidates often confuse scheduled triggers (timer/recurrence) with event-driven triggers, assuming any automated trigger will work, but the requirement for 'as soon as they are uploaded' demands an event-driven service like Event Grid, not a polling-based scheduler.

How to eliminate wrong answers

Option A is wrong because a timer trigger runs on a fixed schedule (e.g., every 5 minutes), which introduces latency and cannot react immediately to uploads; it would require polling the container for new blobs. Option C is wrong because Azure Batch is designed for large-scale parallel compute jobs with job schedules, not for real-time event-driven triggers on individual blob uploads. Option D is wrong because a recurrence trigger in Logic Apps also runs on a schedule, not event-driven, and would similarly require polling, missing the immediate processing requirement.

Full explanation →

228

MCQeasy

You need to monitor usage and costs of your Azure OpenAI Service deployments. Which Azure tool should you use?

A.Azure Cost Management + Billing

B.Azure Monitor

C.Azure Service Health

D.Azure Advisor

AnswerA

Azure Cost Management + Billing provides native cost analysis and budgets for Azure OpenAI Service resources, satisfying the requirement to monitor both usage and spend. It aggregates consumption data per deployment, letting you track token-based charges and set alerts, which dedicated monitoring tools like Azure Monitor cannot do for billing.

Why this answer

Azure Cost Management + Billing is the correct tool for monitoring usage and costs of Azure OpenAI Service deployments because it provides detailed cost analysis, budget tracking, and usage reports across all Azure services. It allows you to set budgets, create cost alerts, and analyze spending patterns specifically for OpenAI model deployments, including per-model and per-region cost breakdowns.

Exam trap

The trap here is that candidates often confuse Azure Monitor (which tracks performance metrics like token usage and latency) with cost monitoring, but Azure Monitor does not provide billing data or cost analysis, which is the specific requirement in this question.

How to eliminate wrong answers

Option B (Azure Monitor) is wrong because it focuses on performance metrics, logs, and alerts for application health and resource utilization, not on cost tracking or billing data. Option C (Azure Service Health) is wrong because it monitors service-level issues, outages, and planned maintenance across Azure services, not usage or cost metrics. Option D (Azure Advisor) is wrong because it provides best-practice recommendations for optimizing cost, performance, and reliability, but it does not directly monitor or report on actual usage and costs in real time.

Full explanation →

229

MCQeasy

You need to translate a large volume of documents from English to French while preserving the original formatting. Which Azure service should you use?

A.Azure AI Language

B.Azure AI Translator (Document Translation)

C.Custom Translator in Azure AI Translator

D.Azure OpenAI Service with GPT-4

AnswerB

Document Translation is the Translator feature that translates whole files while retaining their original layout, tables and formatting. The plain text translation API returns only translated strings, so it cannot preserve document structure across a large batch.

Why this answer

Azure AI Translator's Document Translation feature is specifically designed to translate entire documents while preserving the original structure, layout, and formatting (e.g., tables, headers, bullet points). It supports batch processing of large volumes of documents and maintains the source file's format in the translated output, making it the correct choice for this requirement.

Exam trap

Microsoft often tests the distinction between text translation and document translation, leading candidates to choose Azure AI Language (option A) because they confuse its general language processing capabilities with the specific document translation feature, or to pick Custom Translator (option C) thinking customization is required for formatting preservation.

How to eliminate wrong answers

Option A is wrong because Azure AI Language provides text analytics and language understanding capabilities (e.g., sentiment analysis, key phrase extraction) but does not include document-level translation with formatting preservation. Option C is wrong because Custom Translator is a customization feature within Azure AI Translator that allows you to build custom translation models for domain-specific terminology, but it does not directly handle document translation or formatting preservation; it is used to improve translation quality, not to process documents. Option D is wrong because Azure OpenAI Service with GPT-4 is a general-purpose language model that can translate text but is not optimized for batch document translation with formatting preservation; it lacks native support for handling document structures and may produce inconsistent formatting output.

Full explanation →

230

Multi-Selectmedium

You are building a knowledge mining solution using Azure AI Search with AI enrichment. Which TWO built-in skills can be used to extract information from images embedded in documents?

Select 2 answers

A.Entity Recognition skill

B.Image Analysis skill

C.OCR skill

D.Key Phrase Extraction skill

E.Text Translation skill

AnswersB, C

The Image Analysis skill extracts visual features and generates descriptions or tags from image content, and can also produce text via its OCR capability. It satisfies the requirement to extract information from images embedded within documents during AI enrichment.

Why this answer

The Image Analysis skill (B) is correct because it invokes the Computer Vision service to extract visual features from embedded images, such as descriptions, tags, and celebrity or landmark detection, which is exactly the kind of image-derived information a knowledge mining pipeline needs. The OCR skill (C) is also correct because it extracts printed and handwritten text from image files (including images embedded in documents), producing text that can be mapped into the search index. The Entity Recognition skill (A) operates on text to identify entities like people, places, and organizations, not on image content.

The Key Phrase Extraction skill (D) analyzes text to surface salient phrases and does not process images. The Text Translation skill (E) translates text between languages and likewise does not extract information from images.

Exam trap

The trap here is that candidates often confuse the Image Analysis skill with the OCR skill, thinking only one is needed for image extraction, but the question asks for TWO skills that extract information from images—one for visual content and one for text.

Full explanation →

231

MCQmedium

You are a developer at an e-commerce company. The company wants to build a product search feature that allows customers to search for products using natural language phrases like "red running shoes under $100". The product catalog is stored in Azure Cosmos DB and includes product descriptions, prices, and categories. The solution must use Azure AI Search and must extract entities from product descriptions to enable filtering (e.g., color, size, brand). The search must also support fuzzy matching for misspelled queries. You need to design the indexing pipeline. Which actions should you take?

A.Use Azure AI Language key phrase extraction, and enable vector search

B.Use Azure AI Document Intelligence to extract entities, and enable semantic ranking

C.Use Azure AI Language entity extraction as a custom skill, and enable fuzzy search in the index

D.Use Azure AI Vision OCR to extract text, and enable synonyms

AnswerC

Azure AI Language entity extraction runs as a custom skill during indexing, pulling colour, size and brand from product descriptions into filterable index fields. Enabling fuzzy search on those fields handles misspellings, directly satisfying the natural-language filtering and typo-tolerance requirements.

Why this answer

Azure AI Language entity extraction as a custom skill in an Azure AI Search skillset extracts entities such as color, size, and brand from product descriptions, which can then be mapped to filterable index fields. Enabling fuzzy search (via the Lucene query syntax or the built-in fuzzy matching in the query parser) supports misspelled queries. This combination directly satisfies the requirement to extract entities for filtering and to support fuzzy matching.

Exam trap

AI-102 often tests the confusion between key phrase extraction and entity extraction, and between semantic ranking and fuzzy matching, causing candidates to select key phrase extraction or semantic ranking when the requirement explicitly calls for typed entities and misspelling tolerance.

How to eliminate wrong answers

Option A is wrong because key phrase extraction returns key phrases, not typed entities like color, size, or brand, so it cannot populate filterable entity fields; vector search addresses semantic similarity, not entity extraction or fuzzy matching. Option B is wrong because Azure AI Document Intelligence is designed for OCR and form/document structure extraction, not for entity extraction from short product descriptions, and semantic ranking improves relevance ranking rather than enabling fuzzy matching. Option D is wrong because Azure AI Vision OCR extracts text from images, which is irrelevant to text-based product descriptions, and synonyms expand query terms but do not provide fuzzy matching for misspellings.

Full explanation →

232

MCQeasy

A company wants to build a mobile app that recognizes and tags landmarks in photos taken by tourists. They need a prebuilt model that requires no training and can identify thousands of famous places worldwide. Which Azure AI Vision feature should they use?

A.Image Analysis with the Landmarks feature

B.Face API with the Identify feature to match faces to famous people

C.Computer Vision with the Describe feature to generate captions of the photos

D.Custom Vision with an object detection project trained on landmark images

AnswerA

The Image Analysis service includes a Landmarks feature that can identify thousands of famous landmarks from around the world. It is a prebuilt model, so no training is required, and it returns the name and confidence score for recognized landmarks. This directly meets the requirement for a no-training solution to tag tourist photos.

Why this answer

The Landmarks feature in Image Analysis is a prebuilt model that recognizes thousands of famous landmarks and returns their names. It requires no training and is designed exactly for scenarios like tagging tourist photos. Other options either require training, are for different purposes, or do not provide specific landmark names.

Exam trap

The trap here is confusing the Landmarks feature with general image description, which may mention a landmark but does not provide a structured landmark identifier.

Full explanation →

233

MCQhard

Your organization uses Azure AI Document Intelligence to extract data from invoices. The extraction accuracy for total amounts is low. You have a labeled dataset of 500 invoices. You need to improve the model's accuracy for the 'total amount' field. What should you do?

A.Add additional predefined models for invoice processing.

B.Enable OCR enhancement to improve text recognition.

C.Increase the confidence threshold for the total amount field.

D.Create a custom neural model and train it with the labeled dataset.

AnswerD

A custom neural model learns field-specific patterns from your labelled invoices, including layout and contextual cues around the total amount, which the prebuilt model handles poorly. Training it on the 500 labelled samples directly targets the low-accuracy field, satisfying the requirement to improve extraction for 'total amount'.

Why this answer

Azure AI Document Intelligence's custom neural model is specifically designed to improve extraction accuracy for fields like 'total amount' by training on labeled datasets. Unlike the prebuilt invoice model, a custom neural model learns the unique layout and variations in your invoices, directly addressing low accuracy for a specific field. Training with 500 labeled invoices provides sufficient data to fine-tune the model's extraction capabilities.

Exam trap

Microsoft often tests the misconception that adjusting confidence thresholds or adding more predefined models can improve extraction accuracy, when in fact only custom training with labeled data addresses field-specific low accuracy.

How to eliminate wrong answers

Option A is wrong because adding additional predefined models does not improve accuracy for a specific field; predefined models are fixed and cannot be retrained or customized for your data. Option B is wrong because OCR enhancement improves text recognition quality but does not address the model's ability to correctly interpret and extract the 'total amount' field from the recognized text. Option C is wrong because increasing the confidence threshold only filters out low-confidence predictions, it does not improve the underlying model's extraction accuracy; it may reduce false positives but will not correct mis-extractions.

Full explanation →

234

MCQmedium

A law firm uses Azure Document Intelligence to extract clauses from legal contracts. They have a custom model trained on 15 labeled contracts. The model extracts clauses with high confidence on similar documents but fails to extract correct clauses from a new batch of contracts that have a different font and layout. The firm needs to improve extraction accuracy without retraining the model from scratch. The solution must minimize manual effort and cost. What should they do?

A.Use the prebuilt-layout model to extract clauses instead

B.Increase the OCR confidence threshold in the analysis request

C.Label 15 more contracts with the original layout and retrain the model

D.Create a composed model that includes the existing model and a new model trained on 5 contracts with the new layout

AnswerD

A composed model can handle multiple layouts by combining models.

Why this answer

Creating a composed model in Azure Document Intelligence allows you to combine the existing model (trained on the original layout) with a new model trained on just 5 labeled contracts from the new layout. This approach improves accuracy on the new layout without retraining from scratch, minimizing manual effort and cost by leveraging the composed model's ability to route documents to the appropriate sub-model based on layout similarity.

Exam trap

The trap here is that candidates often assume retraining with more data (Option C) is always the best solution, but they overlook the composed model feature which is specifically designed to handle layout variations with minimal additional labeling and cost.

How to eliminate wrong answers

Option A is wrong because the prebuilt-layout model is designed for extracting text and structure (like tables and selection marks), not for custom clause extraction from legal contracts, and it would not leverage the firm's existing labeled data. Option B is wrong because increasing the OCR confidence threshold only filters out low-confidence text recognition results; it does not improve the model's ability to correctly classify or extract clauses from a different font and layout. Option C is wrong because labeling 15 more contracts with the original layout and retraining the model would not address the new layout variation; it would only reinforce the existing model's performance on the original layout, wasting effort and cost.

Full explanation →

235

MCQhard

You are responsible for an Azure AI multi-agent system built on Microsoft Foundry. The system experiences frequent timeout errors when agents call external APIs. You need to implement a resilient pattern. What should you do?

A.Implement retry logic with exponential backoff in the agent tool definitions

B.Disable retry attempts to avoid duplicate requests

C.Increase the global timeout for all agents

D.Switch to synchronous agent calls

AnswerA

Retry logic with exponential backoff in the agent tool definitions absorbs transient external API failures and rate limiting, preventing them from surfacing as timeouts. Spacing retries avoids hammering the dependency, which is the resilient pattern for intermittent call failures.

Why this answer

Implementing retry logic with exponential backoff in agent tool definitions is a standard resilience pattern for transient failures when calling external APIs. This approach, often using the Retry-After header or a custom backoff strategy, reduces load on the API and prevents cascading timeouts in a multi-agent system built on Microsoft Foundry. It aligns with the recommended practices for building robust AI solutions that depend on external services.

Exam trap

The trap here is that candidates often confuse increasing timeouts with solving transient failures, but timeouts only mask the problem and do not provide resilience against intermittent API errors.

How to eliminate wrong answers

Option B is wrong because disabling retry attempts entirely would cause the system to fail on any transient error, making it less resilient and increasing the likelihood of incomplete agent tasks. Option C is wrong because increasing the global timeout for all agents does not address the root cause of transient API failures; it only delays the timeout, potentially masking underlying issues and wasting resources. Option D is wrong because switching to synchronous agent calls would block agent execution, reducing concurrency and potentially worsening timeout issues by making the system less responsive to failures.

Full explanation →

236

Multi-Selectmedium

You are developing a solution that uses Azure AI Vision to analyze images of products on an assembly line. You need to detect whether each product has a specific logo and also read the serial number printed on it. Which two Azure AI Vision services should you use? (Choose two.)

Select 2 answers

A.Azure AI Custom Vision

B.Azure AI Vision Read API

C.Azure AI Document Intelligence

D.Azure AI Vision Spatial Analysis

E.Azure AI Face API

AnswersA, B

Custom Vision allows you to train a custom image classifier or object detector to recognize specific logos. You can label images of products with and without the logo, train a model, and then use it to detect the presence of the logo on new images. This meets the requirement to detect whether each product has the specific logo.

Why this answer

To detect a specific logo and read a serial number, you need two distinct capabilities: custom object detection and text extraction. Custom Vision can be trained to recognize the logo, while the Read API extracts the serial number from the image. Together, they fulfill both requirements without unnecessary complexity.

Exam trap

The trap here is assuming that a single service like Document Intelligence can handle both tasks, but it is not designed for logo detection on products.

Full explanation →

237

Matchingmedium

Match each Azure AI feature to the service that provides it.

Drag a concept onto its matching description — or click a concept then click the description.

Concepts

Matches

Text Analytics

Face API

Custom Speech

LUIS

Computer Vision

Why these pairings

Azure Cognitive Services offer specialized AI capabilities: Language Service for text analysis (e.g., sentiment), Speech Service for audio processing, Vision Service for image analysis (e.g., OCR), and Translator for translation. Common confusions include swapping features between services.

Full explanation →

238

MCQmedium

Your organization has a large set of PDF invoices stored in Azure Blob Storage. You need to extract line-item details (product names, quantities, prices) and store them in Azure SQL Database for downstream reporting. The invoices have varied layouts. Which Azure AI service should you use?

A.Azure AI Computer Vision

B.Azure AI Language Service

C.Azure AI Search

D.Azure AI Document Intelligence

AnswerD

Azure AI Document Intelligence provides prebuilt and custom invoice models that extract line items such as product names, quantities and prices from varied layouts, then output structured JSON suitable for loading into Azure SQL Database.

Why this answer

Azure AI Document Intelligence (formerly Form Recognizer) is the correct service because it is specifically designed to extract structured data (like line-item details) from documents with varied layouts, such as invoices. Its prebuilt invoice model can parse product names, quantities, and prices from PDFs without requiring manual template configuration, and it outputs the data in a structured format that can be ingested into Azure SQL Database.

Exam trap

The trap here is that candidates often confuse Azure AI Computer Vision's OCR capability with Document Intelligence's document understanding, leading them to choose Computer Vision for any text extraction task, even when the requirement involves structured data extraction from varied-layout documents like invoices.

How to eliminate wrong answers

Option A is wrong because Azure AI Computer Vision is optimized for general image analysis (e.g., object detection, OCR for raw text) but lacks the specialized prebuilt models for extracting line-item tables from invoices with varied layouts. Option B is wrong because Azure AI Language Service focuses on text analytics (e.g., sentiment, key phrase extraction, NER) and is not designed for document structure understanding or table extraction from PDFs. Option C is wrong because Azure AI Search is a search indexing and query service, not a document extraction tool; it can index extracted data but cannot perform the initial extraction of line-item details from invoices.

Full explanation →

239

MCQmedium

A company is building an agent by using Azure AI Foundry Agent Service. The agent must use a file search tool that references a 400-page product manual stored as an uploaded file. The team wants the agent to retrieve relevant passages without you writing chunking or embedding code. Which tool should you configure?

A.A function tool that calls a Logic App workflow to parse the PDF at query time

B.The built-in file_search tool attached to the agent, with the manual uploaded to the agent's vector store

C.Azure AI Search index configured as a connected knowledge source with semantic ranker enabled

D.A code interpreter tool with the PDF mounted in the session file store

AnswerB

The file_search tool in Azure AI Foundry Agent Service automatically chunks, embeds, and indexes uploaded files into a vector store, then performs retrieval during a run. Because the platform handles ingestion and embedding, the team can point the agent at the uploaded manual and get passage-level grounding without writing custom chunking or embedding code, which is exactly what the scenario requires.

Why this answer

The requirement is managed retrieval over an uploaded document without custom chunking or embedding work. The built-in file_search tool ingests uploaded files into a vector store, handles chunking and embedding automatically, and returns relevant passages to the model during runs. Building an Azure AI Search index, writing a function tool, or using code interpreter all shift ingestion and retrieval responsibility back to the team, which the scenario rules out.

Exam trap

The trap here is assuming any grounding option works equally well, when only the built-in file_search tool removes the need for custom chunking and embedding of uploaded files.

Full explanation →

240

MCQhard

A security company uses Azure AI Face API to analyze surveillance footage. They need to detect faces in low-light images and obtain face bounding boxes, but they do not need to identify individuals. They also want to minimize cost and avoid unnecessary features. Which Face API operation should they call?

A.Face - Detect with returnFaceId=true and returnFaceLandmarks=true

B.Face - Verify with two face IDs

C.Face - Detect with returnFaceId=false and returnFaceLandmarks=false

D.Face - Identify with a person group

AnswerC

This operation returns face bounding boxes without generating face IDs or landmarks, which is exactly what is required for detection only. It minimizes cost because face ID generation is a billable feature. By setting both parameters to false, the response includes only the rectangle coordinates and basic attributes if requested, aligning with the need to detect faces in low-light images without identifying individuals.

Why this answer

To detect faces without identification, the Face - Detect operation should be called with returnFaceId and returnFaceLandmarks set to false. This returns bounding boxes and avoids the cost of generating face IDs. Identify and Verify are for matching or comparing identities, which are not needed here, and enabling face IDs or landmarks adds unnecessary expense.

Exam trap

The trap here is thinking that face detection always requires face IDs, when in fact face IDs are only needed for identification or verification and can be disabled to reduce cost.

Full explanation →

241

MCQhard

You are designing a knowledge mining solution that must extract tables from scanned invoices stored in Azure Blob Storage and make the table cells searchable. The invoices are in PDF and JPEG formats. Which Azure AI service should you use to extract the tables before loading the data into Azure AI Search?

A.Azure AI Language

B.Azure AI Search OCR skill

C.Azure AI Vision Image Analysis

D.Azure AI Document Intelligence

AnswerD

Azure AI Document Intelligence is designed to extract structured data from documents, including tables with row and column relationships, using prebuilt or custom models. It supports PDF and image inputs such as JPEG, making it suitable for scanned invoices. Its output can be transformed and loaded into Azure AI Search for searchable table content.

Why this answer

Azure AI Document Intelligence provides prebuilt and custom models that detect tables and return structured cell data with row and column indices. It accepts PDF and image inputs, which matches the scanned invoice formats. The extracted table structure can then be shaped and indexed into Azure AI Search so that table cells are searchable.

Exam trap

The trap here is assuming the Azure AI Search OCR skill can extract tables, when it only produces unstructured text and loses the row and column relationships.

Full explanation →

242

Multi-Selecthard

Which THREE factors should you consider when choosing between a pre-built model and a custom model in Azure AI Language?

Select 3 answers

A.Domain-specific vocabulary coverage

B.Time to develop and deploy

C.Need for a trained endpoint

D.Availability of labeled training data

E.Model size and memory footprint

AnswersA, B, D

Domain-specific vocabulary coverage determines whether a model recognises industry jargon, product names and abbreviations in your text. Pre-built models may miss specialised terms, whereas custom models can be trained on domain data, making coverage a decisive selection factor.

Why this answer

Option A (Domain-specific vocabulary coverage) is correct because pre-built models are trained on general-purpose text and may not recognize industry jargon, acronyms, or specialized entities, so if your scenario requires understanding niche terminology you may need a custom model trained on your own domain data. Option B (Time to develop and deploy) is correct because pre-built models can be consumed immediately via the Azure AI Language service with no training pipeline, whereas custom models require data preparation, training, evaluation, and deployment, which adds significant time. Option D (Availability of labeled training data) is correct because custom models in Azure AI Language (for example custom named entity recognition, custom text classification, or custom question answering) depend on sufficient, high-quality labeled examples; without that data a pre-built model is the practical choice.

Option C (Need for a trained endpoint) is not a deciding factor because both pre-built and custom models are consumed through an endpoint in Azure AI Language, so this does not differentiate the two approaches. Option E (Model size and memory footprint) is not a relevant consideration because Azure AI Language is a managed service that abstracts away model hosting, sizing, and memory concerns from the developer.

Exam trap

The trap here is that candidates confuse 'need for a trained endpoint' (which is always required for custom models but also exists for pre-built models via a shared endpoint) with the decision factor of whether you have labeled training data to build a custom model.

Full explanation →

243

MCQeasy

You need to analyze videos stored in Azure Blob Storage to detect objects and generate timestamps. Which Azure service should you use?

A.Azure Custom Vision

B.Azure Form Recognizer

C.Azure Computer Vision

D.Azure Video Indexer

AnswerD

Azure Video Indexer extracts insights from video, including object detection with timestamps, and can ingest content directly from Azure Blob Storage. This satisfies the stem's requirements for both object detection and timestamp generation on stored videos.

Why this answer

Azure Video Indexer (D) is the correct choice because it is specifically designed to analyze videos, extracting insights such as object detection, scene segmentation, and timestamps. It uses AI models to process video content stored in Azure Blob Storage and generates a timeline of detected objects, making it ideal for this scenario.

Exam trap

The trap here is that candidates often confuse Azure Computer Vision (image analysis) with video analysis, overlooking that Computer Vision lacks native video processing and timestamp generation, while Video Indexer is the dedicated service for end-to-end video insights.

How to eliminate wrong answers

Option A is wrong because Azure Custom Vision is a service for training custom image classification and object detection models on images, not for analyzing pre-recorded videos with timestamp generation. Option B is wrong because Azure Form Recognizer is designed to extract text and structure from documents (e.g., invoices, forms), not for video analysis or object detection. Option C is wrong because Azure Computer Vision provides image analysis APIs (e.g., object detection in static images) but lacks native video processing capabilities and timestamp generation; it would require additional custom logic to handle video frames sequentially.

Full explanation →

244

Multi-Selecteasy

You need to use Azure AI Language to analyze customer feedback. Which THREE analysis types are available in the Text Analytics API?

Select 3 answers

A.Image captioning

B.Speech-to-text

C.Sentiment analysis

D.Entity recognition

E.Key phrase extraction

AnswersC, D, E

Sentiment analysis is a core Text Analytics capability, returning per-document and per-sentence scores between 0 and 1 for positive, neutral and negative tones. It satisfies the stem's requirement for available analysis types, alongside key phrase extraction and named entity recognition, which together form the three standard features.

Why this answer

Sentiment analysis (C) is a core Text Analytics API feature that returns sentiment labels and confidence scores (positive, neutral, negative, mixed) for documents or sentences, making it valid for analyzing customer feedback. Entity recognition (D), specifically Named Entity Recognition (NER), is also available and identifies entities such as people, places, organizations, and dates in text. Key phrase extraction (E) is another supported Text Analytics capability that returns the main concepts or topics in a document, which is useful for summarizing customer feedback.

Image captioning (A) belongs to Azure AI Vision, not Azure AI Language, and speech-to-text (B) is provided by Azure AI Speech, so neither is part of the Text Analytics API.

Exam trap

Azure often tests your ability to distinguish between Azure AI services, so the trap here is that candidates confuse the Text Analytics API with broader AI capabilities like image or speech processing, leading them to select options that belong to other services.

Full explanation →

245

MCQhard

You run the Azure CLI command shown in the exhibit to create an online endpoint for a generative AI model. The deployment fails because the selected VM instance type is not available in the East US region. Which action should you take to resolve the issue?

A.Increase the instance count to 2

B.Specify a different VM type or region that supports Standard_NC6s_v3

C.Use a batch endpoint instead of online endpoint

D.Change --compute-type to CPU

AnswerB

The failure is a regional capacity constraint, not a configuration error. Redeploying with a VM size or region where Standard_NC6s_v3 is offered resolves the unavailability, satisfying the requirement that the endpoint's compute SKU actually exists in the target location.

Why this answer

The deployment failed because the Standard_NC6s_v3 VM instance type is not available in the East US region. The correct action is to either choose a different VM type that is available in East US or deploy to a different region that supports Standard_NC6s_v3. This directly addresses the root cause of the failure, as Azure Machine Learning online endpoints require the selected VM SKU to be available in the target region.

Exam trap

The trap here is that candidates may think increasing instance count or switching to a batch endpoint will bypass the regional SKU limitation, but neither changes the underlying VM type or region, so the deployment will still fail.

How to eliminate wrong answers

Option A is wrong because increasing the instance count does not change the VM type or region; it only scales out the number of instances, which does not resolve the unavailability of the VM SKU. Option C is wrong because switching to a batch endpoint does not address the VM availability issue; batch endpoints also require compatible compute resources and are designed for asynchronous, large-scale inference, not for fixing regional SKU unavailability. Option D is wrong because changing --compute-type to CPU would not help if the model requires GPU acceleration (as implied by the NC-series VM), and it does not solve the regional availability problem for the specified VM type.

Full explanation →

246

MCQeasy

You are using Azure AI Language to perform entity recognition on customer feedback. You need to identify the sentiment expressed towards specific entities. Which feature should you use?

A.Named Entity Recognition (NER)

B.Sentiment analysis with opinion mining

C.Entity linking

D.Key phrase extraction

AnswerB

Opinion mining extends sentiment analysis by associating each detected entity with its own sentiment and the specific words expressing it, rather than returning one document-level score. This directly satisfies the requirement to identify sentiment towards specific entities in the feedback.

Why this answer

Sentiment analysis with opinion mining is the correct feature because it not only detects the overall sentiment of a text but also associates specific sentiments with particular entities or aspects mentioned in the text. This allows you to determine, for example, that a customer feels positively about 'product quality' but negatively about 'customer support', which is exactly what the question requires.

Exam trap

The trap here is that candidates often confuse Named Entity Recognition (NER) with the ability to extract sentiment about entities, but NER only identifies entities without any sentiment analysis, while opinion mining is the specific feature that combines entity detection with sentiment scoring.

How to eliminate wrong answers

Option A is wrong because Named Entity Recognition (NER) only identifies and categorizes entities (e.g., person, organization, location) but does not analyze sentiment or opinion towards those entities. Option C is wrong because Entity Linking disambiguates entities by linking them to a knowledge base (like Wikipedia) and does not perform sentiment analysis. Option D is wrong because Key Phrase Extraction returns a list of important phrases from the text but does not evaluate sentiment or associate opinions with specific entities.

Full explanation →

247

MCQhard

You are building a generative AI application using Azure OpenAI Service. The application must provide factual answers based on your company's internal knowledge base. You need to minimize the risk of the model generating incorrect information (hallucinations). Which approach should you take?

A.Implement Retrieval-Augmented Generation (RAG) with Azure AI Search.

B.Fine-tune the model on your company's documents.

C.Use few-shot prompting with examples of correct answers.

D.Set the max_tokens parameter to a low value.

AnswerA

RAG grounds responses in retrieved documents from Azure AI Search, so the model cites your internal knowledge base rather than relying on parametric memory. This directly minimises hallucination risk, satisfying the requirement for factual answers drawn from company content.

Why this answer

Retrieval-Augmented Generation (RAG) with Azure AI Search grounds the model's responses in your company's internal knowledge base by retrieving relevant documents in real time and injecting them into the prompt. This reduces hallucinations by ensuring the model generates answers based on retrieved facts rather than relying solely on its parametric memory. Azure AI Search provides vector and hybrid search capabilities that efficiently index and query your documents, making RAG the most effective approach for factual accuracy.

Exam trap

The AI-102 exam often tests the misconception that fine-tuning (Option B) is the best way to ground a model in proprietary data, but the trap is that fine-tuning does not provide dynamic, query-specific retrieval and can still produce hallucinations, whereas RAG explicitly forces the model to use retrieved facts.

How to eliminate wrong answers

Option B is wrong because fine-tuning adjusts the model's weights on your documents, which can lead to overfitting and does not guarantee that the model will not hallucinate; it still relies on its internal knowledge and may generate plausible-sounding but incorrect information when queried outside the fine-tuned distribution. Option C is wrong because few-shot prompting provides examples but does not ground the model in your specific knowledge base; the model can still hallucinate if the examples do not cover the exact query context or if it extrapolates incorrectly. Option D is wrong because setting max_tokens to a low value only truncates the output length and does not improve factual accuracy; it may even cause incomplete or misleading answers without addressing the root cause of hallucinations.

Full explanation →

248

MCQeasy

A developer is creating an agent in Azure AI Foundry Agent Service and wants the agent to remember a user's stated preferences across separate conversations that occur days apart. The agent definition already includes a model deployment and instructions. What should the developer add to persist this context?

A.Increase the model's context window by selecting a larger deployment

B.Attach a Foundry memory store to the agent so it can save and retrieve user facts across threads

C.Enable the code interpreter tool so the agent can write preferences to a file

D.Add the preferences to the agent's instructions field in the agent definition

AnswerB

A memory store in Azure AI Foundry Agent Service is designed to persist user-level facts and preferences beyond a single thread. The agent can write memories during one conversation and retrieve them in later conversations, even days apart, as long as the same user identity is used. This directly provides the cross-session continuity the scenario requires.

Why this answer

Cross-conversation memory requires a durable, user-scoped store outside the model and outside static agent configuration. Attaching a Foundry memory store lets the agent save facts during one thread and retrieve them in later threads for the same user. A larger context window, code interpreter session files, and the instructions field all fail because they are either ephemeral, shared, or scoped to a single run.

Exam trap

The trap here is confusing a bigger context window with persistent memory, when context length is per-request and memory must be stored externally.

Full explanation →

249

MCQmedium

You are building a custom question answering solution using Azure AI Language. The knowledge base contains a large number of question-answer pairs. You need to ensure that the solution returns the most relevant answer for a user's query, even when the query contains synonyms not present in the knowledge base. What should you do?

A.Enable the 'enableHierarchicalExtraction' setting in the project configuration.

B.Set the 'defaultAnswer' to a response that asks the user to rephrase using known terms.

C.Add alternative questions for each answer that include common synonyms.

D.Increase the number of documents in the knowledge base to provide more context.

AnswerC

Custom question answering uses the alternative questions to match user queries. Adding synonyms as alternative questions improves recall and ensures that queries with different wording still find the correct answer. This is a supported and straightforward way to enhance relevance without changing the underlying model.

Why this answer

Custom question answering matches user queries to questions in the knowledge base. To handle synonyms, you can add alternative questions that include common synonyms for each answer. This improves the likelihood that a query with different wording will match the correct answer.

Other options do not directly address synonym matching and may not improve relevance.

Exam trap

The trap here is assuming that enabling a configuration setting or adding more documents will automatically handle synonyms, when the correct approach is to explicitly add alternative questions with synonyms.

Full explanation →

250

MCQmedium

A company is deploying a solution using Azure AI Vision to analyze images of products on a retail website. They need to ensure that the image analysis is performed within a specific geographic boundary for data residency compliance. What should they configure?

A.Deploy the Azure AI Vision resource in the desired Azure region

B.Create a private endpoint for the Vision resource

C.Enable multi-region replication on the Vision resource

D.Use the Free tier of Azure AI Vision

AnswerA

Deploying the Azure AI Vision resource in the target Azure region keeps image processing and stored data within that geography, satisfying the data residency constraint. Regional deployment determines where inference occurs; unlike global endpoints, no cross-geography replication happens. Microsoft Entra ID governs access but does not affect processing location.

Why this answer

Azure AI Vision resources are regional Azure resources, meaning all data processing and storage occur within the Azure region where the resource is deployed. By deploying the resource in the desired geographic region, you ensure that image analysis and any derived data remain within that boundary, satisfying data residency compliance requirements. This is the fundamental mechanism for controlling data location in Azure AI services.

Exam trap

The trap here is that candidates confuse network isolation (private endpoints) with data residency, or assume that replication or tier changes can alter where data is stored, when in fact the region of the resource itself is the sole determinant for compliance.

How to eliminate wrong answers

Option B is wrong because a private endpoint restricts network access to the resource via a private IP address in your virtual network, but it does not control the geographic location where data is processed or stored. Option C is wrong because Azure AI Vision does not support multi-region replication; that feature is available for Azure Storage and Cosmos DB, not for AI services. Option D is wrong because the Free tier imposes usage limits (e.g., 20 transactions per minute) and does not provide any data residency guarantees; it still processes data in the region where the resource is created.

Full explanation →

251

MCQmedium

You are building an Azure AI Search enrichment pipeline that extracts key phrases from documents stored in Azure Blob Storage. The documents are plain text files in English. You need to add a built-in skill that identifies the main concepts in each document without writing custom code. Which skill should you use?

A.Entity Recognition skill

B.Key Phrase Extraction skill

C.Language Detection skill

D.Text Translation skill

AnswerB

The Key Phrase Extraction skill is a built-in cognitive skill in Azure AI Search that uses natural language processing to identify the main points in text. It is designed for exactly this scenario: extracting key phrases from unstructured English text without custom code. It outputs a list of key phrases that can be mapped to an index field, enabling users to search or filter by important concepts.

Why this answer

The Key Phrase Extraction skill is the correct choice because it is a built-in Azure AI Search skill that analyzes text and returns a list of key phrases representing the main concepts. It requires no custom code and integrates directly into a skillset. The other skills serve different purposes: entity recognition focuses on named entities, language detection identifies language, and translation converts languages.

Exam trap

The trap here is confusing general key phrase extraction with named entity recognition, assuming that extracting entities will also capture the main concepts of a document.

Full explanation →

252

MCQhard

Your Azure AI Search index is experiencing high query latency. You have enabled semantic search and custom scoring profiles. You need to reduce latency without degrading search quality. Which action should you take?

A.Remove custom scoring profiles.

B.Increase the number of replicas.

C.Reduce the number of partitions.

D.Disable semantic search.

AnswerB

Query latency stems from limited compute, not index size. Replicas add query-processing capacity and enable load balancing across nodes, reducing latency while preserving semantic ranking and scoring profiles. Partition changes would affect storage and indexing, not query throughput.

Why this answer

Increasing the number of replicas distributes query load across multiple copies of the index, allowing parallel processing of search requests. This directly reduces query latency without altering the search logic, scoring profiles, or semantic enrichment, thus preserving search quality.

Exam trap

The trap here is that candidates confuse partitions (which affect storage and indexing speed) with replicas (which affect query throughput), leading them to incorrectly reduce partitions or disable features instead of scaling query capacity.

How to eliminate wrong answers

Option A is wrong because removing custom scoring profiles would degrade search quality by eliminating relevance tuning, and it does not address the root cause of high latency (insufficient query capacity). Option C is wrong because reducing partitions decreases the index's data capacity and can increase latency by forcing more data per partition, while partitions primarily affect indexing speed and storage, not query throughput. Option D is wrong because disabling semantic search would degrade search quality by removing AI-powered ranking and relevance features, and it does not address the underlying query load issue that replicas solve.

Full explanation →

253

MCQeasy

You are developing a solution that uses Azure AI Language to analyze customer feedback. You need to determine whether the sentiment of a given sentence is positive, negative, or neutral. Which Azure AI Language feature should you use?

A.Sentiment Analysis

B.Entity Recognition

C.Language Detection

D.Key Phrase Extraction

AnswerA

Sentiment Analysis in Azure AI Language returns per-sentence labels of positive, negative or neutral together with confidence scores. It directly satisfies the requirement to classify each sentence's sentiment, unlike key phrase extraction or entity recognition, which return different output types.

Why this answer

Sentiment Analysis is the correct Azure AI Language feature because it is specifically designed to evaluate text and determine whether the sentiment expressed is positive, negative, or neutral. This feature uses machine learning classifiers trained on large datasets to assign a sentiment label and confidence scores at the sentence and document level, directly matching the requirement to analyze customer feedback for sentiment polarity.

Exam trap

The trap here is that candidates often confuse Key Phrase Extraction with Sentiment Analysis because both seem to 'analyze' text, but Key Phrase Extraction only identifies topics or terms, not the emotional polarity of the content.

How to eliminate wrong answers

Option B is wrong because Entity Recognition identifies and categorizes named entities (e.g., people, organizations, locations) in text, but it does not evaluate sentiment or polarity. Option C is wrong because Language Detection identifies the language in which the text is written (e.g., English, Spanish), not the sentiment expressed. Option D is wrong because Key Phrase Extraction returns a list of key phrases or main talking points from the text, but it does not classify the overall sentiment as positive, negative, or neutral.

Full explanation →

254

Multi-Selecthard

You are deploying an Azure AI solution that uses Azure AI Language and Azure AI Vision. You need to ensure that the solution can be deployed repeatedly to multiple environments with consistent configuration and that secrets are managed securely. (Choose two.)

Select 2 answers

A.Use Azure Key Vault to store the Azure AI service keys and reference the secrets from the ARM template.

B.Store the Azure AI service keys in the ARM template parameters file.

C.Use the Azure portal to manually create the Azure AI resources in each environment.

D.Store the Azure AI service keys in a Git repository and use a pipeline to inject them during deployment.

E.Use an ARM template to define the Azure AI resources and their configurations.

AnswersA, E

Azure Key Vault provides secure storage for secrets. By referencing Key Vault secrets from an ARM template, you avoid hardcoding keys in the template or parameters. This allows the same template to be deployed to multiple environments while retrieving environment-specific secrets from Key Vault, meeting both consistency and security requirements.

Why this answer

To achieve repeatable deployments with consistent configuration, use infrastructure as code such as ARM templates. To manage secrets securely, store keys in Azure Key Vault and reference them from the templates. Combining these two practices allows the same deployment process to be used across environments while keeping secrets out of source control and deployment definitions.

Exam trap

The trap here is assuming that storing secrets in a parameters file or Git repository is acceptable if access is restricted, when in fact secrets should always be stored in a dedicated secure store like Azure Key Vault.

Full explanation →

255

MCQmedium

You are building an ASP.NET Core web app that must analyze the sentiment of user-submitted product reviews in real time. The app is deployed to an Azure App Service web app. You want to avoid managing API keys in code and prefer to use the app's managed identity for authentication to the Azure AI Language resource. What should you do to enable the app to call the sentiment analysis API securely?

A.Configure the app to use the Language resource's endpoint and a SAS token generated from the resource's keys.

B.Store the Language resource's subscription key in Azure Key Vault and retrieve it at runtime using the app's managed identity.

C.Enable a system-assigned managed identity on the App Service and add the identity's object ID to the Language resource's access control list (ACL).

D.Assign the Cognitive Services User role to the App Service's managed identity on the Language resource, and use DefaultAzureCredential in the app code.

AnswerD

This is correct because the app uses its managed identity to obtain a token from Microsoft Entra ID, and the role assignment grants permission to call the Language service. DefaultAzureCredential automatically picks up the managed identity in Azure, eliminating the need to store keys. This is the recommended passwordless approach for Azure-hosted apps.

Why this answer

The recommended secure approach is to use managed identity with Microsoft Entra ID authentication. Assigning the Cognitive Services User role to the App Service's managed identity grants the necessary permissions, and DefaultAzureCredential in the SDK automatically obtains a token. This eliminates the need to store or manage API keys in code, aligning with the scenario's requirement.

Exam trap

The trap here is assuming that storing keys in Key Vault fully solves the secret management problem, but the app still uses a key for authentication rather than a passwordless identity.

Full explanation →

256

MCQmedium

A company is building a knowledge mining solution using Azure AI Search. They need to extract key phrases from a large set of documents in multiple languages. Which skill should they add to the skillset?

A.Key Phrase Extraction skill

B.Sentiment Analysis skill

C.Language Detection skill

D.Entity Recognition skill

AnswerA

The Key Phrase Extraction skill uses natural language processing to identify salient terms and phrases, and it supports multiple languages within Azure AI Search skillsets. It satisfies the requirement to extract key phrases from a large multilingual document set.

Why this answer

The Key Phrase Extraction skill is the correct choice because it is specifically designed to identify and extract the most important phrases from text, which directly supports the requirement to extract key phrases from documents. Azure AI Search's built-in Key Phrase Extraction skill leverages natural language processing to analyze text and return a list of key phrases, making it the appropriate skill for this knowledge mining solution.

Exam trap

The trap here is that candidates may confuse Entity Recognition (which extracts single-word entities like 'Microsoft') with Key Phrase Extraction (which extracts multi-word phrases like 'Azure AI Search'), leading them to choose Option D instead of A.

How to eliminate wrong answers

Option B (Sentiment Analysis skill) is wrong because it evaluates the emotional tone or sentiment (positive, negative, neutral) of text, not the extraction of key phrases. Option C (Language Detection skill) is wrong because it identifies the language of the text but does not extract key phrases from the content. Option D (Entity Recognition skill) is wrong because it identifies and categorizes named entities (e.g., people, organizations, locations) rather than extracting multi-word key phrases that summarize the document's main topics.

Full explanation →

257

MCQhard

Your company uses Azure Cognitive Search to index millions of documents. Users report that search results include irrelevant documents. You need to improve search relevance by boosting documents that contain the search term in the title field. Which scoring profile configuration should you use?

A.Create a tagging scoring profile that boosts by the title field with a weight of 10.

B.Create a freshness scoring profile with a boosting duration of 30 days.

C.Create a distance scoring profile with a reference point parameter.

D.Create a magnitude scoring profile with a boosting function of 'linear'.

AnswerA

A tagging profile boosts documents that have matching terms in a specific field, like title.

Why this answer

A tagging scoring profile boosts documents based on matching tags from a specific field (like title), and the weight parameter controls the boost magnitude. Option B is incorrect because a freshness scoring profile boosts by recency, not by term presence in a field. Option C is incorrect because a distance scoring profile boosts based on geospatial proximity, not text fields.

Option D is incorrect because a magnitude scoring profile boosts based on numeric field values, not text fields; a linear boosting function is used for magnitude profiles, but this does not apply to boosting by title.

Full explanation →

258

MCQmedium

You manage an Azure AI Search solution that indexes documents from Azure Blob Storage. The index must support real-time updates when documents are added or modified. Which approach should you use?

A.Use the push API to manually upload documents.

B.Configure an indexer with a short schedule and change tracking.

C.Manually reset and rerun the indexer after each change.

D.Add a cognitive skillset to process new documents.

AnswerB

A blob indexer with change tracking detects new and modified documents via the storage layer's timestamps, so each scheduled run reindexes only changed content. A short schedule keeps latency low, satisfying the real-time update constraint without rebuilding the index or reindexing unchanged blobs.

Why this answer

Azure AI Search indexers can be configured with a short schedule (e.g., every 5 minutes) and change tracking (using high-water mark or integrated change detection on Azure Blob Storage) to automatically index new or modified documents without manual intervention. This provides near-real-time updates while leveraging the indexer's built-in change detection capabilities.

Exam trap

The trap here is that candidates often confuse the push API (Option A) as the only way to achieve real-time updates, overlooking that indexers with change tracking and a short schedule can provide automated near-real-time indexing without custom code.

How to eliminate wrong answers

Option A is wrong because the push API requires manual coding to upload documents, which does not automatically detect changes in Azure Blob Storage and is not a 'configured' approach for real-time updates from a data source. Option C is wrong because manually resetting and rerunning the indexer after each change is not automated and does not support real-time updates; it is a manual, batch-oriented process. Option D is wrong because a cognitive skillset enriches documents with AI transformations (e.g., OCR, entity recognition) but does not handle the scheduling or change tracking needed for real-time indexing of new or modified documents.

Full explanation →

259

MCQmedium

A development team is using Azure Cognitive Service for Language to extract key phrases from customer reviews. They notice that some reviews are not being processed, and the API returns a 400 error code. What is the most likely cause?

A.One of the reviews exceeds the maximum character limit for a single document.

B.The reviews contain characters that are not valid UTF-8.

C.The request contains more than 5 documents.

D.The reviews are written in a language not supported by the service.

AnswerA

The service limits each document to 5,120 characters.

Why this answer

The Azure Cognitive Service for Language key phrase extraction API enforces a maximum document size of 5,120 characters per document. When a single review exceeds this limit, the API returns a 400 Bad Request error because the request payload violates the service's input constraints. This is the most common cause of 400 errors in batch text analysis operations.

Exam trap

The trap here is that candidates often assume the 400 error is due to unsupported languages or encoding issues, but the actual constraint is the per-document character limit, which is explicitly documented in the service's input specifications.

How to eliminate wrong answers

Option B is wrong because the service automatically handles UTF-8 encoding validation and would return a different error (e.g., 400 with 'InvalidRequestContent') if characters were not valid UTF-8, but the question states the reviews are standard customer reviews, making invalid UTF-8 unlikely. Option C is wrong because the API supports up to 10 documents per request (not 5), so a request with more than 5 documents would still succeed unless it exceeds the 10-document limit. Option D is wrong because unsupported languages typically result in a successful response with empty key phrases or a warning, not a 400 error; the service supports over 40 languages for key phrase extraction.

Full explanation →

260

MCQhard

You are developing an Azure AI solution that uses Azure Cognitive Search. The solution must index documents from an Azure Blob Storage container. You need to ensure that the indexer can access the Blob Storage container securely without storing credentials in the indexer definition. What should you do?

A.Use a shared access signature (SAS) token for the Blob Storage container and include it in the indexer's data source connection string.

B.Store the Blob Storage account key in Azure Key Vault and reference it in the indexer definition using a Key Vault secret URI.

C.Enable a managed identity for the Azure Cognitive Search service and grant it access to the Blob Storage container.

D.Configure the indexer to use the Azure Cognitive Search service's API key to authenticate to Blob Storage.

AnswerC

Using a managed identity for Azure Cognitive Search allows the indexer to authenticate to Blob Storage without storing credentials. You grant the managed identity the necessary role, such as Storage Blob Data Reader, on the storage account. This is a secure and recommended approach. The indexer can then use the managed identity to access the container. This eliminates the need for connection strings with keys.

Why this answer

A managed identity for Azure Cognitive Search is the correct approach because it allows the indexer to authenticate to Blob Storage without storing any credentials in the indexer definition. You assign the managed identity the appropriate role on the storage account. This is a secure, Azure-native method.

The other options either store credentials (SAS token or account key) or use the wrong key for authentication.

Exam trap

The trap here is assuming that Azure Key Vault can be directly integrated with Cognitive Search indexers to retrieve secrets at runtime.

Full explanation →

261

MCQhard

You are developing a chatbot for a retail company using Azure AI Language's custom question answering. The chatbot must provide answers from a knowledge base of 500 FAQ documents. Users often ask the same question in different wording, and the chatbot fails to return an answer for paraphrased queries. What is the most effective solution?

A.Use Azure AI Bot Service's Direct Line Speech channel to improve accuracy.

B.Enable active learning in the project settings and periodically publish the updated knowledge base.

C.Increase the number of FAQ documents in the knowledge base.

D.Manually add alternate question phrases to the knowledge base for each QnA pair.

AnswerB

Active learning logs paraphrased queries that score low confidence, surfacing them for review so you can add alternative question phrasings to each FAQ pair. Republishing applies those additions, letting the model match the varied wording users actually type instead of only the original phrasing.

Why this answer

Enabling active learning in Azure AI Language's custom question answering allows the system to learn from user interactions. It suggests alternative phrasings for existing QnA pairs, which helps answer paraphrased queries without manual effort. Option A is wrong because Direct Line Speech channel is for voice interactions, not improving answer coverage.

Option C is wrong because simply adding more documents does not address the paraphrasing issue; it may increase redundancy but not handle different wording of the same question. Option D is wrong because manually adding alternate phrases is time-consuming and not scalable; active learning automates this process by identifying and suggesting variations based on user queries.

Full explanation →

262

MCQeasy

A logistics company receives shipping manifests as PDF documents. They need to extract the shipper name, consignee, and total weight from each document. The documents have a consistent layout. Which Azure AI service should they use?

A.Azure AI Document Intelligence with a custom extraction model.

B.Azure AI Translator to convert the PDF to text and then parse with regex.

C.Azure AI Language with custom named entity recognition (NER).

D.Azure AI Vision with OCR and then custom text classification.

AnswerA

Document Intelligence is designed to extract structured data from documents. A custom extraction model can be trained on labeled samples of the manifests to recognize shipper name, consignee, and total weight. Because the layout is consistent, a custom model will achieve high accuracy, making this the appropriate service.

Why this answer

Document Intelligence is built to extract fields from documents, especially forms with consistent layouts. A custom extraction model can be trained to identify shipper name, consignee, and total weight accurately. The other services either do not handle document layout or are not designed for field extraction from PDFs.

Exam trap

The trap here is choosing a text-based service like custom NER or OCR plus classification, when the scenario requires extracting structured fields from documents with layout.

Full explanation →

263

MCQeasy

You are building an Azure AI Search solution to index a collection of technical manuals. Users need to find documents by searching for specific terms and also have the ability to filter by document category. Which feature should you configure in the index to support filtering?

A.Set the 'filterable' property to true on the category field

B.Set the 'facetable' property to true on the category field

C.Set the 'searchable' property to true on the category field

D.Set the 'sortable' property to true on the category field

AnswerA

Filtering in Azure AI Search requires the field to be marked filterable in the index definition. Setting filterable to true on the category field enables OData filter expressions alongside term searches, satisfying the stated requirement.

Why this answer

Setting the 'filterable' property to true on the category field enables Azure AI Search to apply OData filter expressions (e.g., `$filter=category eq 'Networking'`) during query execution. This allows users to narrow search results by document category without requiring the field to be full-text searchable, which is essential for efficient filtering in a technical manuals index.

Exam trap

The trap here is that candidates often confuse 'facetable' with 'filterable' because both are used in search UIs for narrowing results, but faceting only provides aggregation counts for navigation, not the ability to apply server-side OData filters.

How to eliminate wrong answers

Option B is wrong because setting 'facetable' to true enables drill-down navigation (e.g., showing category counts in a UI), but it does not support direct filtering via $filter; faceting and filtering are separate capabilities. Option C is wrong because setting 'searchable' to true enables full-text search on the category field, but filtering does not require searchability—in fact, marking a field as searchable consumes additional storage and processing overhead unnecessarily. Option D is wrong because setting 'sortable' to true allows ordering results by the category field (e.g., $orderby=category), but it does not enable the $filter parameter to restrict results based on category values.

Full explanation →

264

MCQeasy

Your organization needs to monitor Azure AI services for unusual activity patterns that might indicate a security threat. Which Microsoft security solution should you use?

A.Microsoft Intune

B.Microsoft Defender XDR

C.Microsoft Purview

D.Microsoft Sentinel

AnswerD

Microsoft Sentinel is a cloud-native SIEM and SOAR platform that ingests logs from Azure AI services and applies analytics rules and machine learning to detect anomalous activity patterns, satisfying the requirement to monitor for unusual behaviour indicating a security threat.

Why this answer

Microsoft Sentinel is a cloud-native Security Information and Event Management (SIEM) and Security Orchestration Automated Response (SOAR) solution. It is specifically designed to ingest logs from Azure AI services, apply analytics to detect unusual activity patterns, and generate alerts for potential security threats, making it the correct choice for monitoring AI services for security anomalies.

Exam trap

The trap here is that candidates often confuse Microsoft Sentinel with Microsoft Defender XDR, assuming that 'security monitoring' always falls under Defender, but Sentinel is the SIEM solution required for ingesting and analyzing logs from Azure AI services, while Defender XDR focuses on endpoint and identity protection.

How to eliminate wrong answers

Option A is wrong because Microsoft Intune is a mobile device management (MDM) and mobile application management (MAM) solution, focused on managing endpoints and enforcing compliance policies, not on monitoring cloud service activity for security threats. Option B is wrong because Microsoft Defender XDR (Extended Detection and Response) is designed to correlate signals across endpoints, email, and identities, but it does not natively ingest and analyze logs from Azure AI services for SIEM-style threat detection. Option C is wrong because Microsoft Purview is a data governance and compliance solution, primarily used for data cataloging, classification, and policy enforcement, not for real-time security monitoring or threat detection.

Full explanation →

265

Multi-Selectmedium

You are deploying a generative AI chat application that calls an Azure OpenAI model deployment. The application must stream partial completions to the browser as tokens are produced and must reduce perceived latency for long answers. Which two request options should you configure? (Choose two.)

Select 2 answers

A.Increase the frequency_penalty value to discourage repeated tokens

B.Set stream to true on the chat completions request

C.Set the logprobs parameter to return token-level probabilities

D.Set n to a value greater than one to generate multiple candidate completions

E.Consume the response as server-sent events and forward deltas to the client

AnswersB, E

Setting stream to true makes the service return server-sent events containing incremental deltas as the model generates tokens, so the client can render partial text immediately. This directly reduces perceived latency for long answers because the user sees content before the completion finishes. Without streaming, the client waits for the entire response body, which is the behavior the scenario wants to avoid.

Why this answer

Streaming has two cooperating parts: the request must set stream to true so the service emits incremental deltas, and the client must read those server-sent events and forward them to the browser. Together they let users see text as it is generated. Parameters that alter sampling, penalties, or probabilities change content or add metadata but never change delivery timing.

Exam trap

The trap here is treating the stream parameter alone as sufficient without consuming and forwarding the event stream on the client.

Full explanation →

266

Multi-Selectmedium

You are managing an Azure AI solution that uses Azure OpenAI and Azure AI Search. You need to ensure that the solution can be deployed to multiple environments with different configurations while minimizing manual steps. Which two actions should you take? (Choose two.)

Select 2 answers

A.Store environment-specific values in a parameter file for each environment.

B.Manually create resources in the Azure portal for each environment.

C.Use a single hard-coded connection string for all environments.

D.Use Azure Policy to enforce naming conventions and tags after deployment.

E.Define the deployment as a Bicep module and reference it from environment-specific main files.

AnswersA, E

Parameter files allow you to separate environment-specific values such as SKU, location, and capacity from the template logic. By maintaining a parameter file per environment, you can deploy the same template repeatedly with the correct settings for development, test, and production without editing the template itself.

Why this answer

To deploy the same solution to multiple environments with minimal manual steps, you should parameterize environment-specific values using parameter files and modularize the deployment using Bicep modules referenced by environment-specific main files. This combination ensures consistency, reduces duplication, and allows automated, repeatable deployments across development, test, and production.

Exam trap

The trap here is focusing on post-deployment governance tools like Azure Policy or manual portal steps instead of the core infrastructure-as-code practices that actually automate and parameterize multi-environment deployments.

Full explanation →

267

MCQeasy

A company wants to use Azure AI Translator to translate customer emails from English to French. They need to ensure that the translation preserves the tone and formality of the original text. What should they configure in the request?

A.Set the 'category' parameter to 'general' to use a standard translation model.

B.Set the 'scope' parameter to 'document' to ensure context-aware translation.

C.Set the 'formality' parameter to the desired level (e.g., 'formal' or 'informal').

D.Set the 'language' parameter to 'fr' and the 'from' parameter to 'en'.

AnswerC

The formality parameter instructs the neural translation model to bias output toward formal or informal register, preserving the source email's tone in French. Setting it to the desired level satisfies the requirement to maintain formality, which standard translation without this parameter would not reliably control.

Why this answer

Azure AI Translator provides a 'formality' parameter that allows you to specify the desired level of formality (e.g., 'formal' or 'informal') in the translated text. This parameter directly controls the tone and register of the output, ensuring that the translation preserves the original email's tone and formality, which is critical for customer communications.

Exam trap

The trap here is that candidates often confuse the 'formality' parameter with language or category settings, mistakenly thinking that simply specifying the target language (Option D) or using a general category (Option A) is sufficient to control tone, when in fact the formality parameter is the only dedicated mechanism for this purpose.

How to eliminate wrong answers

Option A is wrong because the 'category' parameter is used to select a custom translation model or domain (e.g., 'general' for standard translations), but it does not control tone or formality; it affects terminology and style based on the training domain. Option B is wrong because Azure AI Translator does not have a 'scope' parameter; context-aware translation for documents is handled by the Document Translation feature, not by a request parameter in the standard Translate operation. Option D is wrong because while setting 'language' to 'fr' and 'from' to 'en' is necessary for specifying the source and target languages, it does not address the requirement to preserve tone and formality; it only defines the language pair.

Full explanation →

268

MCQeasy

You need to summarize a large document using Azure AI Language. Which feature should you use?

A.Document summarization

B.Key phrase extraction

C.Entity recognition

D.Sentiment analysis

AnswerA

Document summarization is the Azure AI Language feature purpose-built for condensing long text into extractive or abstractive summaries, handling documents exceeding token limits. Sentiment analysis, entity recognition and key phrase extraction return labels or phrases rather than coherent summaries, so they cannot satisfy the requirement.

Why this answer

Document summarization is the correct feature because it is specifically designed to generate concise summaries of large documents, extracting the most important information. Azure AI Language's document summarization uses extractive or abstractive techniques to produce a summary, directly addressing the requirement to summarize a large document.

Exam trap

The trap here is that candidates may confuse key phrase extraction with summarization, thinking that extracting important phrases is equivalent to summarizing the document, but key phrase extraction lacks the narrative structure and coherence of a true summary.

How to eliminate wrong answers

Option B is wrong because key phrase extraction identifies and returns a list of key terms or phrases from the text, but it does not produce a coherent summary or reduce the document's length. Option C is wrong because entity recognition identifies and categorizes named entities (e.g., people, organizations, locations) but does not summarize the content. Option D is wrong because sentiment analysis determines the overall emotional tone (positive, negative, neutral) of the text, not a summary of its content.

Full explanation →

269

MCQhard

You are designing a solution to detect brand logos in social media images. The logos vary in size and orientation. You need to achieve high accuracy with minimal false positives. Which approach should you recommend?

A.Use Azure Computer Vision Describe API to generate captions and filter by logo mentions.

B.Train an Azure Custom Vision object detection model with labeled logo images.

C.Use Azure Computer Vision Analyze API with domain-specific models.

D.Use Azure Form Recognizer to extract logo positions from images.

AnswerB

Object detection returns bounding boxes, so it handles logos that vary in size and orientation, unlike classification which only labels the whole image. Training on your labelled logo images tunes the model to your brands, satisfying the high-accuracy, low-false-positive constraint.

Why this answer

Azure Custom Vision allows you to train a custom object detection model with your own labeled dataset of brand logos, enabling high accuracy for specific logo shapes, sizes, and orientations. This approach directly addresses the need for minimal false positives by learning the exact visual features of the logos, unlike generic pre-built models.

Exam trap

The trap here is that candidates confuse Azure Computer Vision's pre-built domain-specific models (which cover only landmarks, celebrities, and general objects) with the ability to detect custom logos, leading them to choose option C instead of recognizing that Custom Vision is required for custom object detection.

How to eliminate wrong answers

Option A is wrong because the Describe API generates natural language captions and is not designed for precise object detection or localization; filtering by logo mentions would be unreliable and produce many false positives. Option C is wrong because the Analyze API with domain-specific models (e.g., landmarks, celebrities) does not include a pre-built model for brand logos, so it cannot detect arbitrary logos with high accuracy. Option D is wrong because Azure Form Recognizer is specialized for extracting text and structured data from documents (e.g., invoices, forms), not for detecting or localizing visual objects like logos in images.

Full explanation →

270

MCQmedium

You are testing an Azure OpenAI chat completion. The response shown in the exhibit is returned. What does the finish_reason of 'content_filter' indicate?

A.The model's response was blocked by the content filter.

B.There was a system error during processing.

C.The user's prompt was flagged by the content filter.

D.The model refused to answer due to insufficient data.

AnswerA

The content_filter finish reason means Azure OpenAI's responsible AI filtering intercepted the completion, so no usable text was returned. The prompt itself was accepted; the generated output was flagged and suppressed, which is distinct from length or stop-sequence termination.

Why this answer

The 'content_filter' finish_reason indicates that the Azure OpenAI content filtering system detected that the model's generated response violated one of the configured content policies (e.g., hate, violence, self-harm, sexual content). The response was therefore blocked before being returned to the user, and the finish_reason explicitly signals this filtering action rather than a normal completion or a stop due to token limits.

Exam trap

The trap here is that candidates often confuse 'content_filter' with prompt rejection, but the finish_reason specifically indicates the model's output was blocked, not the user's input.

How to eliminate wrong answers

Option B is wrong because a system error during processing would return a different finish_reason such as 'error' or an HTTP 500 status, not 'content_filter'. Option C is wrong because the content_filter finish_reason applies to the model's response, not the user's prompt; if the prompt were flagged, the API would typically return a 400 error with a content filter violation message before any generation occurs. Option D is wrong because the model refusing to answer due to insufficient data would be indicated by a finish_reason of 'stop' (if the model generated a refusal message) or by a specific response text, not by the 'content_filter' reason.

Full explanation →

271

Matchingmedium

Match each Azure Cognitive Search skill to its capability.

Drag a concept onto its matching description — or click a concept then click the description.

Concepts

Matches

Extract text from images

Identify entities like people or organizations

Extract key phrases from text

Detect language of text

Determine sentiment of text

Why these pairings

Common Azure Cognitive Search skills include Entity Recognition, Key Phrase Extraction, OCR (for image text extraction), and Sentiment Analysis. The correct matches are A and B. Options C and D are mismatched definitions.

Full explanation →

272

Multi-Selecthard

You are architecting an Azure AI solution that uses Azure AI Language to analyze text for sentiment and key phrases. The solution must handle bursts of up to 500 requests per second but average only 50 requests per second. You need to ensure cost efficiency while meeting performance requirements. Which THREE actions should you take?

Select 3 answers

A.Use Azure Queue Storage to buffer requests during spikes

B.Select the Free tier and implement queuing

C.Use the S0 pricing tier

D.Implement client-side throttling and retry logic

E.Deploy the service in multiple regions

AnswersA, C, D

Queue Storage decouples ingestion from processing, absorbing the 500 requests-per-second bursts so the Language service is called at a sustainable rate. This smooths spikes and avoids provisioning for peak throughput, keeping costs aligned with the 50 requests-per-second average.

Why this answer

Option A is correct because Azure Queue Storage decouples the request producers from the Azure AI Language service, buffering the burst of up to 500 requests per second so the service can process them at a sustainable rate rather than being overwhelmed. Option C is correct because the S0 (Standard) pricing tier is the paid tier that supports the throughput and quota needed for production workloads, whereas the Free tier is limited to 5,000 transactions per month at 20 transactions per minute, which cannot handle 50 requests per second on average. Option D is correct because client-side throttling and retry logic (for example, honoring HTTP 429 responses and using exponential backoff) prevents the application from exceeding the service's rate limits and gracefully handles transient throttling during spikes.

Option B is not correct because the Free tier's low transaction and rate limits make it unsuitable for this workload even with queuing. Option E is not correct because deploying in multiple regions increases cost and complexity without addressing the burst-handling and cost-efficiency requirements, since the service's per-region rate limits would still apply.

Exam trap

The trap here is that candidates often assume a higher pricing tier (like S0) alone can handle bursts, but without queuing and retry logic, the service will still throttle requests, leading to failures or the need for costly over-provisioning.

Full explanation →

273

MCQhard

Your company is deploying an Azure AI solution that uses multiple AI services. You need to ensure that all API calls are authenticated securely using managed identities. Which of the following steps is required to enable managed identity authentication for an Azure AI service?

A.Enable system-assigned managed identity at the subscription level

B.Assign a managed identity to the Azure AI service and grant it the required RBAC role

C.Store the authentication key in Azure Key Vault and reference it in the application

D.Configure a shared access key in the Azure AI service

AnswerB

Managed identity authentication requires two elements: a managed identity assigned to the Azure AI resource, plus an RBAC role assignment granting that identity access to the target service. Without the role grant, token acquisition succeeds but authorisation fails.

Why this answer

Managed identity authentication for Azure AI services requires assigning either a system-assigned or user-assigned managed identity to the resource, and then granting that identity the appropriate RBAC role (e.g., Cognitive Services User) on the target AI service. This eliminates the need for keys or secrets in the code and leverages Azure AD tokens for secure, passwordless authentication.

Exam trap

The trap here is that candidates often confuse managed identity with key management solutions like Key Vault, or assume that enabling a managed identity at a higher scope (subscription) automatically applies to all resources, when in fact the identity must be explicitly assigned to each resource and granted RBAC permissions.

How to eliminate wrong answers

Option A is wrong because managed identities are assigned at the resource level, not the subscription level; enabling a system-assigned identity at the subscription scope is not a valid operation. Option C is wrong because storing the authentication key in Key Vault is a valid security practice but does not use managed identity authentication—it still relies on a static key, not an Azure AD token. Option D is wrong because shared access keys are the traditional key-based authentication method, which managed identities are designed to replace; configuring a shared access key does not enable managed identity authentication.

Full explanation →

274

MCQhard

A healthcare portal must summarize patient feedback stored in Azure Blob Storage. Documents are plain UTF-8 text and the summaries must be generated without sending content to a public endpoint. You are building the pipeline with Azure AI Language and a private network. Which configuration should you use to call extractive summarization while meeting the network requirement?

A.Create an Azure AI Language resource with a private endpoint and disable public network access, then call the REST API from a client inside the virtual network.

B.Deploy the open-source summarization container for Azure AI Language on an Azure Kubernetes Service cluster and call its local endpoint.

C.Create a standard Azure AI Language resource and authenticate with an account key, relying on TLS to protect the content in transit.

D.Use the Azure AI Language resource with a service tag firewall rule that allows only the portal's app service outbound IP addresses.

AnswerA

A private endpoint places the Azure AI Language resource inside your virtual network and disabling public network access blocks internet traffic, so REST calls from an in-network client never traverse a public endpoint. This satisfies the requirement while still allowing the extractive summarization operation, which is available through the standard Language REST API.

Why this answer

Meeting a no-public-endpoint requirement means the Azure AI Language resource must be reachable only through private networking. Combining a private endpoint with disabled public network access forces all calls, including extractive summarization, to travel inside the virtual network, which keeps document content off the public internet while preserving full API functionality.

Exam trap

The trap here is treating encryption in transit or IP allowlisting as equivalent to eliminating the public endpoint entirely.

Full explanation →

275

Multi-Selectmedium

You plan to deploy a custom named entity recognition (NER) project in Azure AI Language. The project must extract supplier names and contract identifiers from procurement documents. You need to prepare the project so that the model can be trained and evaluated before deployment. Which two actions should you perform? (Choose two.)

Select 2 answers

A.Define a list of intents that correspond to each procurement document type.

B.Upload and label training documents by selecting the text spans for SupplierName and ContractId entities.

C.Create a question answering knowledge base that contains the procurement documents.

D.Create a project of type Custom named entity recognition in Azure AI Language and select the language for the documents.

E.Enable opinion mining so the model can detect sentiment around supplier mentions.

AnswersB, D

Custom NER learns from labeled spans, so documents must be uploaded and the exact text for each entity marked. Labeling SupplierName and ContractId spans teaches the model the boundaries and context of those entities, which is required before training and evaluation can produce meaningful results.

Why this answer

To train a custom NER model, you create a project of the custom named entity recognition type with the correct language, then upload and label documents by marking the spans for each entity such as SupplierName and ContractId. Training and evaluation follow, and only then can the model be deployed for extraction.

Exam trap

The trap here is mixing features from other Azure AI Language project types, such as intents or knowledge bases, into a custom NER workflow that only requires labeled entity spans.

Full explanation →

276

MCQmedium

You run the Azure CLI command 'az search indexer list --search-service mysearch --query "[].{name:name, status:status, lastResult:lastResult}"' and get the above output. Your indexer shows 5 warnings. What should you do to investigate the warnings?

A.Run 'az search indexer run --name myindexer' to trigger a new run.

B.Run 'az search indexer show --name myindexer' and review the 'warnings' array in the output.

C.Ignore the warnings because they are not errors.

D.Run 'az search indexer reset --name myindexer' to reset the indexer.

AnswerB

Indexer list output omits warning detail. Running az search indexer show for that indexer returns the full execution history, including the warnings array, which names each affected document and the underlying skill or mapping issue.

Why this answer

The 'az search indexer list' command with the query you used returns a summary of indexer status and last result, but it does not include the detailed 'warnings' array. To investigate the 5 warnings, you need to use 'az search indexer show --name myindexer', which returns the full indexer execution history, including a 'warnings' array that lists each warning with its message and details. This allows you to understand the nature of each warning and take corrective action if needed.

Exam trap

The trap here is that candidates assume the 'list' command provides full details, but it only returns a filtered projection; the 'show' command is required to access the nested 'warnings' array, which is a common pattern in Azure CLI where list commands return summaries and show commands return full objects.

How to eliminate wrong answers

Option A is wrong because 'az search indexer run' triggers a new execution but does not retrieve or display existing warnings; it would only produce a new set of warnings or errors. Option C is wrong because warnings in Azure Cognitive Search indexers often indicate issues like field mapping conflicts, data truncation, or unsupported types that can degrade indexing quality or cause silent data loss, so they should not be ignored. Option D is wrong because 'az search indexer reset' resets the indexer's change tracking state, forcing a full reindex of all documents, which is an aggressive action unrelated to investigating warnings.

Full explanation →

277

MCQmedium

A company is deploying an agent built with Azure AI Foundry Agent Service to production. The agent must log all interactions for auditing and compliance. The team needs to capture the full conversation history, including tool calls and their results. Which approach should they use?

A.Use the agent's built-in conversation history feature and periodically export it manually via the portal.

B.Configure the agent to write conversation logs to a local file on the client application.

C.Instruct the agent to include a summary of each interaction in its response for the user to save.

D.Enable diagnostic settings on the Azure AI Foundry resource to send logs to Azure Monitor and Log Analytics.

AnswerD

Enabling diagnostic settings on the Azure AI Foundry resource allows you to stream logs, including agent interactions and tool calls, to Azure Monitor and Log Analytics. This provides a centralized, durable audit trail. You can then query and analyze logs for compliance. This is the recommended approach for auditing agent interactions in production.

Why this answer

Enabling diagnostic settings to send logs to Azure Monitor and Log Analytics is the correct approach because it provides a centralized, durable, and queryable audit trail of all agent interactions, including tool calls. Other methods are either local, manual, or incomplete, and do not meet compliance needs for production auditing.

Exam trap

The trap here is relying on the agent's built-in conversation history for auditing, but that history is session-scoped and not designed for compliance logging.

Full explanation →

278

MCQhard

You are deploying an Azure AI solution that uses Azure AI Search. The solution must be able to access the search service from an Azure virtual network without traversing the public internet. You need to configure the search service to meet this requirement. What should you do?

A.Configure the search service to use an IP firewall and add the virtual network's public IP address.

B.Create a private endpoint for the search service and configure the virtual network to use it.

C.Deploy the search service with a public endpoint and use Azure Front Door to route traffic privately.

D.Enable a service endpoint for Azure AI Search on the virtual network subnet.

AnswerB

A private endpoint creates a network interface in your virtual network, allowing private connectivity to Azure AI Search. This enables access from within the virtual network without going over the public internet. You must also configure the search service to deny public access and allow the private endpoint. This satisfies the requirement for private network access.

Why this answer

To access Azure AI Search privately from a virtual network, you must use a private endpoint, which assigns a private IP address from your subnet to the search service. This ensures all traffic remains within the Azure backbone. Other options either use public internet or do not provide a private IP.

Exam trap

The trap here is assuming that service endpoints or IP firewalls provide private connectivity, when they actually still use public IP addresses.

Full explanation →

279

MCQhard

A company uses Azure AI Vision Image Analysis to generate captions for product photos. The solution must return a caption in English and a confidence score for each image. You call the Image Analysis API with the caption feature. The response does not include a confidence score. What should you do to obtain confidence scores for the captions?

A.Use the older Computer Vision v3.2 Analyze Image API with the 'description' visual feature.

B.Call the Image Analysis API twice and compare the captions to derive a confidence score.

C.Use the denseCaptions feature instead of captions.

D.Set the language parameter to 'en' and include the 'confidence' query parameter.

AnswerA

The legacy Computer Vision v3.2 Analyze Image API with the description visual feature returns a caption along with a confidence score between 0 and 1. The newer Image Analysis 4.0 caption feature does not include a confidence score. To meet the requirement, you must call the v3.2 endpoint, which still provides the confidence value for the generated description.

Why this answer

The legacy Computer Vision v3.2 Analyze Image API with the description feature returns a caption and a confidence score, which the newer Image Analysis 4.0 caption feature does not. To obtain confidence scores for captions, you must use the older API version. Other options either use features that lack confidence scores or rely on invalid parameters or workarounds that cannot produce a true confidence value.

Exam trap

The trap here is assuming that the newest Image Analysis caption feature includes a confidence score, when only the legacy v3.2 description feature returns one.

Full explanation →

280

MCQeasy

You are creating an agent in Microsoft Foundry that answers questions about internal HR policies. The policy documents are already indexed in Azure AI Search, and you want the fastest path to a working agent without writing retrieval code. What should you do first?

A.Upload the policy documents to the agent's file storage and enable file search.

B.Export the policies to CSV and attach them as a code interpreter file.

C.Build an Azure Functions app that queries the index and expose it as a function tool.

D.Create the agent, then connect the existing Azure AI Search index as a knowledge source for the agent.

AnswerD

Because the documents are already indexed, connecting that index as a knowledge source is the minimal-effort path. The agent gains retrieval over the policy content immediately, with the service handling queries and context injection, so no custom retrieval code is needed and the existing index investment is reused.

Why this answer

Connecting the existing Azure AI Search index as a knowledge source gives the agent retrieval over the HR policies with no custom retrieval code. It reuses the index the organization already maintains, so content updates flow through automatically and the agent is functional with minimal configuration.

Exam trap

The trap here is reaching for a custom function or file upload when an existing search index can be connected directly as a knowledge source.

Full explanation →

281

MCQmedium

A company uses Microsoft Copilot Studio to create an agent that helps employees schedule meetings. The agent must access the user's calendar to find free time slots and book meetings. The agent should only work for users who have granted consent. Which authentication and authorization approach should be used?

A.Configure OAuth 2.0 authentication with Microsoft Entra ID and request delegated permissions for Microsoft Graph.

B.Use certificate-based authentication for the agent.

C.Use API key authentication to call Microsoft Graph.

D.Use OAuth 2.0 client credentials flow with application permissions.

AnswerA

OAuth 2.0 with Microsoft Entra ID and delegated Microsoft Graph permissions ensures the agent acts only on behalf of users who granted consent, scoping calendar access to their own free/busy data. Application permissions would bypass per-user consent, violating the requirement.

Why this answer

The agent needs to act on behalf of a signed-in user (delegated identity) to access their calendar. OAuth 2.0 with Microsoft Entra ID and delegated permissions for Microsoft Graph allows the agent to request only the scopes (e.g., Calendars.ReadWrite) that the user has consented to, ensuring the agent operates within the user's granted permissions.

Exam trap

The trap here is that candidates often confuse delegated permissions (user-context) with application permissions (tenant-wide), and mistakenly choose the client credentials flow (Option D) because it seems simpler, but it violates the explicit requirement for per-user consent.

How to eliminate wrong answers

Option B is wrong because certificate-based authentication is a method for establishing the identity of the agent itself (client credential), not for obtaining user-delegated access to a resource like a calendar; it does not support per-user consent. Option C is wrong because API key authentication is not supported by Microsoft Graph; Microsoft Graph requires OAuth 2.0 tokens and does not accept static API keys. Option D is wrong because the OAuth 2.0 client credentials flow uses application permissions, which grant the agent tenant-wide access to all users' calendars without per-user consent, violating the requirement that the agent should only work for users who have granted consent.

Full explanation →

282

MCQhard

You are responsible for an Azure AI solution that uses Custom Vision to classify manufacturing defects. The model must achieve high recall to avoid missing defects. The current model has high precision but low recall. Which action should you take?

A.Add more images of non-defective items to the training set

B.Increase the number of training iterations

C.Lower the probability threshold for the defect class

D.Increase the probability threshold for the defect class

AnswerC

Lowering the probability threshold classifies more samples as defects, raising recall by catching defects previously missed, at the cost of some precision. This directly addresses the high-precision, low-recall imbalance, prioritising detection of defects over avoiding false positives.

Why this answer

Lowering the probability threshold for the defect class means the model will classify an image as defective even when its confidence score is lower. This increases the number of true positives (defects caught), directly improving recall at the cost of potentially more false positives. In Custom Vision, the default probability threshold is 50%, and adjusting it downward is the standard technique to prioritize recall over precision.

Exam trap

The trap here is that candidates confuse precision and recall, often assuming that increasing the threshold (making the model stricter) will improve overall performance, when in fact it reduces recall by missing more defects.

How to eliminate wrong answers

Option A is wrong because adding more images of non-defective items would bias the model toward the non-defective class, likely reducing recall further by making the model more conservative in predicting defects. Option B is wrong because increasing the number of training iterations (epochs) primarily helps the model converge better on the training data but does not directly control the precision-recall trade-off; it may even lead to overfitting without improving recall. Option D is wrong because increasing the probability threshold for the defect class would require higher confidence to classify a defect, which reduces false positives but also reduces true positives, thereby lowering recall even further.

Full explanation →

283

MCQhard

You are implementing a conversational language understanding (CLU) project in Azure AI Language. Your utterances include entities that are sometimes a single word and sometimes a multi-word phrase, such as 'New York' and 'San Francisco'. You need the model to correctly capture these multi-word entities during training and prediction. Which entity component type should you use?

A.Prebuilt entity component

B.Learned entity component

C.Regex entity component

D.List entity component

AnswerB

Learned entities use labeled examples to train the model to recognize entity spans, including multi-word phrases, based on context. This allows the model to generalize to new phrasings such as 'New York' or 'San Francisco' without requiring exact matches in a list.

Why this answer

Learned entity components in CLU are trained from labeled utterances, allowing the model to identify entity spans based on surrounding context. This is the correct choice for multi-word entities such as city names that vary in phrasing and cannot be captured reliably by exact-match lists or fixed regex patterns.

Exam trap

The trap here is confusing list or regex entities, which require explicit patterns or synonyms, with learned entities that generalize from labeled examples.

Full explanation →

284

MCQeasy

A developer is building a generative AI application with Azure OpenAI Service. The application must stream partial responses to users as tokens are generated, rather than waiting for the complete response. Which parameter should the developer set in the chat completion request?

A.Set 'temperature' to 0.

B.Set 'stream' to true.

C.Set 'max_tokens' to a large value.

D.Set 'n' to a value greater than 1.

AnswerB

Setting stream to true enables server-sent events so the service returns incremental chunks as tokens are produced. The client can render partial text immediately, which improves perceived latency for long generations. This is the documented parameter for streaming chat completions in Azure OpenAI Service.

Why this answer

Streaming is controlled by the stream parameter, which when true causes the service to emit incremental delta chunks over a persistent connection. This lets the application display tokens as they are generated, reducing perceived latency while keeping the same model and prompt configuration.

Exam trap

The trap here is confusing parameters that shape output content, such as temperature or max_tokens, with the parameter that changes how output is delivered over the network.

Full explanation →

285

MCQhard

You are using Azure OpenAI Service to generate product descriptions. You notice that the model occasionally outputs descriptions that contain factual inaccuracies about product specifications. You want to reduce these hallucinations without changing the model. What should you do?

A.Increase the frequency_penalty parameter.

B.Decrease the temperature parameter.

C.Increase the max_tokens parameter.

D.Provide the product specifications in the prompt and use the system message to instruct the model to base answers on them.

AnswerD

Grounding the model with factual data in the prompt reduces hallucinations by providing accurate context.

Why this answer

Providing the product specifications directly in the prompt and using the system message to instruct the model to base its answers on them grounds the generation in factual data, reducing hallucinations. This technique, known as 'grounding' or 'retrieval-augmented generation' (RAG), does not modify the model itself but constrains its output to the provided context, which is the only way to reduce factual inaccuracies without changing model parameters.

Exam trap

The trap here is that candidates often confuse hyperparameter tuning (like temperature or frequency_penalty) with prompt engineering techniques, mistakenly believing that adjusting randomness or repetition penalties can fix factual hallucinations, when only providing the correct context in the prompt can do so without model changes.

How to eliminate wrong answers

Option A is wrong because increasing frequency_penalty reduces repetition of tokens by penalizing tokens that have already appeared, which does not address factual accuracy or hallucinations. Option B is wrong because decreasing temperature makes the model more deterministic and less creative, but it does not prevent the model from generating plausible-sounding but factually incorrect statements about product specifications. Option C is wrong because increasing max_tokens only allows longer responses, which can actually increase the chance of hallucinations by giving the model more opportunity to generate unsupported content.

Full explanation →

286

MCQeasy

Your team is building a mobile app that uses Azure Custom Vision to classify plant species. The app must work offline and sync labeled images when connectivity is restored. Which SDK feature should you use?

A.Azure IoT Edge runtime on the phone

B.Azure API Management with caching

C.Export the model as a TensorFlow or CoreML model for on-device inference

D.Continuous deployment integration

AnswerC

Exported model runs offline.

Why this answer

Azure Custom Vision allows exporting trained models to formats like TensorFlow, CoreML, ONNX, or Docker for on-device inference. This enables the mobile app to run classification locally without network connectivity, and the Custom Vision SDK includes a method to upload labeled images for offline training sync when connectivity is restored.

Exam trap

The trap here is that candidates confuse offline inference with edge computing (IoT Edge) or API caching, not realizing that Custom Vision's export feature is the only option that provides a local model for on-device classification without requiring a network connection.

How to eliminate wrong answers

Option A is wrong because Azure IoT Edge runtime is designed for edge devices like gateways or industrial controllers, not for mobile phones, and it does not provide offline inference or image sync capabilities for Custom Vision. Option B is wrong because Azure API Management with caching only caches API responses to reduce latency, but it does not enable offline model execution or local image storage and sync. Option D is wrong because continuous deployment integration automates model deployment pipelines but does not address offline inference or offline image labeling and sync on a mobile device.

Full explanation →

287

MCQhard

A developer is building an agent using the Microsoft Bot Framework SDK in C#. The agent must authenticate users via Microsoft Entra ID and maintain state across conversations. The solution must store user preferences (e.g., language, timezone) in Azure Cosmos DB. Which state management approach should the developer use?

A.Use the Bot State Service (deprecated).

B.Use UserState with Blob Storage.

C.Use ConversationState with Memory Storage.

D.Use UserState with Cosmos DB Storage.

AnswerD

UserState scopes data per user across conversations, and Cosmos DB Storage persists it durably, satisfying both the Entra ID authentication and cross-conversation preference requirements. ConversationState alone would lose language and timezone settings once the session ends.

Why this answer

The developer needs to persist user preferences across conversations, which requires UserState (not ConversationState, which is scoped to a single conversation). Cosmos DB Storage is the appropriate choice for durable, scalable, and low-latency storage of user-specific data, and it integrates directly with the Bot Framework SDK's `CosmosDbPartitionedStorage` class.

Exam trap

The trap here is confusing UserState (persistent across conversations) with ConversationState (temporary per conversation), leading candidates to incorrectly choose ConversationState with Memory Storage, which loses data when the bot restarts.

How to eliminate wrong answers

Option A is wrong because the Bot State Service was deprecated and is no longer supported; using it would violate the requirement for a modern, supported solution. Option B is wrong because Blob Storage is designed for large unstructured data (e.g., files, images) and is not optimized for the small, frequent read/write operations typical of user state in a bot; Cosmos DB is the recommended storage for state data. Option C is wrong because ConversationState is scoped to a single conversation and does not persist across conversations, so it cannot store user preferences that must be available across multiple sessions.

Full explanation →

288

MCQmedium

Your organization uses Microsoft Entra ID for identity management. You are building an AI solution that uses Azure AI Vision to analyze images. The solution must use managed identities to authenticate to the Vision resource. Which RBAC role should you assign to the managed identity?

A.Owner

B.Cognitive Services User

C.Reader

D.Contributor

AnswerB

Cognitive Services User grants data-plane access to call Azure AI Vision's analyse operations without granting key management or resource administration. Assigning it to the managed identity lets the solution authenticate to the Vision resource via Microsoft Entra ID, meeting the managed-identity constraint.

Why this answer

The Cognitive Services User role (B) is the correct RBAC role because it grants the minimum required permissions for a managed identity to call Azure AI Vision APIs (e.g., analyze image, OCR) without allowing any write or management operations. This role is specifically designed for accessing Azure Cognitive Services endpoints, and it aligns with the principle of least privilege for authentication via managed identities.

Exam trap

The trap here is that candidates often confuse the Reader role (which grants read access to the resource's Azure Resource Manager properties) with the ability to read data from the service, but Reader does not include data-plane permissions for Cognitive Services APIs.

How to eliminate wrong answers

Option A (Owner) is wrong because it grants full control over the Vision resource, including the ability to delete or modify the resource itself, which is excessive and violates security best practices for a managed identity that only needs to call APIs. Option C (Reader) is wrong because it only allows read access to the resource's metadata and configuration (e.g., viewing keys or endpoints) but does not grant permission to call the Vision API endpoints for image analysis. Option D (Contributor) is wrong because it allows creating and managing resources (e.g., deploying models or changing settings) but does not include the specific 'Cognitive Services User' data-plane permission required to authenticate and invoke the Vision API.

Full explanation →

289

MCQeasy

You are planning an Azure AI solution that will use several Azure AI services, including Azure AI Vision and Azure AI Language. The solution must be deployed to multiple regions and must provide a single endpoint and key for all the services. Which Azure resource type should you create?

A.An Azure API Management instance with a backend for each service

B.An Azure Machine Learning workspace with online endpoints

C.An Azure AI services multi-service account

D.A separate Azure AI services resource for each service in each region

AnswerC

A multi-service account provides one endpoint and one key that can be used across supported Azure AI services such as Vision and Language. It also supports multi-region deployment by creating the account in each region while keeping a consistent management model. This directly satisfies the requirement for a single endpoint and key across several services.

Why this answer

A multi-service account is designed to expose several Azure AI services through one endpoint and one key. It supports deployment in multiple regions while keeping a consistent management and authentication model. Separate resources or an API Management layer would still leave you managing per-service credentials, so they do not satisfy the single-key requirement.

Exam trap

The trap here is confusing an API gateway such as API Management with a unified Azure AI services key, when the gateway still requires managing a separate credential for each backend.

Full explanation →

290

MCQhard

You are deploying an Azure OpenAI model for a healthcare application. You need to ensure that the model does not generate medical advice and that all responses include a disclaimer. Which configuration should you use?

A.Ground the model with your own medical documents.

B.Set max_tokens to 50 to limit response length.

C.Use Azure AI Content Safety to filter medical terms.

D.Configure a system message with instructions and enable content filtering.

AnswerD

A system message sets the model's behavioural guardrails, instructing it to refuse medical advice and append a disclaimer to every response. Content filtering adds a safety layer blocking harmful outputs. Together they satisfy both constraints: no medical advice and mandatory disclaimers in a healthcare deployment.

Why this answer

Configuring a system message with explicit instructions (e.g., 'Do not provide medical advice; always include a disclaimer') combined with Azure AI Content Safety's content filtering allows you to enforce behavioral guardrails and block harmful outputs at the application layer. The system message sets the model's behavior, while content filtering provides a secondary safety net to catch policy violations, ensuring compliance in a regulated healthcare environment.

Exam trap

The trap here is that candidates confuse content filtering with behavioral control, assuming Azure AI Content Safety can enforce custom rules like 'do not generate medical advice' when it only filters predefined harmful categories, not domain-specific instructions.

How to eliminate wrong answers

Option A is wrong because grounding the model with medical documents (e.g., via Azure OpenAI on your data) does not prevent the model from generating medical advice; it only improves factual accuracy by referencing your data, but the model can still produce advice or omit disclaimers. Option B is wrong because setting max_tokens to 50 limits response length but does not control the content or ensure a disclaimer is included; the model could still generate medical advice within that token limit. Option C is wrong because Azure AI Content Safety filters harmful content based on predefined categories (e.g., hate, violence), but it does not have a built-in 'medical terms' filter; it cannot enforce a custom rule like 'do not generate medical advice' or 'include a disclaimer'.

Full explanation →

291

Multi-Selecthard

A media company is building an Azure AI Foundry agent that generates summaries of news articles. The agent uses a tool to fetch articles from an internal CMS. The company wants to ensure the agent respects content usage rights and does not summarize articles that are marked as restricted. The agent must also log every article it accesses for auditing. Which two actions should the team take? (Choose two.)

Select 2 answers

A.Instruct the agent in its system message to ignore articles that are marked restricted.

B.Enable diagnostic logging on the Azure AI Foundry resource to capture all tool calls and their responses.

C.Use a content filter to block articles that contain certain keywords associated with restricted content.

D.Store a copy of every article in the agent's conversation history for later review.

E.Configure the CMS tool to return only articles with a usage rights field set to 'public' or 'licensed', and have the agent filter based on that field.

AnswersB, E

Diagnostic logging captures tool invocations and responses, providing an audit trail of every article the agent accesses. This satisfies the auditing requirement by recording which articles were fetched and when. It works in conjunction with access controls to ensure compliance, and it does not expose restricted content because the filtering already prevents such articles from being returned.

Why this answer

To respect usage rights, the CMS tool should filter articles based on the usage rights field before they reach the agent, ensuring only permissible content is summarized. To audit access, diagnostic logging on the Azure AI Foundry resource captures all tool calls and responses. Together, these actions enforce policy at the data layer and provide a verifiable audit trail without exposing restricted content to the model.

Exam trap

The trap here is relying on prompt instructions to enforce content restrictions, which is not deterministic, instead of implementing access control at the tool or data source level.

Full explanation →

292

MCQeasy

A news organization wants to automatically summarize long articles into short, coherent summaries. The solution must preserve the original meaning and key points. Which Azure AI service should be used?

A.Azure AI Document Intelligence

B.Azure AI Language - Key Phrase Extraction

C.Azure AI Language - Extractive Summarization

D.Azure AI Translator

AnswerC

Extractive summarization selects and ranks the most salient existing sentences from the source article, guaranteeing the summary preserves original wording and key points. This directly satisfies the requirement to maintain meaning without generative paraphrasing, which could introduce distortion.

Why this answer

Azure AI Language's Extractive Summarization is specifically designed to generate concise summaries by extracting the most important sentences from a document while preserving the original meaning and key points. This service uses natural language processing to rank sentences based on relevance and coherence, making it ideal for summarizing long articles without altering the original content.

Exam trap

The trap here is that candidates often confuse Key Phrase Extraction (Option B) with summarization, but Key Phrase Extraction only returns isolated terms, not a coherent summary, whereas Extractive Summarization returns full sentences that preserve meaning.

How to eliminate wrong answers

Option A is wrong because Azure AI Document Intelligence (formerly Form Recognizer) is optimized for extracting structured data (e.g., tables, key-value pairs) from documents, not for generating textual summaries. Option B is wrong because Azure AI Language - Key Phrase Extraction identifies individual keywords or phrases, not coherent summaries; it lacks the sentence-level extraction and ranking needed for summarization. Option D is wrong because Azure AI Translator focuses on translating text between languages, not summarizing content in the same language.

Full explanation →

293

MCQeasy

You are deploying a generative AI feature that drafts marketing copy with an Azure OpenAI GPT-4o deployment. The application must stream partial responses to the browser so users see text as it is generated, rather than waiting for the full completion. What should you enable in your API call?

A.Set the stream parameter to true and consume the server-sent event chunks returned by the chat completions endpoint.

B.Enable the asynchronous batch API and retrieve results from the output file once the job finishes.

C.Set a high max_tokens value and poll the operation status endpoint until the completion is marked complete.

D.Increase the deployment's tokens-per-minute quota so responses are generated faster.

AnswerA

The chat completions API supports streaming by setting stream to true, after which the service returns incremental server-sent event chunks containing delta content. The client renders each delta as it arrives, which produces the typewriter effect the scenario requires. This is the standard, supported mechanism for partial response delivery in Azure OpenAI.

Why this answer

Streaming in Azure OpenAI is opt-in per request: setting the stream flag makes the chat completions endpoint return incremental deltas over server-sent events, which the browser can render progressively. Polling, batch processing, and quota changes all describe non-interactive or throughput-oriented behaviors that do not deliver partial tokens as they are generated.

Exam trap

The trap here is confusing throughput tuning, such as raising quota or max tokens, with response delivery mode, which only the stream parameter changes.

Full explanation →

294

Multi-Selecteasy

A company is planning to use Azure AI services. They require the ability to audit all API calls for compliance. Which THREE components should they enable?

Select 3 answers

A.Azure Monitor

B.Log Analytics workspace

C.Diagnostic settings for the AI service

D.Azure RBAC roles for the service

E.Managed identity for the service

AnswersA, B, C

Azure Monitor captures diagnostic logs and activity logs for Azure AI services, recording every API call with caller identity, timestamp and operation. Enabling diagnostic settings routes these logs to a Log Analytics workspace, satisfying the compliance requirement to audit all API calls. It also supports alerting on suspicious access patterns.

Why this answer

Azure Monitor (A) is correct because it is the platform that collects, stores, and surfaces activity and resource logs, enabling the audit trail of API calls required for compliance. A Log Analytics workspace (B) is correct because it is the destination where diagnostic logs are ingested and queried with KQL, allowing retention and analysis of all API call records. Diagnostic settings for the AI service (C) is correct because it is the mechanism that routes the service's resource logs (e.g., Audit, RequestResponse) and metrics to Azure Monitor/Log Analytics, which is what actually captures the API calls.

Azure RBAC roles (D) only control authorization to the resource and do not record API call history, and a managed identity (E) only provides an authentication credential for the service, so neither produces the audit data required.

Exam trap

The trap here is that candidates often confuse auditing (logging API calls) with security controls like RBAC or Managed Identity, mistakenly thinking that controlling access or identity automatically provides an audit trail, whereas auditing requires explicit diagnostic logging and a monitoring pipeline.

Full explanation →

295

MCQhard

Your Azure AI Search index stores customer support tickets. You need to implement a search feature that returns semantically similar results even if the query uses different wording. Which configuration should you enable?

A.Use simple query parsing with searchMode=any

B.Add a synonym map with custom entries

C.Enable semantic search and configure a semantic configuration

D.Enable fuzzy search on the index

AnswerC

Semantic search applies a reranking model over results, using a semantic configuration that maps title, content and keyword fields. This lets queries match by meaning rather than exact terms, satisfying the requirement for semantically similar results despite different wording.

Why this answer

Semantic search in Azure AI Search uses deep neural networks to understand the intent and context of a query, returning results that are semantically similar even when the wording differs. By enabling semantic search and configuring a semantic configuration, you define which fields are used for summarization and ranking, which directly addresses the requirement for meaning-based matching rather than keyword matching.

Exam trap

The trap here is that candidates often confuse synonym maps (which handle predefined word equivalence) with semantic search (which handles contextual meaning), leading them to choose synonym maps when the question explicitly requires handling of different wording beyond simple synonyms.

How to eliminate wrong answers

Option A is wrong because simple query parsing with searchMode=any only controls how terms are combined (OR logic) and does not provide any semantic understanding or synonym expansion. Option B is wrong because a synonym map expands queries to include predefined equivalent terms, but it cannot handle novel or context-dependent paraphrasing that semantic search can. Option D is wrong because fuzzy search corrects for typos and minor spelling variations by using Levenshtein distance, but it does not capture semantic similarity between different words or phrases.

Full explanation →

296

MCQmedium

You are building a knowledge mining solution for legal documents using Azure AI Search. The solution must extract entities like dates, organizations, and persons from PDF files and index them. Which built-in skill should you add to the skillset to perform this extraction?

A.Named Entity Recognition skill

B.Language Detection skill

C.Optical Character Recognition (OCR) skill

D.Key Phrase Extraction skill

AnswerA

The Named Entity Recognition skill calls Azure AI Language to extract persons, organisations and dates from document text, writing them into the index as enriched fields. This directly satisfies the requirement to index those entity types from PDFs.

Why this answer

The Named Entity Recognition (NER) skill in Azure AI Search is specifically designed to extract entities such as dates, organizations, and persons from text. When added to a skillset, it processes the content extracted from PDF files and outputs structured entity information that can be indexed and queried. This directly matches the requirement to extract and index named entities from legal documents.

Exam trap

The trap here is that candidates often confuse entity extraction with key phrase extraction or OCR, mistakenly thinking that extracting 'important terms' or 'text from images' is equivalent to identifying specific named entities like dates and organizations.

How to eliminate wrong answers

Option B is wrong because Language Detection skill identifies the language of text (e.g., English, French) but does not extract specific entities like dates or organizations. Option C is wrong because Optical Character Recognition (OCR) skill extracts text from images or scanned PDFs, but it does not perform entity extraction; it only converts visual text into machine-readable text. Option D is wrong because Key Phrase Extraction skill identifies important phrases or topics in text, not specific named entities such as persons, organizations, or dates.

Full explanation →

297

MCQmedium

Refer to the exhibit. You have trained an object detection model in Azure Custom Vision. The model is published as 'defect-model'. You need to deploy this model to a Docker container for on-premises inference using the Azure IoT Edge runtime. What should you do first?

A.Create an Azure Container Registry and push the Custom Vision base image.

B.Export the model as a Docker container (e.g., TensorFlow) using the Custom Vision portal.

C.Use the Custom Vision prediction API to call the published endpoint from the edge device.

D.Retrain the model with more images to improve mAP.

AnswerB

IoT Edge requires a containerised model image, so the model must first be exported from Custom Vision as a Docker container for a supported framework such as TensorFlow, producing the artefact later deployed to the edge device.

Why this answer

To deploy a Custom Vision model to an Azure IoT Edge device, you must first export the model as a Docker container (e.g., TensorFlow, ONNX, or DockerFile) from the Custom Vision portal. This export creates a container image that can be deployed to Azure Container Registry and then used as a module in an IoT Edge deployment. Without this export step, you cannot create the containerized module required for on-premises inference.

Exam trap

The trap here is that candidates may think they can directly use the cloud prediction endpoint on an edge device, but Azure IoT Edge requires a containerized module for local execution, making the export step mandatory before any deployment.

How to eliminate wrong answers

Option A is wrong because you do not push the Custom Vision base image; instead, you export the trained model as a container from the portal, which generates a Docker image that you then push to Azure Container Registry. Option C is wrong because calling the prediction API from the edge device would require internet connectivity and defeats the purpose of on-premises inference; IoT Edge runs modules locally without constant cloud access. Option D is wrong because retraining the model to improve mAP is a separate optimization step and does not address the immediate deployment requirement to create a container for IoT Edge.

Full explanation →

298

MCQmedium

You are building a generative AI application with Azure OpenAI Service that summarizes long legal contracts. Legal reviewers report that the summaries omit obligations buried in the middle of very long documents, even though the documents fit within the model's maximum context window. You need to improve recall of those mid-document clauses while minimizing added latency. What should you do first?

A.Add a system message instructing the model to read the entire document carefully before summarizing.

B.Increase the model's temperature setting so the model explores more of the document content.

C.Split the contract into overlapping chunks, summarize each chunk, then summarize the combined chunk summaries.

D.Switch the deployment to a model with a larger maximum context window and resend the full contract.

AnswerC

Chunking with overlap, followed by a map-reduce style summarization, ensures every clause is processed in a short context where attention is strongest. Overlap prevents obligations that straddle chunk boundaries from being lost. This directly improves mid-document recall and keeps each model call small, so latency stays reasonable compared with sending the entire contract repeatedly. It is the standard mitigation for weak recall in long inputs.

Why this answer

Long inputs suffer from uneven attention, so obligations in the middle can be missed even when they fit the context window. Chunking with overlap and summarizing hierarchically places every clause in a short, high-attention context and preserves clauses that cross boundaries. Raising temperature, enlarging the window, or adding instructions does not change how attention is distributed, so those approaches leave the recall gap unresolved.

Exam trap

The trap here is assuming that fitting within the context window means the model reliably uses every part of the input.

Full explanation →

299

MCQhard

You are developing a solution that uses Azure AI Language's custom named entity recognition (NER) to extract product names from technical support tickets. After training a model, you notice it performs well on the training set but poorly on new tickets. You need to improve the model's ability to generalize. What should you do?

A.Add more labeled examples that include variations in phrasing, context, and entity placement.

B.Reduce the training dataset size to prevent the model from memorizing too much information.

C.Increase the number of training epochs to allow the model to learn more from the existing data.

D.Switch to a prebuilt entity extraction model, because custom models cannot generalize.

AnswerA

Overfitting occurs when the model memorizes training data rather than learning general patterns. Adding diverse labeled examples that vary sentence structure, context, and entity position helps the model generalize to unseen tickets. This is the recommended approach to improve performance on new data without changing the model architecture.

Why this answer

Poor generalization on new data indicates overfitting. The best remedy is to expand the training set with varied examples that reflect real-world diversity, helping the model learn robust patterns. Increasing epochs or reducing data would exacerbate overfitting, and prebuilt models cannot extract custom entities like product names.

Exam trap

The trap here is thinking that more training on the same data will improve performance, when it actually worsens overfitting.

Full explanation →

300

MCQeasy

You are deploying an Azure AI solution that uses Azure AI Language and Azure AI Vision. You need to ensure that the solution can be monitored for performance and health. You want to collect metrics and logs from these services and analyze them in a central location. What should you use?

A.Azure Resource Health

B.Azure Service Health

C.Azure Advisor

D.Azure Monitor

AnswerD

Azure Monitor is the centralized platform for collecting, analyzing, and acting on telemetry from Azure resources. It can collect metrics and logs from Azure AI Language and Azure AI Vision, and allows you to create alerts, dashboards, and queries using Log Analytics. This satisfies the requirement for centralized monitoring of performance and health.

Why this answer

Azure Monitor is the correct choice because it is the comprehensive monitoring solution for Azure resources. It can collect metrics and logs from Azure AI Language and Azure AI Vision, store them in Log Analytics, and enable analysis, alerting, and visualization. The other services provide health advisories or recommendations but do not offer centralized telemetry collection and analysis.

Exam trap

The trap here is confusing Azure Monitor with services like Azure Service Health or Azure Advisor, which provide health status and recommendations but do not collect resource-level metrics and logs for performance analysis.

Full explanation →

Microsoft Azure AI Engineer Associate AI-102 (AI-102) — Questions 226–300