Microsoft Azure AI Engineer Associate AI-102 AI-102 Questions 601–675 | Page 9/11

601

MCQeasy

You are deploying a chatbot using Azure AI Bot Service and Language Understanding (LUIS). The bot must understand user intent from free-text input. Which component should you train?

A.Language Understanding (LUIS) model

B.Speech-to-text model

C.QnA Maker knowledge base

D.Computer Vision model

AnswerA

LUIS is the component that maps free-text utterances to intents, so its model must be trained with example utterances and labelled intents. Azure AI Bot Service only orchestrates conversation; it performs no intent classification itself.

Why this answer

The Language Understanding (LUIS) model is the correct component to train because the bot needs to interpret free-text user input and extract intent. LUIS is a natural language processing service specifically designed for intent recognition and entity extraction from conversational phrases. Training the LUIS model with labeled utterances teaches it to map user expressions to predefined intents, enabling the bot to understand and respond appropriately.

Exam trap

The trap here is that candidates may confuse the role of LUIS with QnA Maker, assuming both handle any text input, but LUIS is for intent classification from free-text conversation, while QnA Maker is for retrieving answers from a fixed knowledge base, not for understanding dynamic user intents.

How to eliminate wrong answers

Option B is wrong because a Speech-to-text model converts audio to text, but the question specifies free-text input, not spoken input; training this model would be unnecessary and irrelevant for text-based intent understanding. Option C is wrong because QnA Maker knowledge base is designed for answering factual questions from a structured FAQ or document, not for understanding free-form intents from conversational input; it lacks the intent classification capability required here. Option D is wrong because a Computer Vision model processes images and video, not text; it has no role in interpreting user intent from free-text input.

Full explanation →

602

MCQeasy

A team is prototyping an Azure OpenAI chat application and wants to iterate quickly on prompt wording without changing application code or redeploying. They also want to compare outputs from different prompt variants side by side. Which Azure AI Foundry feature should they use?

A.The Azure OpenAI resource's keys and endpoint blade

B.The model deployment quota page in the Azure portal

C.The prompt flow authoring experience in Azure AI Foundry

D.The content filter configuration page for the deployment

AnswerC

Prompt flow provides a visual authoring surface where prompts, inputs, and model configurations are defined as a flow and can be edited and re-run without changing application code. It supports comparing variants and evaluating outputs, which matches the goal of fast prompt iteration and side-by-side comparison. This is the intended tool for that workflow.

Why this answer

The requirement is to iterate on prompts and compare variants without code changes. Prompt flow in Azure AI Foundry is the authoring and orchestration surface designed for exactly that, letting prompts be edited, run, and evaluated visually. Quota pages, key and endpoint blades, and content filter settings are administrative or safety configuration surfaces and provide no prompt editing or comparison capability.

Exam trap

The trap here is confusing administrative configuration surfaces, such as quota or content filter pages, with the authoring environment where prompts are actually developed and evaluated.

Full explanation →

603

Multi-Selectmedium

You are building a solution that must translate customer chat messages from Spanish to English in real-time. The solution must also detect the language of incoming messages to handle cases where users write in other languages. Which TWO Azure AI service features should you use?

Select 2 answers

A.Azure AI Translator - Language Detection

B.Azure AI Translator - Translation

C.Azure AI Speech - Speech Translation

D.Azure AI Language Understanding (LUIS)

E.Azure AI Custom Question Answering

AnswersA, B

Language Detection identifies the language of each incoming chat message, satisfying the requirement to handle users writing in languages other than Spanish before translation is applied. It returns the detected language code and confidence score, enabling conditional routing.

Why this answer

Azure AI Translator's Language Detection feature (option A) is correct because it identifies the language of incoming chat text, which is exactly what is needed to handle messages written in languages other than Spanish. Azure AI Translator's Translation feature (option B) is correct because it performs the actual text-to-text translation from Spanish to English in real time. Together, these two features satisfy both requirements: detecting the source language and translating the chat messages.

Option C (Azure AI Speech - Speech Translation) is not appropriate because the scenario involves chat text, not spoken audio. Option D (LUIS) is a natural language understanding service for intent and entity extraction, not translation or language identification. Option E (Custom Question Answering) builds FAQ-style knowledge bases and does not provide translation or language detection.

Exam trap

Microsoft Azure exams often test the distinction between text-based and speech-based services, so the trap here is assuming that Speech Translation (Option C) is appropriate for text chat, when it is actually designed for audio input and would require unnecessary speech-to-text conversion.

Full explanation →

604

MCQhard

You are planning to deploy an Azure AI solution that uses an Azure AI Services multi-service resource. The solution must be deployed across multiple Azure regions to provide high availability. You need to ensure that the solution can fail over automatically if one region becomes unavailable. What should you do?

A.Deploy the Azure AI Services resource in one region and configure a custom domain with multiple CNAME records.

B.Deploy the Azure AI Services resource in two regions and configure a Traffic Manager profile with priority routing.

C.Deploy the Azure AI Services resource in two regions and use Azure Front Door with session affinity enabled.

D.Deploy the Azure AI Services resource in one region and enable geo-redundant storage for the resource.

AnswerB

Traffic Manager with priority routing directs all traffic to the primary region and automatically fails over to the secondary region if the primary becomes unavailable. This provides high availability and automatic failover for the Azure AI Services resource across regions, meeting the requirement.

Why this answer

Deploying the resource in two regions and using Traffic Manager with priority routing ensures that traffic is directed to the primary region and automatically redirected to the secondary if the primary fails. This provides the required high availability and automatic failover. Other options either replicate data without failover or rely on DNS without health checks.

Exam trap

The trap here is assuming that data replication or DNS-based load balancing alone provides automatic failover, when health-checked routing such as Traffic Manager priority routing is needed for automatic region failover.

Full explanation →

605

MCQeasy

A developer is building a mobile app that uses Azure Computer Vision to analyze images. The app needs to handle many requests with low latency. Which pricing tier should they choose?

A.Video Analyzer S1 tier.

B.Free F0 tier.

C.Custom Vision S0 tier.

D.Computer Vision S1 tier.

AnswerD

The S1 tier provides higher request throughput and lower latency than the free or F0 tier, which is rate-limited and unsuitable for production mobile traffic. S1 satisfies the constraint of handling many concurrent requests with low latency.

Why this answer

The Computer Vision S1 tier is designed for production workloads requiring high throughput and low latency, making it suitable for a mobile app that handles many image analysis requests. Unlike the Free F0 tier, which has strict rate limits (e.g., 20 transactions per minute), the S1 tier offers higher transactions per second (e.g., up to 10 TPS) and guaranteed performance for real-time scenarios.

Exam trap

The trap here is that candidates may confuse the Computer Vision S1 tier with the Custom Vision S0 tier or the Video Analyzer S1 tier, not realizing that each Azure AI service has its own distinct pricing tiers and that the question specifically targets the Computer Vision service for image analysis.

How to eliminate wrong answers

Option A is wrong because Video Analyzer S1 is a pricing tier for Azure Video Analyzer, a service for video indexing and analysis, not for image analysis with Computer Vision. Option B is wrong because the Free F0 tier has severe rate limits (e.g., 20 calls per minute) and is intended only for evaluation and small-scale testing, not for production apps with many requests. Option C is wrong because Custom Vision S0 is a tier for the Custom Vision service, which is used to train and deploy custom image classification models, not the general-purpose Computer Vision API for analyzing images.

Full explanation →

606

MCQmedium

You are designing an Azure AI solution that uses Language Understanding (LUIS) for intent detection. The solution must handle multiple languages dynamically based on the user's locale. What should you do?

A.Use Azure Translator to translate user input to English before sending to LUIS.

B.Create separate LUIS applications for each language and route based on locale.

C.Train a single LUIS app with utterances in all languages.

D.Enable the 'Multi-Language' feature in the LUIS app.

AnswerB

LUIS applications are language-specific, so a single app cannot serve multiple locales. Creating one app per language and routing by locale satisfies the dynamic multi-language constraint, since each app trains on its own language's utterances.

Why this answer

LUIS does not natively support multi-language within a single application; each LUIS app is designed for a single language. To handle multiple languages dynamically, you must create separate LUIS applications for each language and route user utterances based on the detected locale, ensuring accurate intent and entity recognition per language.

Exam trap

The trap here is that candidates assume LUIS has a built-in multi-language feature or that translation is a viable shortcut, but Microsoft explicitly requires separate LUIS apps per language and does not support multi-language training within a single app.

How to eliminate wrong answers

Option A is wrong because translating user input to English before sending to LUIS introduces translation latency, potential loss of nuance, and inaccuracies in intent detection, as LUIS is optimized for native language patterns. Option C is wrong because training a single LUIS app with utterances in multiple languages degrades performance, as LUIS expects consistent language structure and cannot distinguish between languages during prediction. Option D is wrong because there is no 'Multi-Language' feature in LUIS; the platform requires separate apps for each language, and enabling such a feature would not resolve the fundamental single-language limitation.

Full explanation →

607

Multi-Selecthard

Which TWO of the following are best practices for managing Azure AI services costs?

Select 2 answers

A.Use the S0 pricing tier for production workloads

B.Always use the Free tier to avoid charges

C.Scale up partitions to improve performance

D.Increase batch size to reduce number of API calls

E.Set up budget alerts in Azure Cost Management

AnswersA, E

The S0 standard tier provides the throughput, SLA and feature set required for production workloads, avoiding the rate limits and lack of SLA that constrain the free F0 tier. This satisfies the production workload requirement.

Why this answer

Option A is correct because the S0 (Standard) tier is the production-grade pricing tier for Azure AI services, offering higher throughput, SLA-backed availability, and pay-as-you-go billing that lets you match capacity to actual usage rather than being capped by Free-tier limits. Option E is correct because configuring budget alerts in Azure Cost Management (Microsoft Cost Management + Billing) proactively notifies you when spending approaches or exceeds defined thresholds, enabling early corrective action before costs escalate. Option B is incorrect because the Free tier (F0) has strict transaction and rate limits and is intended only for trials and evaluation, not production workloads.

Option C is incorrect because scaling up partitions increases provisioned capacity and therefore cost, and it is a performance/scalability action rather than a cost-management best practice. Option D is incorrect because increasing batch size is a throughput optimization for supported batch APIs, not a general cost-control practice, and it does not reduce charges for services billed per transaction in the way implied.

Exam trap

The trap here is that candidates often confuse cost-saving strategies (like using the Free tier or batching) with best practices for managing costs in production, overlooking that the Free tier is not for production and that batching may not be applicable or effective for all services.

Full explanation →

608

MCQeasy

You are building a solution that uses Azure AI Language's sentiment analysis to monitor customer feedback. The feedback includes text in multiple languages, and you need to obtain sentiment scores at both the document level and the sentence level. Which API endpoint should you call?

A.POST /language/:analyze-text with kind set to "KeyPhraseExtraction" and include sentiment scores.

B.POST /text/analytics/v3.1/sentiment with the showStats parameter set to true.

C.POST /language/:analyze-conversations with kind set to "SentimentAnalysis".

D.POST /language/:analyze-text with kind set to "SentimentAnalysis" and include opinion mining.

AnswerD

The /language/:analyze-text endpoint is the unified endpoint for Azure AI Language features. Setting kind to SentimentAnalysis performs sentiment analysis. By default, it returns document-level sentiment and sentence-level sentiment when the input contains multiple sentences. Opinion mining can be enabled to extract aspects and opinions. This endpoint meets the requirement for both document and sentence level scores.

Why this answer

The unified Azure AI Language endpoint /language/:analyze-text supports sentiment analysis when the kind parameter is set to SentimentAnalysis. It returns both document-level and sentence-level sentiment, and can optionally include opinion mining. This is the current recommended approach, replacing the older Text Analytics API endpoints.

The other options either use legacy endpoints or specify incorrect task kinds.

Exam trap

The trap here is assuming that the older Text Analytics API endpoint is still the primary way to perform sentiment analysis, when the unified Language endpoint is now preferred.

Full explanation →

609

MCQmedium

Refer to the exhibit. You send this request to the Azure AI Language Service for custom entity recognition. The response returns no entities. What is the most likely reason?

A.The text input is too short for entity recognition

B.The language parameter is set incorrectly

C.The model version is not specified

D.The projectName and deploymentName are missing from the body parameters

AnswerD

Custom entity recognition requests must specify both projectName and deploymentName in the body so the service knows which trained model to invoke. Omitting them means no deployment is resolved, so the response returns an empty entities array.

Why this answer

The Azure AI Language Service for custom entity recognition requires both `projectName` and `deploymentName` in the request body to identify which trained custom model to invoke. Without these parameters, the service cannot route the request to the correct custom model, so it returns no entities. The standard pre-built entity recognition does not require these fields, but custom entity recognition mandates them.

Exam trap

The trap here is that candidates assume the request is valid because it includes text and a language parameter, overlooking that custom entity recognition requires explicit project and deployment identifiers to invoke the trained model.

How to eliminate wrong answers

Option A is wrong because the Azure AI Language Service does not impose a minimum text length for entity recognition; even very short text can return entities if they are present. Option B is wrong because the language parameter, while important for language detection, is not the cause of returning no entities in a custom entity recognition request; the service will still attempt to process the text and return entities from the custom model if properly configured. Option C is wrong because the model version is an optional parameter; if not specified, the service uses the latest available version of the custom model, so its absence does not cause a failure to return entities.

Full explanation →

610

Multi-Selectmedium

Which TWO configurations are required to enable Azure AI Search to index content from an Azure SQL database?

Select 2 answers

A.Create a custom skillset for data enrichment

B.Configure semantic ranking on the index

C.Enable change tracking on the Azure SQL table

D.Define a data source connection to the Azure SQL database

E.Enable high availability on the Azure SQL database

AnswersC, D

Change tracking lets the indexer detect which rows were inserted, updated, or deleted since the last run, so incremental re-indexing stays accurate without full reloads. Without it, the SQL indexer cannot identify changed rows, and the required high-water mark column for incremental indexing is unavailable.

Why this answer

Option C is correct because Azure AI Search's SQL indexer relies on change tracking (or a rowversion/timestamp column) to detect which rows have been inserted, updated, or deleted since the last indexing run, so enabling change tracking on the Azure SQL table is required for incremental indexing. Option D is correct because the indexer must be given a data source object that specifies the connection string, table or view, and change-tracking policy for the Azure SQL database, which is the mandatory link between the search service and the SQL data. Option A is not required because a skillset is only needed for AI enrichment (for example OCR, key phrase extraction, or embedding generation), not for basic SQL-to-index ingestion.

Option B is not required because semantic ranking is an optional query-time feature that improves relevance; it does not affect whether content can be indexed. Option E is not required because high availability is a resilience/uptime configuration for the SQL database and is unrelated to the indexer's ability to read and index data.

Exam trap

The trap here is that candidates often confuse optional enrichment features (like custom skillsets or semantic ranking) with mandatory infrastructure requirements for data ingestion in Azure AI Search, leading them to select those options instead of the core connectivity and change tracking configurations.

Full explanation →

611

MCQmedium

Your organization is using Azure OpenAI Service to generate content. You need to ensure that the content meets safety guidelines by filtering harmful outputs. What should you configure?

A.Enable the Responsible AI dashboard.

B.Configure the content filters in the Azure OpenAI Studio.

C.Use Azure AI Content Safety APIs to analyze outputs.

D.Set the system message to instruct the model to avoid harmful content.

AnswerB

Content filters in Azure OpenAI Studio apply configurable severity thresholds across hate, violence, sexual and self-harm categories, blocking or annotating harmful model outputs. This satisfies the stem's safety requirement by enforcing filtering at the deployment level before responses reach users.

Why this answer

Content filters in Azure OpenAI Studio allow you to define severity levels (safe, low, medium, high) for categories like hate, sexual, violence, and self-harm, which are enforced at the inference API level to block or flag harmful outputs before they reach the user. This is the primary configuration for filtering model-generated content in Azure OpenAI Service.

Exam trap

The trap here is that candidates often confuse the Responsible AI dashboard (a monitoring tool) with active content filtering, or they assume that system messages alone are sufficient for safety, when in fact content filters provide the only guaranteed enforcement layer at the API level.

How to eliminate wrong answers

Option A is wrong because the Responsible AI dashboard is a monitoring and reporting tool that provides visibility into model behavior and fairness metrics, but it does not actively filter or block harmful outputs in real-time. Option C is wrong because Azure AI Content Safety APIs are a separate service for analyzing user-generated or third-party content, not for filtering outputs from Azure OpenAI models directly; they would require an additional integration layer. Option D is wrong because system messages are instructional prompts that guide model behavior but are not a reliable enforcement mechanism—they can be overridden by adversarial inputs or model quirks, and they lack the deterministic filtering capabilities of content filters.

Full explanation →

612

MCQmedium

You need to create a solution that extracts key-value pairs from scanned invoices using Azure AI Document Intelligence. The invoices have varying layouts. Which model should you use?

A.Layout model

B.Custom extraction model

C.Read model

D.Prebuilt invoice model

AnswerD

The prebuilt invoice model returns structured key-value pairs for fields such as invoice date, vendor and total, handling the layout variance that custom templates cannot. It satisfies the stem's requirement to extract key-value pairs from scanned invoices without training, since Microsoft's pretrained model already covers common invoice schemas.

Why this answer

The Prebuilt invoice model (Option D) is specifically designed to extract key-value pairs, line items, and other structured fields from invoices, even when layouts vary. It is trained on thousands of invoice samples and uses deep learning to handle diverse formats without requiring custom training, making it the correct choice for this scenario.

Exam trap

The trap here is that candidates often confuse the Layout model's ability to extract tables and text with the specific key-value pair extraction needed for invoices, leading them to choose Option A instead of the purpose-built Prebuilt invoice model.

How to eliminate wrong answers

Option A is wrong because the Layout model extracts text, tables, and selection marks but does not extract key-value pairs or invoice-specific fields like invoice number or vendor details. Option B is wrong because a Custom extraction model requires labeled training data and is overkill when a prebuilt model already exists for invoices; it is intended for documents not covered by prebuilt models. Option C is wrong because the Read model only extracts printed and handwritten text (OCR) without any structure or key-value pair extraction.

Full explanation →

613

MCQmedium

A retail company uses Azure Computer Vision to analyze customer traffic in stores. They deploy a custom object detection model to count customers and detect occupancy. After deployment, the model consistently underestimates the number of customers during peak hours. The company has retrained the model with more data but the issue persists. What is the most likely cause?

A.The model is not being batch-processed for inference.

B.The training data does not adequately represent peak-hour scenarios.

C.The model is overfitting to the training data.

D.The Computer Vision API version is outdated.

AnswerB

Persistent underestimation despite retraining indicates the training data under-represents peak-hour conditions such as crowding and occlusion, so the model never learns those patterns. The cause is a data representation gap, not model architecture or inference configuration.

Why this answer

The model consistently underestimates customer counts during peak hours, which indicates a distribution shift between the training data and the inference environment. Even after retraining with more data, the issue persists because the additional data likely still lacks sufficient representation of peak-hour scenarios (e.g., high density, occlusion, rapid movement). In Azure Custom Vision, object detection models learn from labeled examples; if the training set does not include diverse peak-hour images with varied lighting, crowd densities, and angles, the model will fail to generalize to those conditions.

Exam trap

The trap here is that candidates may assume retraining with 'more data' automatically fixes the issue, but the key is that the additional data must be representative of the specific failure scenario (peak hours), not just any data.

How to eliminate wrong answers

Option A is wrong because batch processing affects throughput and latency, not the accuracy of individual inference results; the model's underestimation is a precision/recall issue, not a processing mode issue. Option C is wrong because overfitting would cause the model to perform well on training data but poorly on new data in general, not specifically during peak hours; the consistent underestimation only in peak hours points to a data distribution mismatch, not overfitting. Option D is wrong because the Computer Vision API version affects available features and endpoints, not the learned weights of a custom object detection model; the model's behavior is determined by its training data and architecture, not the API version used for deployment.

Full explanation →

614

MCQeasy

You are building a generative AI solution using Azure Machine Learning prompt flow. The solution must allow business analysts without coding experience to modify prompts and evaluate different model versions. What should you do?

A.Provide the analysts with a Jupyter notebook using the OpenAI Python SDK

B.Deploy a chatbot in Microsoft Copilot Studio and let analysts configure it

C.Implement a custom web UI using Azure Static Web Apps and Azure Functions

D.Use Azure Machine Learning prompt flow with the visual designer and variant management

AnswerD

The visual designer gives non-coders a drag-and-drop canvas to edit prompts, while variant management lets them compare model versions side by side. Together these satisfy the requirement that business analysts modify prompts and evaluate model versions without writing code.

Why this answer

Azure Machine Learning prompt flow provides a visual designer that enables non-technical users to modify prompts without coding, and its variant management feature allows them to evaluate different model versions side-by-side. This directly addresses the requirement for business analysts to iteratively refine prompts and compare model outputs in a controlled, no-code environment.

Exam trap

The trap here is that candidates may confuse the no-code visual designer in Azure Machine Learning prompt flow with other low-code tools like Copilot Studio or assume that a custom web UI is simpler, but the question specifically requires a solution that allows prompt modification and model evaluation without coding, which only prompt flow's variant management and visual designer provide.

How to eliminate wrong answers

Option A is wrong because Jupyter notebooks require Python coding skills and familiarity with the OpenAI SDK, which business analysts without coding experience cannot use. Option B is wrong because Microsoft Copilot Studio is designed for building conversational agents with pre-built templates, not for modifying prompts or evaluating different model versions in a generative AI pipeline. Option C is wrong because implementing a custom web UI with Azure Static Web Apps and Azure Functions requires significant development effort and coding, which defeats the purpose of enabling non-technical analysts to modify prompts directly.

Full explanation →

615

MCQeasy

A developer is creating a custom text classification model using Azure AI Language. The dataset has 10,000 documents across 50 categories. Which method is most suitable for labeling the data efficiently?

A.Use a prebuilt model from the Azure AI Language service

B.Use active learning in the custom text classification project

C.Manually label all documents in the Language Studio

D.Use Azure Machine Learning designer to auto-label

AnswerB

Active learning surfaces the unlabelled documents the model is least certain about, so annotators label only the most informative examples. With 10,000 documents across 50 categories, this sharply reduces labelling effort compared with exhaustive manual annotation.

Why this answer

Active learning in custom text classification projects automatically selects the most informative unlabeled documents for manual review, reducing labeling effort while maximizing model accuracy. With 10,000 documents across 50 categories, active learning prioritizes ambiguous or high-uncertainty samples, making it the most efficient approach for iterative labeling.

Exam trap

The trap here is that candidates assume 'prebuilt models' (Option A) can be adapted to custom categories via fine-tuning, but Microsoft Azure AI Language custom text classification requires a dedicated project with active learning—prebuilt models are static and cannot learn new labels.

How to eliminate wrong answers

Option A is wrong because prebuilt models are designed for general-purpose classification (e.g., sentiment, key phrases) and cannot be customized to 50 specific categories; they lack the ability to learn custom labels. Option C is wrong because manually labeling all 10,000 documents is inefficient and time-consuming, especially when active learning can achieve comparable accuracy with far fewer labeled examples. Option D is wrong because Azure Machine Learning designer does not provide auto-labeling for custom text classification; it focuses on automated ML pipelines for structured data, not active learning for text labeling.

Full explanation →

616

MCQhard

You are using Azure AI Language to analyze customer feedback. You need to identify the sentiment of each sentence within a review, not just the overall document sentiment. Which feature should you use?

A.Sentiment analysis with opinion mining enabled.

B.Key phrase extraction.

C.Custom text classification with sentiment labels.

D.Entity linking.

AnswerA

Sentiment analysis with opinion mining provides sentence-level sentiment and also extracts opinions and aspects. It returns sentiment for each sentence, which directly meets the requirement of identifying sentiment per sentence. Opinion mining adds details about what the sentiment refers to, but the sentence-level sentiment is the key output.

Why this answer

Sentiment analysis in Azure AI Language can return sentiment at the document and sentence level when opinion mining is enabled. This provides the granularity needed to see sentiment per sentence. Key phrase extraction, custom classification, and entity linking do not offer built-in sentence-level sentiment.

Exam trap

The trap here is assuming that key phrase extraction or custom classification can provide sentiment, when only the sentiment analysis feature with opinion mining returns sentence-level sentiment.

Full explanation →

617

Multi-Selecthard

Which THREE factors should you consider when designing a knowledge mining solution that uses Azure AI Search and custom skills to extract insights from large volumes of documents?

Select 3 answers

A.The number of knowledge store projections affects indexing speed

B.The maximum execution time of the custom skill must fit within the indexer timeout

C.Incremental enrichment should be enabled to avoid reprocessing unchanged documents

D.Semantic ranking configuration must be included in the skillset

E.The custom skill should be stateless and idempotent to allow parallel execution

AnswersB, C, E

Custom skills execute synchronously inside the indexer's enrichment pipeline, so a skill exceeding the indexer timeout aborts the run and leaves documents partially enriched. Sizing skill execution against that timeout is therefore a core design constraint for large document volumes.

Why this answer

Option B is correct because custom skills run inside the indexer execution pipeline, and each skill invocation must complete within the indexer's timeout limits (for example, the default HTTP timeout of 3 minutes 30 seconds for a WebApiSkill); a long-running custom skill will cause the indexer to fail or time out. Option C is correct because enabling incremental enrichment (by setting the indexer's data source change detection and using a high-water mark) lets Azure AI Search skip documents whose content has not changed, avoiding unnecessary re-invocation of expensive custom skills and reducing cost and processing time. Option E is correct because the indexer can invoke skills concurrently across documents, so a custom skill must not depend on shared mutable state or on a specific call order; making it stateless and idempotent ensures consistent, repeatable enrichment results under parallel execution.

Option A is not a primary design factor here because knowledge store projections are an output concern and do not fundamentally constrain the design of the custom-skill enrichment pipeline in the way timeout, incremental enrichment, and statelessness do. Option D is incorrect because semantic ranking is configured on the search index/query side (semantic configuration), not as a required element of the skillset, so it is not a factor in designing the custom-skill extraction pipeline.

Exam trap

The trap here is that candidates often confuse knowledge store projections (output storage) with indexing performance, or assume semantic ranking is a mandatory skillset component, when in fact it is an optional query-time feature.

Full explanation →

618

MCQhard

You are designing a solution to extract customer names and addresses from scanned handwritten forms. The forms are stored as images in Azure Blob Storage. The extraction must achieve high accuracy with minimal manual review. Which combination of Azure AI services should you use?

A.Azure AI Document Intelligence with prebuilt invoice and receipt models

B.Azure AI Document Intelligence with a custom model trained on handwritten forms

C.Azure AI Language Service with custom Named Entity Recognition (NER)

D.Azure AI Computer Vision with OCR and Azure AI Search

AnswerB

Azure AI Document Intelligence custom models learn your forms' specific layout and handwriting variations from labelled samples, directly satisfying the high-accuracy, minimal-review constraint. Unlike the prebuilt read model, a custom neural model handles the unpredictable field positions and cursive styles typical of scanned handwritten forms stored in Blob Storage.

Why this answer

Azure AI Document Intelligence's custom model capability allows you to train a model specifically on handwritten forms, enabling it to learn the unique handwriting patterns and layout structures present in your scanned documents. This tailored approach achieves high accuracy with minimal manual review, as the model is optimized for your specific form type rather than generic invoice or receipt templates.

Exam trap

The trap here is that candidates often confuse prebuilt models (which work well for printed documents) with custom models (which are necessary for handwritten forms), or they assume OCR alone is sufficient without considering the need for structured field extraction.

How to eliminate wrong answers

Option A is wrong because prebuilt invoice and receipt models are designed for structured, printed documents and cannot reliably extract handwritten text with high accuracy, leading to increased manual review. Option C is wrong because Azure AI Language Service with custom NER extracts entities from text but does not perform OCR or handle image-based handwritten input, so it cannot process scanned forms directly. Option D is wrong because Azure AI Computer Vision with OCR provides raw text extraction but lacks the document understanding and field-level extraction capabilities needed to accurately parse structured fields like customer names and addresses from forms, and Azure AI Search is for indexing and querying, not extraction.

Full explanation →

619

Multi-Selectmedium

You are building an agentic solution using Microsoft Semantic Kernel. The agent uses a planner to orchestrate multiple functions. You want to improve the planner's ability to handle complex user requests that involve multiple steps. Which THREE strategies should you implement?

Select 3 answers

A.Limit the number of available functions to reduce planning overhead

B.Enable the planner to ask the user for clarification when the request is ambiguous

C.Create composite functions that encapsulate common multi-step sub-tasks

D.Use a simple, generic prompt to avoid overfitting

E.Provide few-shot examples of multi-step workflows in the planner prompt

AnswersB, C, E

Clarification prompting lets the planner resolve ambiguous intent before committing to a function sequence, preventing mis-ordered or missing steps in multi-step orchestration. This directly satisfies the stem's goal of handling complex, multi-step requests, since ambiguity is a primary cause of planner failure in Semantic Kernel pipelines.

Why this answer

Option B is correct because allowing the planner to ask the user for clarification when a request is ambiguous prevents it from guessing at missing parameters or intent, which is essential for correctly decomposing complex multi-step requests. Option C is correct because composite functions encapsulate common multi-step sub-tasks into a single callable unit, reducing the planner's reasoning burden and making orchestration of complex workflows more reliable. Option E is correct because few-shot examples of multi-step workflows in the planner prompt demonstrate the expected decomposition and sequencing pattern, improving the planner's ability to generate correct multi-step plans.

Option A is not appropriate because arbitrarily limiting available functions removes capabilities the agent needs for complex requests rather than improving planning quality. Option D is not appropriate because a simple, generic prompt provides no guidance on multi-step decomposition and would degrade, not improve, planning performance for complex tasks.

Exam trap

Microsoft often tests the misconception that reducing function count (Option A) or using simpler prompts (Option D) improves planning, when in fact these strategies limit the planner's expressiveness and ability to handle complex, multi-step requests.

Full explanation →

620

Multi-Selectmedium

Which TWO Azure services can be used to perform optical character recognition (OCR) on images?

Select 2 answers

A.Azure Computer Vision Read API

B.Azure Face API

C.Azure Video Indexer

D.Azure Custom Vision

E.Azure Form Recognizer

AnswersA, E

The Computer Vision Read API is purpose-built for OCR, extracting printed and handwritten text from images and documents via synchronous or asynchronous read operations. This directly satisfies the stem's requirement for an Azure service performing optical character recognition on images.

Why this answer

Azure Computer Vision Read API is correct because it provides a dedicated OCR capability that extracts printed and handwritten text from images and documents. It uses deep learning models to detect text regions, recognize characters, and return structured output with bounding boxes and confidence scores.

Exam trap

Candidates may mistakenly think that only the Computer Vision Read API can perform OCR. However, Azure Form Recognizer also uses OCR technology to extract text from documents, though it is optimized for structured forms and tables. The correct answers are both A and E.

A common mistake is to choose the Face API or Custom Vision, which do not provide OCR capabilities.

Full explanation →

621

Multi-Selectmedium

You are implementing an agent with Azure AI Agent Service that must run a multi-step task: retrieve a customer record, then create a support ticket containing that record. You want the agent to complete both steps in a single run and to be able to report intermediate progress. Which two capabilities should you rely on? (Choose two.)

Select 2 answers

A.File search over the support policy documents, so the agent can justify the ticket priority.

B.Tool calling, so the model can request the retrieval and ticket-creation functions during the run.

C.Run steps, so the application can observe each tool call and its result as the run progresses.

D.Vector embeddings of the ticket schema, so the model can match the customer record to the correct ticket fields.

E.A separate thread per step, so each tool call is isolated from the others.

AnswersB, C

Tool calling is how an agent takes actions in Azure AI Agent Service. The model emits a tool call, the service or your code executes it, and the result returns to the model so it can decide the next step. Without tool calling, the agent could only produce text and could not retrieve the customer record or create the ticket.

Why this answer

Tool calling lets the model request the customer lookup and the ticket creation as part of one run, while run steps give the application visibility into each invocation and result. Together they enable a single multi-step run with observable progress, which is exactly what the scenario requires.

Exam trap

The trap here is confusing retrieval-augmented features such as embeddings or file search with the action-execution and observability features that actually drive a multi-step agent run.

Full explanation →

622

MCQmedium

You are building an Azure AI Search knowledge mining pipeline that enriches scanned PDF invoices. The PDFs are stored in Azure Blob Storage, and you need to extract text from each page before running downstream entity recognition. The solution must minimize development effort and rely on a built-in cognitive skill. Which skill should you add to the skillset?

A.OcrSkill

B.EntityRecognitionSkill

C.ImageAnalysisSkill

D.KeyPhraseExtractionSkill

AnswerA

OcrSkill is the built-in cognitive skill that extracts text from image files and embedded images in PDFs, producing a text output that downstream skills can consume. In this scenario, the scanned invoices contain image-based text, so OCR is required before entity recognition. It minimizes development effort because no custom code or external endpoint is needed.

Why this answer

The pipeline must convert image-based invoice content into text before any language skill can run. OcrSkill is the built-in Azure AI Search cognitive skill designed for that conversion and outputs a text field that downstream skills consume. Image analysis, entity recognition, and key phrase extraction all assume text already exists, so they cannot satisfy the extraction requirement on their own.

Exam trap

The trap here is assuming any cognitive skill that produces text can read scanned images, when only the OCR skill performs image-to-text extraction.

Full explanation →

623

MCQeasy

You need to monitor the costs of your Azure AI services across multiple subscriptions. Which Azure tool should you use to track spending and set budgets?

A.Azure Cost Management

B.Azure Portal

C.Azure Monitor

D.Azure Advisor

AnswerA

Azure Cost Management aggregates spend across subscriptions and resource groups, letting you analyse costs and configure budgets with alerts. This satisfies the requirement to track spending for Azure AI services across multiple subscriptions in one place.

Why this answer

Azure Cost Management is the dedicated tool for monitoring, analyzing, and controlling cloud spending across multiple subscriptions. It provides cost analysis, budget creation, and alerting capabilities specifically designed for tracking Azure AI services costs at scale.

Exam trap

The trap here is that candidates often confuse Azure Monitor (which tracks resource metrics and logs) with cost monitoring, but Azure Monitor has no native capability to track financial spend or set budgets.

How to eliminate wrong answers

Option B is wrong because Azure Portal is the web-based management interface for provisioning and configuring resources, not a dedicated cost tracking and budgeting tool. Option C is wrong because Azure Monitor focuses on performance metrics, logs, and alerts for resource health and application diagnostics, not financial cost tracking. Option D is wrong because Azure Advisor provides best-practice recommendations for optimizing resource usage, security, and reliability, but it does not offer direct cost tracking or budget management features.

Full explanation →

624

MCQmedium

A company uses Azure OpenAI to generate marketing copy. They want to ensure that the generated content does not contain offensive language. Which feature should they enable?

A.Use DALL-E to generate images instead of text.

B.Use a system message instructing the model to avoid offensive language.

C.Enable diagnostic logging to review all outputs.

D.Enable content filtering at the deployment level.

AnswerD

Content filtering at the deployment level applies Azure OpenAI's classification models to both prompts and completions, blocking hate, violence, sexual and self-harm categories. This directly satisfies the requirement that generated marketing copy must not contain offensive language, enforcing the restriction on every request routed through that deployment.

Why this answer

Azure OpenAI provides built-in content filtering at the deployment level that automatically detects and blocks offensive or harmful language in both input prompts and generated outputs. This feature uses Microsoft's Responsible AI models to enforce safety policies without requiring custom code or manual review, making it the most reliable and scalable solution for preventing offensive content in marketing copy.

Exam trap

The trap here is that candidates often assume prompt engineering (system messages) is sufficient for safety, but Azure OpenAI requires explicit content filtering at the deployment level to enforce policies reliably and prevent bypassing via prompt injection.

How to eliminate wrong answers

Option A is wrong because DALL-E is an image generation model, not a text filtering mechanism; switching to images does not address the requirement to prevent offensive language in text outputs. Option B is wrong because a system message is a prompt engineering technique that provides guidance to the model but does not guarantee enforcement; the model may still generate offensive content if the instruction is not followed or if the model is manipulated. Option C is wrong because diagnostic logging only records outputs for review after generation, not preventing offensive content in real-time; it is a monitoring tool, not a content filter.

Full explanation →

625

MCQeasy

You are planning to deploy an Azure AI solution that uses Azure AI Language to analyze text. The solution must be able to process a high volume of requests and provide a service-level agreement (SLA) for availability. You need to choose the appropriate pricing tier. What should you do?

A.Use the Free (F0) tier for each Azure AI Language resource to minimize costs.

B.Use the Standard (S) tier only for development and switch to Free (F0) for production to save costs.

C.Use multiple Free (F0) resources and load-balance requests across them.

D.Use the Standard (S) tier for the Azure AI Language resource.

AnswerD

The Standard tier supports high-volume requests, provides an SLA for availability, and is designed for production workloads. It allows you to scale as needed and ensures that the service meets performance requirements. This is the appropriate choice for a solution that must process many requests reliably.

Why this answer

The Standard tier is designed for production workloads, offering higher throughput limits and an SLA. The Free tier is for evaluation only and lacks an SLA. Therefore, the Standard tier is necessary to meet the high-volume and availability requirements.

Exam trap

The trap here is assuming that multiple Free tier resources can collectively meet production needs, but they do not provide an SLA and have per-resource limits.

Full explanation →

626

Multi-Selectmedium

A media company is building a knowledge mining solution with Azure AI Search. They need to enrich video assets by extracting spoken words from the audio track and then indexing that transcript for search. Which two components must be included in the enrichment pipeline? (Choose two.)

Select 2 answers

A.A custom skill that calls Azure AI Video Indexer or the Speech service to transcribe the audio.

B.A scoring profile that boosts video documents by duration.

C.An indexer configured for Azure Blob Storage with the video files in a container.

D.A synonym map that maps spoken words to their text equivalents.

E.A skillset entry that calls the built-in OCR skill on the video file.

AnswersA, C

Azure AI Search does not include a built-in skill that transcribes video audio, so a custom skill is required to invoke a speech-to-text service such as Azure AI Video Indexer or the Speech service. The custom skill returns the transcript as enriched text that can then be mapped into an index field for search.

Why this answer

To index spoken words from video, the pipeline needs a data source and indexer to pull video assets from Blob Storage, plus a custom skill that calls a speech-to-text service because Azure AI Search has no built-in audio transcription skill. OCR, synonym maps, and scoring profiles operate on images, queries, or ranking respectively and cannot produce a transcript.

Exam trap

The trap here is assuming Azure AI Search has a built-in skill for audio transcription, when video or speech transcription must be handled by a custom skill.

Full explanation →

627

MCQmedium

A company uses Azure AI Language's custom named entity recognition (NER) to extract product names from support tickets. They have trained a model with 500 labeled entities across 200 documents. During evaluation, they notice the model has high precision but low recall for a specific product category. What should they do to improve recall for that category?

A.Add more labeled examples of the product category to the training dataset, ensuring they cover different contexts and sentence structures.

B.Increase the model's confidence threshold for entity extraction to reduce false positives.

C.Use the model's evaluation metrics to identify and remove documents that contain ambiguous entity mentions.

D.Retrain the model with a higher learning rate to help it converge faster on the underrepresented category.

AnswerA

This is correct because low recall indicates the model is missing many instances of the entity. Adding more labeled examples for that category, especially in varied contexts, helps the model learn to recognize it more consistently. This directly addresses the gap in training data for that specific entity type.

Why this answer

Low recall for a specific category means the model is failing to identify many true instances. The most effective solution is to add more labeled examples of that category in diverse contexts, which helps the model learn the patterns. Adjusting thresholds or removing data does not address the underlying data gap.

Exam trap

The trap here is thinking that adjusting the confidence threshold or hyperparameters will fix recall, but the real issue is insufficient training examples for the underrepresented entity category.

Full explanation →

628

MCQeasy

A company needs to extract personally identifiable information (PII) from customer support transcripts stored in Azure Blob Storage. Which Azure AI service should they use?

A.Azure AI Speech

B.Azure AI Language Service

C.Azure AI Translator

D.Azure AI Vision

AnswerB

Azure AI Language provides a prebuilt PII extraction feature that identifies and redacts entities such as names, addresses and phone numbers in text. It processes transcript text directly, satisfying the requirement to extract personally identifiable information without training a custom model.

Why this answer

Azure AI Language Service (formerly Text Analytics) includes a pre-built PII detection feature that can identify, categorize, and redact personally identifiable information from unstructured text. This service is specifically designed for text-based extraction tasks, making it the correct choice for processing customer support transcripts stored in Azure Blob Storage.

Exam trap

In the AI-102 exam, candidates often confuse Azure AI services that process text (Language Service) versus those that process audio (Speech), images (Vision), or translation (Translator), leading them to incorrectly select Azure AI Speech when the question involves text extraction from stored files.

How to eliminate wrong answers

Option A is wrong because Azure AI Speech is focused on converting audio to text (speech-to-text) and text to speech, not on extracting PII from existing text transcripts. Option C is wrong because Azure AI Translator is designed for language translation, not for identifying or redacting PII within text. Option D is wrong because Azure AI Vision handles image and video analysis (e.g., OCR, object detection), not text-based PII extraction from documents or transcripts.

Full explanation →

629

MCQeasy

You are deploying an Azure AI solution that uses Azure Cognitive Services. You need to ensure that the API keys are stored securely and can be rotated automatically. What should you use?

A.Azure Key Vault

B.Azure Storage account with SAS tokens

C.Environment variables in Azure App Service

D.Azure App Configuration

AnswerA

Azure Key Vault securely stores secrets such as API keys and supports automated rotation through integration with Azure services. You can configure Key Vault to manage the lifecycle of keys and rotate them on a schedule or on-demand. This meets the requirement for secure storage and automatic rotation.

Why this answer

Azure Key Vault is the correct choice because it provides secure storage for secrets and supports automated rotation. It integrates with other Azure services and allows you to manage keys centrally. The other options do not offer the same level of security and rotation capabilities.

Exam trap

The trap here is assuming App Configuration can securely store secrets; it is meant for non-sensitive configuration data.

Full explanation →

630

MCQeasy

You are using Azure AI Search to index a set of contracts. You need to extract named entities such as organizations, people, and dates from the contract text and store them as separate fields in the index. Which skill should you add to the skillset?

A.Text Translation skill

B.Language Detection skill

C.Key Phrases skill

D.Entity Recognition skill

AnswerD

The Entity Recognition skill uses Azure AI Language to detect named entities in text and returns them with type and subtype information, such as Organization, Person, and DateTime. This allows each entity category to be mapped to its own index field. It is the correct skill for extracting typed entities from contract text.

Why this answer

Entity Recognition is the built-in Azure AI Search skill that calls Azure AI Language to detect named entities with type and subtype metadata. Its output includes categorized entities such as organizations, people, and dates, which can be projected into separate index fields. The other language skills perform different functions and do not produce typed entity output.

Exam trap

The trap here is confusing Key Phrases, which returns a flat list of salient phrases, with Entity Recognition, which returns categorized entities with type information.

Full explanation →

631

Multi-Selecteasy

Which Azure AI service can be used to analyze sentiment in text data?

Select 1 answer

A.Azure AI Translator

B.Azure AI Language Service

C.Azure AI Vision

D.Azure AI Content Safety

E.Azure AI Speech

AnswersB

Azure AI Language Service provides prebuilt sentiment analysis that returns positive, negative, neutral or mixed labels with confidence scores for supplied text. It is the dedicated text analytics offering, unlike Vision or Speech, which handle images and audio respectively.

Why this answer

Azure AI Language Service (formerly Text Analytics) includes a built-in sentiment analysis feature that evaluates text and returns sentiment labels (positive, negative, neutral, mixed) along with confidence scores at the sentence and document level. This makes it the primary service for analyzing sentiment in text data. Azure AI Content Safety is not designed for sentiment analysis; it is intended for harmful content moderation.

Exam trap

A common mistake is to think that Azure AI Content Safety can perform sentiment analysis, but it is actually designed to detect harmful content such as hate speech, threats, and self-harm, not to gauge sentiment. Only Azure AI Language Service provides native sentiment analysis capabilities.

Full explanation →

632

MCQeasy

A developer is creating a custom question answering project in Azure AI Language. The knowledge base contains product manuals in PDF format. Which step is essential before importing the PDFs?

A.Ensure PDFs are in a supported format and accessible

B.Create an Azure AI Search index

C.Deploy a QnA Maker service

D.Translate PDFs to English

AnswerA

Azure AI Language's question answering ingestion accepts PDFs only when they are text-based and within size and page limits; scanned image-only files must be converted first. Ensuring supported formatting and accessible storage satisfies the prerequisite for successful import, since unsupported or inaccessible documents fail extraction before any knowledge base training can begin.

Why this answer

Before importing PDFs into a custom question answering project in Azure AI Language, the essential step is to ensure the PDFs are in a supported format (e.g., searchable PDF, not scanned images without OCR) and accessible via a valid URL or local path. This is because the import process relies on the service being able to read and extract text from the documents; unsupported or inaccessible files will cause the import to fail.

Exam trap

The trap here is that candidates might assume creating an Azure AI Search index is required because custom question answering uses search under the hood, but the Azure AI Language service manages its own index automatically, making Option B a distractor that tests knowledge of Azure AI Language service boundaries.

How to eliminate wrong answers

Option B is wrong because creating an Azure AI Search index is not a prerequisite for importing PDFs into a custom question answering project; the project uses its own built-in indexing and storage, not an external Azure AI Search index. Option C is wrong because QnA Maker is a deprecated service; the current solution is custom question answering within Azure AI Language, which does not require deploying a separate QnA Maker service. Option D is wrong because translation to English is not mandatory; Azure AI Language supports multiple languages for question answering, and PDFs can be imported in their original language as long as the project's language setting matches.

Full explanation →

633

MCQeasy

You are developing a mobile app that allows users to take a photo of a product and get information about it. The app must identify the product from the image. Which Azure AI service should you use?

A.Azure AI Vision OCR

B.Azure AI Face API

C.Azure AI Custom Vision with image classification

D.Azure AI Custom Vision with object detection

AnswerC

Custom Vision image classification trains a model on your own labelled product images, returning the predicted product class for a photo. This satisfies identifying a specific product, which a pre-trained general service cannot do without custom training.

Why this answer

Azure AI Custom Vision with image classification is specifically designed to identify and categorize products or objects within an image based on trained labels. This service allows you to upload images of products, train a model to recognize them, and then use the model to classify new product photos, making it ideal for a product identification app.

Exam trap

The trap here is that candidates often confuse image classification with object detection, thinking that identifying a product requires bounding boxes, when in fact classification alone suffices for determining the product type without needing its location in the image.

How to eliminate wrong answers

Option A is wrong because Azure AI Vision OCR (Optical Character Recognition) extracts text from images, not product identification; it cannot recognize or classify objects like a specific product. Option B is wrong because Azure AI Face API is specialized for detecting, analyzing, and recognizing human faces, not general products or objects. Option D is wrong because Azure AI Custom Vision with object detection identifies and locates multiple objects within an image by drawing bounding boxes around them, which is overkill for simply identifying a single product; image classification is more appropriate for determining what the product is without needing spatial coordinates.

Full explanation →

634

MCQmedium

A bank is deploying an Azure AI Foundry agent that provides account balance information over the phone. The agent must authenticate callers by using voice biometrics before revealing any account details. The bank wants to integrate this authentication step into the agent's conversation flow without exposing sensitive data to the language model. Which approach should the bank use?

A.Include the caller's voiceprint in the system message so the model can compare it to the live audio.

B.Use Azure Communication Services to record the call and send the audio to the agent as a text transcript for authentication.

C.Add a voice biometrics tool to the agent and configure a pre-action that runs before any account-related tool call.

D.Configure the agent to ask for the caller's account number and PIN, then validate them against a database before answering.

AnswerC

Azure AI Foundry agents support tools and pre-actions that can enforce authentication before executing sensitive operations. A voice biometrics tool can verify the caller, and a pre-action ensures the verification runs before account tools are invoked. This keeps sensitive data out of the language model because the authentication result gates access to account information.

Why this answer

The bank needs voice biometric authentication integrated into the agent flow without exposing sensitive data to the language model. A voice biometrics tool with a pre-action that gates account-related tool calls achieves this: the biometric check happens outside the model, and only a success token is passed to the orchestration layer, allowing account tools to run. This satisfies security and integration requirements.

Exam trap

The trap here is assuming the language model itself can perform biometric verification or that any authentication method, such as PIN, is acceptable when voice biometrics is explicitly required.

Full explanation →

635

MCQeasy

You are developing a generative AI application that uses Azure OpenAI Service. The application must generate responses that are grounded in a specific set of documents stored in Azure Blob Storage. You want to use the simplest approach that allows the model to reference these documents without building a custom retrieval pipeline. What should you use?

A.Fine-tune the model with the documents and deploy the fine-tuned model.

B.Implement a custom RAG pipeline using Azure Cognitive Search and the Chat Completions API.

C.Azure OpenAI On Your Data with Azure Blob Storage as the data source.

D.Use the Completions API with a prompt that includes the full text of all documents.

AnswerC

Azure OpenAI On Your Data natively supports Azure Blob Storage as a data source. It handles indexing, retrieval, and citation generation automatically, requiring minimal custom code. This is the simplest way to ground responses in documents stored in Blob Storage without building a retrieval pipeline.

Why this answer

Azure OpenAI On Your Data is a managed feature that integrates with Azure Blob Storage, enabling the model to retrieve and cite documents without custom code. It handles indexing and retrieval automatically, making it the simplest solution for grounding responses in a specific document set. Other options require more effort or are not designed for this purpose.

Exam trap

The trap here is assuming that fine-tuning or manual prompt stuffing can achieve grounding, when a managed retrieval service is specifically designed for this.

Full explanation →

636

MCQhard

You deploy a Custom Vision object detection model to classify vehicles. The model works well in good lighting but fails in low-light conditions. What is the most appropriate action?

A.Add images with different lighting conditions to the training set

B.Increase the probability threshold

C.Increase the number of training iterations

D.Use a domain-specific model for vehicles

AnswerA

Low-light failure is a data coverage gap, not a model or deployment fault. Custom Vision learns only from labelled examples, so adding images captured under varied lighting lets the detector learn illumination-invariant features. This directly satisfies the stem's constraint of poor performance in low-light conditions.

Why this answer

The core issue is a data distribution mismatch: the model was trained primarily on well-lit images and lacks exposure to low-light examples. Adding images with diverse lighting conditions directly addresses this by enriching the training dataset, enabling the model to learn robust features for low-light scenarios. This aligns with the fundamental principle that Custom Vision models are only as good as the training data they receive.

Exam trap

The trap here is that candidates often confuse model performance tuning (threshold, iterations) with data quality issues, mistakenly believing that adjusting hyperparameters can compensate for missing training scenarios.

How to eliminate wrong answers

Option B is wrong because increasing the probability threshold only adjusts the confidence level required to return a prediction; it does not improve the model's ability to detect objects in low light, and may actually reduce recall by filtering out correct but lower-confidence detections. Option C is wrong because increasing the number of training iterations (epochs) on the same dataset does not introduce new visual patterns; it risks overfitting to the existing well-lit images without addressing the low-light deficiency. Option D is wrong because domain-specific models in Custom Vision are pre-trained on generic vehicle images and do not inherently compensate for lighting variations; the problem is not the domain but the lack of representative lighting conditions in the training set.

Full explanation →

637

MCQmedium

You are a security engineer for a financial services company. The company uses Azure AI Language to analyze customer communications for compliance. The solution processes sensitive personal data. You need to ensure that all data transmitted to the Azure AI Language service is encrypted in transit and that the service endpoint is not accessible from the public internet. Additionally, you must use Microsoft Entra ID for authentication. The current implementation uses API keys and the public endpoint. You need to reconfigure the solution. What should you do?

A.Configure a private endpoint and continue using the public endpoint for redundancy

B.Disable the public network access without configuring a private endpoint

C.Enable Microsoft Entra ID authentication but keep the public endpoint and API keys

D.Disable the public network access, configure a private endpoint, enable managed identity, and enforce HTTPS

AnswerD

Disabling public network access with a private endpoint removes the service from the public internet, satisfying the private connectivity constraint. Enabling managed identity replaces API keys with Microsoft Entra ID authentication, while enforced HTTPS guarantees encryption in transit for sensitive personal data.

Why this answer

It addresses all three requirements: disabling public network access removes internet exposure, configuring a private endpoint ensures traffic stays within the Azure backbone and your virtual network, enabling managed identity allows Microsoft Entra ID authentication without API keys, and enforcing HTTPS guarantees encryption in transit via TLS. This combination fully secures the Azure AI Language service for sensitive personal data.

Exam trap

The trap here is that candidates may think disabling public network access alone is sufficient (Option B), but without a private endpoint, the service becomes unreachable, and they may overlook that managed identity is required to replace API keys for Microsoft Entra ID authentication.

How to eliminate wrong answers

Option A is wrong because continuing to use the public endpoint for redundancy still exposes the service to the public internet, violating the requirement that the endpoint not be accessible from the public internet. Option B is wrong because disabling public network access without a private endpoint leaves no way to connect to the service, as the service would be unreachable. Option C is wrong because keeping the public endpoint and API keys fails to restrict public internet access and does not eliminate the use of API keys, contradicting the requirement to use Microsoft Entra ID authentication exclusively.

Full explanation →

638

Drag & Dropmedium

Drag and drop the steps to implement an Azure AI Bot Service with QnA Maker into the correct order.

Drag or tap steps into the slots.

Steps

Order

1Step 1

2Step 2

3Step 3

4Step 4

Why this order

Start with QnA Maker, build the knowledge base, create the bot, connect it, and test.

Full explanation →

639

Multi-Selectmedium

You are building a conversational language understanding (CLU) project in Azure AI Language. You need to ensure the model can correctly interpret user utterances that include both an intent and multiple entities. Which two actions should you perform? (Choose two.)

Select 2 answers

A.Add a prebuilt entity component for each custom entity to supplement training data.

B.Add utterances that contain multiple entities and label each entity with the correct entity type.

C.Train the model and review the evaluation metrics for entity precision and recall.

D.Enable the "Extract multiple entities" option in the project settings.

E.Define a separate intent for each entity type to improve extraction accuracy.

AnswersB, C

Labeling utterances with multiple entities teaches the model to recognize and extract each entity type in context. This is essential for multi-entity extraction because the model learns from examples where entities co-occur. Without such examples, the model may miss entities or confuse types, so this action directly supports the requirement.

Why this answer

To handle utterances with multiple entities, you need labeled examples that include multiple entities with correct types, and you must train and evaluate the model to verify performance. Creating intents per entity, looking for a non-existent setting, or using prebuilt entities for custom types will not achieve the goal.

Exam trap

The trap here is assuming there is a project setting to enable multiple entity extraction or that intents should map to entity types, when the real work is in labeled data and evaluation.

Full explanation →

640

MCQmedium

You are building a chatbot that must understand user intents from free-text input. You have a small set of labeled examples. Which Azure AI Language feature should you use to classify intents with minimal effort?

A.Entity Linking

B.Custom Text Classification

C.Language Detection

D.Conversational Language Understanding (CLU)

AnswerD

Conversational Language Understanding trains intent classification and entity extraction from labelled utterances, so a small set of labelled examples is sufficient to build a working model. This directly meets the requirement to classify intents from free-text input with minimal effort.

Why this answer

Conversational Language Understanding (CLU) is the correct choice because it is specifically designed to extract intents and entities from free-text input in a conversational context, and it can be trained with a small set of labeled examples to classify user intents with minimal effort. CLU provides a pre-built pipeline for intent recognition and entity extraction, making it the most efficient option for building a chatbot that understands user intents.

Exam trap

The trap here is that candidates often confuse Custom Text Classification (option B) with intent classification, but CLU is the dedicated Azure AI Language feature for conversational intent recognition, while Custom Text Classification is better suited for static document categorization without dialog context.

How to eliminate wrong answers

Option A is wrong because Entity Linking is used to identify and disambiguate named entities by linking them to a knowledge base (e.g., Wikipedia), not for classifying intents from free-text input. Option B is wrong because Custom Text Classification is designed for categorizing whole documents or sentences into predefined classes, but it does not natively handle conversational context or extract intents and entities in a dialog flow, requiring more custom effort for chatbot scenarios. Option C is wrong because Language Detection identifies the language of the input text, not the user's intent, and is irrelevant to intent classification.

Full explanation →

641

MCQhard

A company uses Azure AI Search to index documents from an Azure SQL Database. They need to ensure that deleted rows in the database are also removed from the search index during incremental indexing. They have configured the data source with change detection policies. What should they do to enable deletion detection?

A.Configure the indexer to use the high water mark change detection policy and set the deletion detection policy to 'none'.

B.Create a SQL trigger that calls the Azure AI Search REST API to delete the document when a row is deleted.

C.Use a view that filters out deleted rows and configure the indexer to use that view as the data source.

D.Add a soft delete column to the table that indicates when a row is deleted, and configure the data source with a soft deletion policy that references that column.

AnswerD

Azure AI Search supports soft delete detection for Azure SQL Database. You must add a column (e.g., IsDeleted) to the table and set its value to true or a timestamp when a row is deleted. Then, in the data source definition, configure a soft deletion policy that specifies this column. The indexer will then remove corresponding documents from the index during incremental runs.

Why this answer

To enable deletion detection for Azure SQL Database in Azure AI Search, you must implement a soft delete column in the table and configure the data source with a soft deletion policy that references that column. The indexer then uses this column to identify and remove deleted documents from the index during incremental indexing.

Exam trap

The trap here is assuming that change detection policies automatically handle deletions, when in fact a separate soft delete policy is required.

Full explanation →

642

MCQmedium

You are using Azure OpenAI Service to generate code snippets for a development team. You notice that the generated code sometimes contains security vulnerabilities. You need to minimize the risk of generating insecure code while maintaining productivity. What should you do?

A.Use system messages to instruct the model to prioritize security

B.Fine-tune the model on a dataset of secure code

C.Set the temperature parameter to 0

D.Disable content filtering to allow more flexibility

AnswerA

System messages set persistent behavioural instructions applied to every request, so embedding a security-first directive steers the model away from vulnerable patterns such as unsanitised input handling, satisfying the requirement to reduce insecure output without adding review overhead.

Why this answer

System messages in Azure OpenAI Service allow you to set the context and behavior of the model, including instructing it to prioritize security when generating code. This approach directly influences the model's output without requiring retraining or sacrificing flexibility, making it the most effective way to reduce security vulnerabilities while maintaining productivity.

Exam trap

The trap here is that candidates may overestimate the effectiveness of fine-tuning (Option B) for security, not realizing that system messages are a simpler, more practical first-line defense in Azure OpenAI Service, while fine-tuning is better suited for domain-specific style or knowledge rather than real-time safety constraints.

How to eliminate wrong answers

Option B is wrong because fine-tuning requires a curated dataset of secure code and significant computational resources, which is time-consuming and may not generalize well to all scenarios; it also reduces the model's flexibility for other tasks. Option C is wrong because setting the temperature parameter to 0 makes the model deterministic and less creative, which can hinder code generation quality and does not inherently address security vulnerabilities. Option D is wrong because disabling content filtering removes safety guardrails that help block harmful or insecure outputs, increasing the risk of generating vulnerable code rather than reducing it.

Full explanation →

643

MCQmedium

You are deploying an Azure AI solution that uses Azure OpenAI Service and Azure AI Language. The solution must ensure that each service has its own managed identity and that access to keys is restricted. You need to configure authentication for the services. What should you do?

A.Configure the services to use API keys stored in Azure App Configuration and enable Azure AD authentication for the application.

B.Use Azure AD service principals with client secrets for each service and store the secrets in Azure Key Vault.

C.Store the service keys in Azure Key Vault and configure the application to retrieve them at runtime using the Azure SDK.

D.Enable managed identities for each Azure AI service and grant them access to Azure Key Vault.

AnswerD

Each Azure AI service can have a system-assigned or user-assigned managed identity. By enabling managed identities and granting them access to Key Vault, you avoid storing credentials in code and can securely retrieve secrets. This aligns with least privilege and Azure best practices for authentication.

Why this answer

Managed identities provide an identity for the service in Azure AD, allowing it to authenticate to resources like Key Vault without storing credentials. Granting each service's managed identity access to Key Vault ensures secure retrieval of secrets and aligns with the principle of least privilege. Other options either rely on shared secrets or do not provide per-service identities.

Exam trap

The trap here is confusing managed identities with service principals or assuming that storing keys in Key Vault alone satisfies the requirement for per-service identities.

Full explanation →

644

MCQhard

Your Azure AI Agent Service agent occasionally calls a 'createOrder' tool with plausible but incorrect item codes. You must ensure that the order is created only when the item code exists in the product database, and that invalid calls are rejected before any order is written. Which approach should you take?

A.Validate the item code inside the function tool implementation and return an error result to the agent when the code is unknown.

B.Add the full product database to the agent's instructions so the model can check codes itself before calling the tool.

C.Enable content filtering on the model deployment and add a blocklist entry for malformed item codes.

D.Increase the model temperature to zero and add a system message telling the model to verify item codes before calling the tool.

AnswerA

Executing validation in the tool implementation puts a deterministic check between the model's proposal and the side effect. If the code is unknown, the tool refuses to create the order and returns an error the model can reason about, so no invalid order is written regardless of what the model suggested. This enforces the constraint where it cannot be bypassed.

Why this answer

Validation belongs in the tool implementation, where it runs deterministically before any write occurs. The model may still propose a bad code, but the tool rejects it and returns an error the agent can act on, guaranteeing that only codes present in the product database result in orders.

Exam trap

The trap here is believing that prompt instructions or lower temperature can enforce a business rule, when only code executing at the side-effect boundary can guarantee it.

Full explanation →

645

MCQhard

You are using Azure AI Foundry to fine-tune a GPT-3.5 model on a dataset of customer service conversations. The fine-tuning job fails with an error indicating that the training data format is invalid. What is the most likely issue?

A.The training data is not in JSONL format with the correct structure.

B.The training data is in CSV format instead of JSON.

C.The training data contains only one conversation example.

D.The training data does not include the assistant's responses.

AnswerA

Fine-tuning requires training data as JSONL, with each line holding a valid chat completion object containing messages with role and content pairs. Any deviation, such as plain JSON arrays or CSV, triggers the invalid format error.

Why this answer

Azure AI Foundry requires fine-tuning data to be in JSONL format with a specific structure: each line must be a JSON object containing a 'messages' array with 'role' and 'content' fields for system, user, and assistant turns. The error indicates the training data format is invalid, and the most likely cause is that the data is not in this required JSONL structure, as JSONL is the only accepted format for GPT-3.5 fine-tuning in Azure OpenAI Service.

Exam trap

The trap here is that candidates confuse the general requirement for 'JSON format' with the specific requirement for 'JSONL format with a messages array,' leading them to incorrectly select CSV or plain JSON as the issue, when the real problem is the lack of the correct conversational structure.

How to eliminate wrong answers

Option B is wrong because CSV format is not supported for fine-tuning GPT-3.5 models in Azure AI Foundry; the service requires JSONL, not JSON or CSV, and CSV lacks the nested 'messages' structure needed for conversational data. Option C is wrong because having only one conversation example does not cause a format error; it may lead to poor model performance but the format itself would still be valid if structured correctly. Option D is wrong because while missing assistant responses would make the data unusable for training, the error specifically indicates a format issue, not a content issue; the JSONL structure could still be technically valid without assistant responses.

Full explanation →

646

MCQhard

You are building a generative AI solution using Azure OpenAI Service. The application must retrieve information from a large private knowledge base. You need to ensure the model uses only relevant documents from the knowledge base to generate answers. Which feature should you configure?

A.Implement a custom prompt flow

B.Use Azure OpenAI On Your Data with vector search

C.Configure a content filter

D.Fine-tune the model with the knowledge base

AnswerB

Azure OpenAI On Your Data with vector search indexes the private knowledge base and retrieves semantically relevant chunks, grounding responses in those documents rather than the model's parametric knowledge. This constrains generation to the supplied content.

Why this answer

B is correct because Azure OpenAI On Your Data with vector search enables the model to retrieve only the most semantically relevant documents from a private knowledge base by converting both the user query and the documents into high-dimensional vectors and performing similarity search. This ensures the model's responses are grounded in the specific, relevant information without exposing the entire knowledge base to the model.

Exam trap

The trap here is that candidates often confuse fine-tuning (D) with retrieval-augmented generation (RAG), assuming that training the model on the knowledge base is the best way to ground answers, when in fact RAG with vector search is the correct pattern for dynamic, relevant document retrieval without modifying the base model.

How to eliminate wrong answers

Option A is wrong because implementing a custom prompt flow does not inherently include a retrieval mechanism; it only orchestrates the sequence of calls and prompts, so it cannot ensure that only relevant documents are used from the knowledge base. Option C is wrong because configuring a content filter is a safety mechanism to block harmful or inappropriate content, not a retrieval or grounding feature to select relevant documents. Option D is wrong because fine-tuning the model with the knowledge base would bake the entire knowledge into the model's weights, which is inefficient, costly, and does not allow dynamic retrieval of only relevant documents per query; it also risks overfitting and cannot handle updates to the knowledge base without retraining.

Full explanation →

647

MCQeasy

A financial services company uses Azure AI Language to analyze customer support transcripts. They want to identify the main topics discussed in each conversation and generate a summary of the key points. The solution must minimize development effort and use prebuilt functionality. You need to recommend the appropriate Azure AI Language features. What should you use?

A.Custom Named Entity Recognition (NER) and conversation summarization.

B.Key phrase extraction and conversation summarization.

C.Entity linking and conversation summarization.

D.Sentiment analysis and key phrase extraction.

AnswerB

Key phrase extraction surfaces the salient terms per transcript, while conversation summarization condenses the key points. Both are prebuilt Azure AI Language capabilities, satisfying the stem's requirement to minimise development effort rather than train custom models.

Why this answer

Key phrase extraction identifies the main topics in text, and conversation summarization produces a summary of key points from a conversation. Both are prebuilt Azure AI Language features requiring no custom training, minimizing development effort. Together they meet the requirement to identify topics and summarize transcripts.

Exam trap

AI-102 often tests whether candidates know which Language features are prebuilt versus custom, so they pick Custom NER or entity linking when the scenario explicitly asks for prebuilt functionality with minimal effort.

How to eliminate wrong answers

Option A is wrong because Custom NER requires training a custom model, which adds development effort and is not needed for topic identification. Option C is wrong because entity linking resolves entities to a knowledge base but does not identify main topics or summarize. Option D is wrong because sentiment analysis and key phrase extraction do not provide conversation summarization, so the summary requirement is unmet.

Full explanation →

648

MCQhard

Your team is using Azure AI Search to index a large collection of technical manuals. Users report that searches for 'disk failure' do not return relevant results because the manuals use terms like 'hard drive crash'. Which feature should you implement to improve recall?

A.Apply a filter

B.Configure a scoring profile

C.Enable semantic search

D.Add a synonym map to the index

AnswerD

A synonym map expands queries so 'disk failure' also matches 'hard drive crash', raising recall for terminology mismatches. It applies at query time against the index, directly addressing the vocabulary gap between user phrasing and manual wording.

Why this answer

A synonym map in Azure AI Search allows you to define equivalent terms (e.g., 'disk failure' = 'hard drive crash') so that queries automatically expand to include synonyms. This directly addresses the vocabulary mismatch between user queries and indexed content, improving recall without requiring changes to the documents or queries.

Exam trap

The trap here is that candidates often confuse semantic search (which improves ranking via language models) with synonym expansion (which directly addresses vocabulary mismatch by broadening the query), leading them to choose option C instead of D.

How to eliminate wrong answers

Option A is wrong because a filter narrows results based on structured field criteria (e.g., date range, category) and does not expand query terms to match synonyms. Option B is wrong because a scoring profile boosts relevance ranking based on fields or functions (e.g., freshness, magnitude) but does not alter which documents match the query. Option C is wrong because semantic search re-ranks results using language understanding to improve relevance, but it does not expand the query to include synonymous terms; it still relies on the original query tokens for matching.

Full explanation →

649

MCQeasy

A company wants to moderate user-generated images for adult content. Which Azure AI Vision feature should they use?

A.Custom Vision with a custom adult classifier

B.Face API

C.Analyze Image API with moderation categories

D.OCR

AnswerC

The Analyze Image API returns adult and racy classification flags directly from its moderation categories, satisfying the requirement to detect adult content in user-generated images. Unlike standalone classifiers, it performs this assessment within a single vision call, so no separate moderation service or custom model is needed.

Why this answer

The Analyze Image API in Azure AI Vision includes built-in moderation categories for detecting adult, racy, and gory content in images. This feature is specifically designed for content moderation without requiring custom training, making it the correct choice for moderating user-generated images for adult content.

Exam trap

The trap here is that candidates may assume Custom Vision is needed for any custom moderation task, but Azure AI Vision's Analyze Image API already includes built-in adult content detection, making custom training unnecessary for this specific use case.

How to eliminate wrong answers

Option A is wrong because Custom Vision requires training a custom classifier with labeled data, which is unnecessary when Azure AI Vision already provides pre-built adult content moderation categories. Option B is wrong because Face API is designed for face detection, recognition, and analysis, not for general adult content moderation. Option D is wrong because OCR (Optical Character Recognition) extracts text from images and does not analyze visual content for adult themes.

Full explanation →

650

Multi-Selectmedium

Which TWO Azure AI services can be used to perform optical character recognition (OCR) on images? (Choose two.)

Select 2 answers

A.Azure AI Document Intelligence Read model

B.Azure Video Indexer

C.Azure AI Custom Vision

D.Azure AI Face API

E.Azure AI Vision OCR (Read API)

AnswersA, E

Azure AI Document Intelligence's Read model extracts printed and handwritten text from images and documents via OCR, returning lines, words and page layout. It satisfies the stem's requirement to perform optical character recognition on images, operating on image files directly rather than requiring search-index ingestion or translation pipelines.

Why this answer

Azure AI Document Intelligence Read model (option A) is correct because its Read model is specifically designed to extract printed and handwritten text from documents and images, returning lines, words, and page layout via the Analyze Document operation. Azure AI Vision OCR (Read API) (option E) is also correct because the Read API in Azure AI Vision performs OCR on images and PDFs, returning extracted text and bounding boxes asynchronously. Azure Video Indexer (option B) focuses on video/audio insights such as transcription and face tracking, not general image OCR.

Azure AI Custom Vision (option C) is an image classification and object detection service, not an OCR text extractor. Azure AI Face API (option D) detects and analyzes faces (landmarks, attributes), and does not perform optical character recognition.

Exam trap

The trap here is that candidates may confuse Azure AI Custom Vision (option C) with OCR capabilities, assuming it can read text from images, when it is actually limited to classifying and detecting objects based on custom training data.

Full explanation →

651

MCQhard

You are implementing an agentic solution using Azure AI Agent Service with multiple agents that need to collaborate. Each agent has access to different knowledge bases. You want to ensure that the agents can share context and hand off tasks to each other seamlessly. Which architecture should you use?

A.Create a single monolithic agent that includes all knowledge bases

B.Deploy each agent independently and configure them to call each other via HTTP

C.Use a supervisor agent that delegates to specialized agents, with a shared context store in Azure Cosmos DB

D.Chain the agents sequentially, passing output from one to the next

AnswerC

A supervisor agent orchestrates delegation to specialised agents while a shared Cosmos DB context store lets each agent read and write common state, enabling seamless handoffs. Point-to-point messaging between agents would fragment context and complicate routing.

Why this answer

The supervisor agent pattern with a shared context store (e.g., Azure Cosmos DB) enables multiple agents to maintain a consistent conversation state and hand off tasks seamlessly. The supervisor orchestrates specialized agents, each with its own knowledge base, while the shared store ensures context is preserved across agent boundaries, which is essential for collaborative agentic workflows in Azure AI Agent Service.

Exam trap

The trap here is that candidates often assume sequential chaining (Option D) is sufficient for handoffs, but they overlook the need for a shared context store to maintain state across agent boundaries, which is a core requirement for seamless collaboration in agentic solutions.

How to eliminate wrong answers

Option A is wrong because a single monolithic agent that includes all knowledge bases violates the principle of separation of concerns and does not allow specialized agents to collaborate or share context dynamically; it also creates a single point of failure and scalability bottleneck. Option B is wrong because deploying each agent independently and configuring them to call each other via HTTP introduces tight coupling, latency, and no built-in mechanism for shared context or state management, leading to inconsistent handoffs. Option D is wrong because chaining agents sequentially passes output from one to the next without a shared context store, which prevents agents from accessing the full conversation history or collaborating in a non-linear fashion, breaking seamless handoff.

Full explanation →

652

MCQeasy

A support team wants to build a bot that answers employee questions by using a set of internal HR policy documents. The team does not want to author question-and-answer pairs manually and needs the bot to return the most relevant passage from the documents, with the source cited. The documents are in English and are updated frequently in a blob container. Which Azure AI Language feature should the team use?

A.Named Entity Recognition with a custom model trained on HR policies

B.Conversational language understanding with an intent for each HR topic

C.Custom text classification with a single-label project

D.Custom question answering with a project that imports the documents and enables the option to extract answers from the source content

AnswerD

Custom question answering can ingest documents, automatically generate question-and-answer pairs, and extract answers from the source content. It returns the best passage with a source citation, which matches the requirement to avoid manual authoring and to cite the document. It also supports refreshing the knowledge base when documents change.

Why this answer

Custom question answering supports importing documents and automatically generating question-and-answer pairs, plus extracting answers directly from source content. It returns the best matching passage with a citation, and it can refresh from a blob container as documents change, meeting the no-manual-authoring and source-citation requirements.

Exam trap

The trap here is confusing question answering with intent classification, where only question answering stores documents and returns cited passages.

Full explanation →

653

MCQhard

A healthcare company is using Azure AI Document Intelligence to extract patient data from forms. They need to ensure that all extracted data is encrypted at rest using a customer-managed key (CMK) and that the service endpoint is restricted to a specific virtual network. Which combination of steps should they take?

A.Use a service endpoint and configure a managed identity

B.Disable public network access and enable CMK via Azure Key Vault

C.Configure IP firewall rules and enable CMK via Azure Key Vault

D.Create a private endpoint and associate a customer-managed key in the resource encryption settings

AnswerD

A private endpoint restricts the Document Intelligence endpoint to a specific virtual network, while associating a customer-managed key in the resource's encryption settings enforces CMK encryption at rest. Together these satisfy both the network isolation and key management constraints in the stem.

Why this answer

It combines a private endpoint (which restricts the service endpoint to a specific virtual network by providing a private IP address within that VNet, eliminating public internet exposure) with a customer-managed key (CMK) in the resource encryption settings, which ensures data at rest is encrypted using a key stored in Azure Key Vault that the customer controls. This directly meets both requirements: network isolation via private endpoint and CMK-based encryption at rest.

Exam trap

The trap here is that candidates often confuse 'service endpoint' or 'IP firewall rules' with 'private endpoint' for VNet-specific access, but only a private endpoint provides a fully private IP within the VNet and meets the 'restricted to a specific virtual network' requirement, while the other options either allow public exposure or do not enforce VNet-level isolation.

How to eliminate wrong answers

Option A is wrong because using a service endpoint with a managed identity only secures network access at the subnet level and provides identity-based authentication, but it does not restrict the endpoint to a specific virtual network in the same way a private endpoint does, and it does not enable CMK for encryption at rest. Option B is wrong because disabling public network access alone does not restrict access to a specific virtual network; it only blocks all public traffic, and while enabling CMK via Azure Key Vault is correct for encryption, the network requirement is not met. Option C is wrong because configuring IP firewall rules only restricts access based on source IP addresses, not to a specific virtual network, and while CMK via Azure Key Vault is correct, the network isolation is insufficient for a VNet-specific restriction.

Full explanation →

654

MCQmedium

A company is developing a conversational AI solution using Microsoft Copilot Studio. They want the copilot to answer questions based on a knowledge base of technical documents. Which data source integration should they use?

A.Azure AI Search

B.Azure Blob Storage

C.Azure SQL Database

D.Microsoft Lists

AnswerA

Azure AI Search indexes the technical documents and exposes them to Copilot Studio as a knowledge source, enabling retrieval-augmented answers grounded in that content. It satisfies the requirement to answer from a document knowledge base rather than relying solely on the model's pretrained knowledge.

Why this answer

Azure AI Search is the correct data source because it provides a search index that can be queried by Copilot Studio using the 'Azure AI Search' connector. This allows the copilot to perform semantic or keyword-based retrieval over indexed technical documents, enabling accurate question-answering from a knowledge base. Copilot Studio natively supports Azure AI Search as a data source for generative answers, making it the optimal choice for this scenario.

Exam trap

The trap here is that candidates often confuse data storage (Blob Storage, SQL Database) with data retrieval and search capabilities, assuming any storage service can be directly used for Q&A, but Copilot Studio requires a search-optimized index like Azure AI Search to perform effective knowledge base queries.

How to eliminate wrong answers

Option B is wrong because Azure Blob Storage is a raw object storage service that does not provide built-in search capabilities; Copilot Studio cannot directly query blobs for question-answering without an indexing layer like Azure AI Search. Option C is wrong because Azure SQL Database is a relational database designed for transactional workloads, not for full-text or semantic search over unstructured technical documents; while it can be queried, it lacks the optimized search and ranking features needed for knowledge base retrieval. Option D is wrong because Microsoft Lists is a simple data-tracking tool for small-scale lists and lacks the indexing, scoring, and natural language query support required for a production knowledge base; it is not designed for document-based Q&A.

Full explanation →

655

MCQeasy

A company is building an agent that needs to perform tasks like sending emails and updating a CRM system. The agent uses Azure OpenAI with function calling. The team defines functions for these tasks. When the agent is tested, it sometimes calls the wrong function or invents function names. What should the team do to improve the reliability of function calling?

A.Fine-tune the model on a dataset of correct function calls.

B.Reduce the number of functions to only the most common ones.

C.Set the temperature parameter to 0 for deterministic output.

D.Provide better function descriptions with examples of when to use each function.

AnswerD

Function-calling accuracy depends on the model's understanding of each function's purpose. Richer descriptions stating when to invoke each function, plus concrete examples, reduce ambiguity and stop the model inventing or mis-selecting names, directly improving reliability.

Why this answer

Providing better function descriptions with examples directly improves the model's ability to select the appropriate function. Azure OpenAI's function calling relies on the semantic understanding of the function definitions; clear descriptions and usage examples reduce ambiguity, helping the model map user intent to the correct function signature without hallucinating names.

Exam trap

The trap here is that candidates often assume deterministic output (temperature=0) or reducing complexity (fewer functions) will fix reliability, when the real issue is semantic ambiguity in function definitions that the model cannot resolve without better descriptions.

How to eliminate wrong answers

Option A is wrong because fine-tuning on a dataset of correct function calls is unnecessary and inefficient; Azure OpenAI's base models already understand function calling patterns, and fine-tuning would require a large, curated dataset and could introduce overfitting or degrade general performance. Option B is wrong because reducing the number of functions limits the agent's capabilities and does not address the root cause of incorrect selection; the model may still invent names if descriptions are poor. Option C is wrong because setting temperature to 0 makes output deterministic but does not fix ambiguous or poorly defined function descriptions; the model will still confidently choose the wrong function if it misinterprets the intent.

Full explanation →

656

MCQmedium

You are building an Azure AI solution that uses Azure OpenAI Service to generate text. You need to ensure that the solution can handle up to 10,000 requests per minute. You also need to monitor the usage and set up alerts when the request rate exceeds 80% of the quota. What should you do?

A.Configure autoscaling for the Azure OpenAI Service in the Azure portal and set up alerts using Azure Service Health.

B.Create a deployment in Azure OpenAI Service with a high tokens-per-minute (TPM) quota and configure Azure Monitor alerts on the Azure OpenAI resource metrics.

C.Deploy multiple Azure OpenAI resources in different regions and use Azure Front Door to load-balance requests.

D.Use Azure API Management to throttle requests to the Azure OpenAI Service and configure Application Insights for monitoring.

AnswerB

Azure OpenAI Service deployments have quota limits measured in tokens per minute (TPM). To handle high request rates, you need to request a quota increase and deploy a model with sufficient TPM. Azure Monitor can track metrics such as total calls and token usage. You can create alert rules based on these metrics to notify when usage approaches the quota. This approach directly addresses both capacity and monitoring.

Why this answer

To handle high request rates, you must ensure the Azure OpenAI deployment has a sufficient tokens-per-minute quota. You can request a quota increase and deploy a model with adequate TPM. Azure Monitor can track metrics like total calls and token usage, and you can create alert rules to notify when usage exceeds a threshold.

The other options do not directly address quota management and monitoring of the service.

Exam trap

The trap here is assuming that Azure OpenAI Service supports autoscaling like other Azure services, when in fact quotas are fixed and require manual increases.

Full explanation →

657

MCQeasy

You are designing an Azure AI Search solution that indexes documents from an Azure SQL Database. The documents include a field named 'content' that contains HTML markup. You need to strip the HTML tags and extract only the plain text before applying further enrichment. Which built-in skill should you use?

A.Text Merger skill

B.Text Split skill

C.HTML Strip skill

D.Language Detection skill

AnswerC

The HTML Strip skill is a built-in cognitive skill that removes HTML tags from a string and returns plain text. It is designed exactly for this scenario: cleaning HTML content before further processing. By using this skill, you ensure that subsequent enrichment skills receive clean text, improving the accuracy of language detection, entity recognition, and other NLP tasks.

Why this answer

The HTML Strip skill is specifically designed to remove HTML markup and return plain text. It is the correct choice for cleaning HTML content before applying other enrichment skills. The Text Merger and Text Split skills manipulate text structure but do not remove HTML.

Language Detection analyzes language but does not alter the text content.

Exam trap

The trap here is assuming that Text Split or Text Merger can also clean HTML, when they only restructure text without removing markup.

Full explanation →

658

MCQhard

You are using Azure AI Language's conversational language understanding (CLU). The above JSON is a request to a CLU endpoint. What is the purpose of this request?

A.To predict the intent and entities from the user utterance

B.To query a knowledge base for answers

C.To deploy the CLU model to production

D.To train a new CLU model

AnswerA

The request body supplies a user utterance to the CLU prediction endpoint, which returns the top-scoring intent plus any extracted entities. This satisfies the scenario's need to interpret a conversational input, since CLU's runtime API performs intent classification and entity extraction in a single call.

Why this answer

The JSON request is sent to the Azure AI Language CLU endpoint with a 'query' field containing the user utterance. The 'kind' field is set to 'Conversation', which triggers the CLU runtime to analyze the utterance against the deployed model. The purpose is to return a prediction of the top intent and any extracted entities, which is the core function of a conversational language understanding endpoint.

Exam trap

The trap here is that candidates confuse the CLU prediction endpoint with the training or deployment endpoints, mistakenly thinking a request with a 'query' field is used for model management rather than runtime inference.

How to eliminate wrong answers

Option B is wrong because querying a knowledge base for answers is the purpose of Azure AI Language's custom question answering (QnA Maker) or Azure Cognitive Search, not CLU. Option C is wrong because deploying a CLU model is a separate operation performed via the Azure portal, REST API (e.g., PUT on the deployment resource), or SDK; this request is a prediction call, not a deployment action. Option D is wrong because training a new CLU model requires a training API call (e.g., POST to the /train endpoint with a training dataset), not a prediction request to the runtime endpoint.

Full explanation →

659

MCQhard

Your Azure AI Search index contains millions of documents. Users report that search results are slow for complex queries. You need to improve query performance without reducing result quality. Which action should you take?

A.Reduce the maximum number of results returned per query

B.Increase the number of replicas

C.Remove all facet fields from the index

D.Disable complex query types such as fuzzy and regex

AnswerB

Adding replicas increases the number of copies of the index that serve queries, distributing concurrent search load and reducing latency. This satisfies the stem's constraint of improving complex query performance without altering analysers, scoring profiles or result quality.

Why this answer

Increasing the number of replicas in Azure AI Search distributes query load across multiple copies of the index, enabling parallel processing of complex queries. This directly improves query throughput and latency without altering the index schema or reducing result quality, as replicas provide dedicated resources for query execution.

Exam trap

The trap here is that candidates confuse replicas (which improve query performance and availability) with partitions (which improve indexing speed and storage capacity), leading them to choose options that degrade functionality instead of scaling resources.

How to eliminate wrong answers

Option A is wrong because reducing the maximum number of results per query (e.g., via $top) only limits the response payload and does not address the underlying computational cost of complex queries; it can also degrade user experience by hiding relevant results. Option C is wrong because removing facet fields eliminates aggregation capabilities and does not improve query performance—facets are computed during indexing, not at query time, and their removal would reduce result quality by removing navigation aids. Option D is wrong because disabling complex query types (fuzzy, regex) restricts search functionality and may reduce result relevance; while these queries are resource-intensive, the correct approach is to scale out via replicas rather than sacrifice search capabilities.

Full explanation →

660

MCQeasy

You are implementing a chatbot using Microsoft Copilot Studio that helps employees find company policies. The chatbot must: - Use generative answers based on a SharePoint Online site. - Only respond with information from approved policy documents. - Include citations in responses. - Be accessible from Microsoft Teams. - Require no custom code. What should you do?

A.Use Power Automate to retrieve documents and feed them to Azure OpenAI. Build a custom connector for Teams.

B.In Copilot Studio, create a new copilot. Add the SharePoint site as a knowledge source. Enable generative answers with citations. Publish to Teams.

C.Build a bot using Azure Bot Service and QnA Maker. Train it with the policy documents. Deploy to Teams.

D.Create a custom GPT in Azure OpenAI Studio. Upload the policy documents. Deploy via Azure API Management and expose to Teams.

AnswerB

Adding the SharePoint site as a knowledge source with generative answers and citations enabled restricts responses to approved policy documents, and publishing to Teams delivers the required channel. No custom code is needed, meeting every stated constraint.

Why this answer

Microsoft Copilot Studio natively supports adding a SharePoint Online site as a knowledge source, enabling generative answers that retrieve and cite only approved policy documents. It requires no custom code, automatically includes citations in responses, and can be published directly to Microsoft Teams, fulfilling all stated requirements.

Exam trap

The trap here is that candidates may overcomplicate the solution by choosing Azure OpenAI or Azure Bot Service options, missing that Copilot Studio is the no-code, fully integrated tool designed specifically for this scenario with built-in SharePoint knowledge sources, citations, and Teams deployment.

How to eliminate wrong answers

Option A is wrong because it requires custom code (Power Automate flow, custom connector) and Azure OpenAI, which violates the 'no custom code' requirement and adds unnecessary complexity. Option C is wrong because QnA Maker is deprecated and does not support generative answers with citations from SharePoint; it also requires manual training and custom deployment to Teams. Option D is wrong because creating a custom GPT in Azure OpenAI Studio and deploying via Azure API Management involves custom code and infrastructure management, contradicting the 'no custom code' and 'accessible from Teams' requirements without additional integration.

Full explanation →

661

MCQeasy

Your organization needs to analyze customer feedback from social media posts to determine the sentiment (positive, negative, neutral). The solution must process up to 10,000 posts per day and provide a confidence score for each sentiment. Which Azure AI service should you use?

A.Azure AI Speech Service

B.Azure AI Language Service

C.Azure AI Translator

D.Azure AI Language Understanding (LUIS)

AnswerB

Azure AI Language Service provides sentiment analysis with per-document confidence scores for positive, negative, and neutral labels, directly meeting the stem's scoring requirement. Its hosted endpoint scales to handle 10,000 daily social media posts without custom model training, unlike Azure AI Vision or Speech, which address different modalities.

Why this answer

Azure AI Language Service provides pre-built sentiment analysis capabilities that can process up to 10,000 posts per day and return a confidence score for each sentiment (positive, negative, neutral). This service is specifically designed for natural language processing tasks like sentiment analysis, making it the correct choice for analyzing customer feedback from social media posts.

Exam trap

The trap here is that candidates often confuse Azure AI Language Service with LUIS, assuming both are for language understanding, but LUIS is specifically for intent and entity extraction in conversational AI, not for general sentiment analysis with confidence scores.

How to eliminate wrong answers

Option A is wrong because Azure AI Speech Service is designed for speech-to-text, text-to-speech, and speech translation, not for analyzing text sentiment from social media posts. Option C is wrong because Azure AI Translator focuses on translating text between languages, not on determining sentiment or providing confidence scores. Option D is wrong because Azure AI Language Understanding (LUIS) is a conversational AI service for intent recognition and entity extraction in chatbots, not for general-purpose sentiment analysis with confidence scores.

Full explanation →

662

MCQmedium

You are using Azure OpenAI Service to generate marketing copy. The marketing team reports that the generated content sometimes contains factual inaccuracies. You need to improve the factual accuracy of the generated content. What should you do?

A.Increase the max_tokens parameter

B.Include relevant context and facts in the prompt

C.Decrease the temperature parameter

D.Disable content filtering

AnswerB

Grounding the model with relevant facts and context in the prompt constrains generation to supplied information, reducing hallucinated claims. This directly addresses the factual inaccuracy problem by giving the model authoritative source material rather than relying on parametric knowledge alone.

Why this answer

The most effective way to improve factual accuracy in Azure OpenAI generations is to ground the model with relevant context and facts in the prompt — this is the core of Retrieval-Augmented Generation (RAG). By supplying authoritative source content, the model conditions its output on verified information rather than relying solely on parametric memory, which reduces hallucinations. This directly addresses the marketing team's complaint about factual inaccuracies.

Exam trap

AI-102 often tests the misconception that lowering temperature or increasing tokens improves factual accuracy — candidates must recognize that grounding with context (RAG) is the correct approach to reduce hallucinations.

How to eliminate wrong answers

Option A is wrong because max_tokens only controls the length of the generated output, not its factual correctness — increasing it can even allow more room for fabricated content. Option C is wrong because lowering temperature reduces randomness and makes outputs more deterministic, but a deterministic model can still confidently state incorrect facts; temperature does not inject or remove knowledge. Option D is wrong because disabling content filtering removes safety guardrails against harmful content and has no bearing on factual accuracy — it may even increase risk of inappropriate outputs.

Full explanation →

663

Multi-Selecthard

Which THREE factors should you consider when choosing between Azure AI Document Intelligence prebuilt models and custom models for invoice processing?

Select 3 answers

A.Both model types can be deployed on-premises.

B.Prebuilt models require no training data.

C.Prebuilt models are always less accurate than custom models.

D.Custom models require a large set of labeled training invoices.

E.Custom models can handle non-standard invoice layouts.

AnswersB, D, E

Prebuilt models are trained by Microsoft on large document corpora, so you supply no labelled samples and can call them immediately. This removes data collection and training effort, a decisive factor when your documents match the supported schema and you need fast deployment.

Why this answer

Option B is correct because Azure AI Document Intelligence prebuilt invoice models are pretrained by Microsoft and can be invoked immediately without supplying any labeled training data, which is ideal when you want fast time-to-value on standard invoices. Option D is correct because custom models (template or neural) are trained on your own labeled invoice samples, and although the exact count varies by model type, a meaningful set of labeled invoices is required to teach the model your specific fields and layouts. Option E is correct because custom models are specifically designed to handle non-standard, supplier-specific, or unusual invoice layouts that prebuilt models may not parse accurately.

Option A is not correct because these are Azure cloud services accessed via the Document Intelligence endpoint/API, not on-premises deployable models. Option C is not correct because prebuilt models are not always less accurate than custom models; for standard invoice formats they can perform very well, and accuracy depends on the document set and training quality.

Exam trap

The trap here is that candidates assume prebuilt models are always less accurate than custom models, but accuracy depends on the document's similarity to the training data; prebuilt models can outperform custom ones on standard layouts, especially when training data is limited.

Full explanation →

664

MCQeasy

You need to extract key-value pairs from scanned forms as part of a knowledge mining solution. Which Azure AI service should you use?

A.Azure AI Vision

B.Azure AI Language

C.Azure AI Search

D.Azure AI Document Intelligence

AnswerD

Document Intelligence provides prebuilt and custom models that return structured key-value pairs from scanned forms, unlike pure OCR which yields unstructured text. This directly satisfies the requirement to extract key-value pairs from scanned forms within the knowledge mining pipeline.

Why this answer

Azure AI Document Intelligence (formerly Form Recognizer) is the correct service because it is specifically designed to extract key-value pairs, tables, and structured data from scanned forms and documents using prebuilt and custom models. This aligns directly with the requirement for knowledge mining from scanned forms.

Exam trap

The trap here is that candidates often confuse Azure AI Vision's OCR capability with form-specific extraction, not realizing that Document Intelligence is the dedicated service for key-value pair extraction from scanned forms, while Vision only provides raw text coordinates without semantic understanding.

How to eliminate wrong answers

Option A is wrong because Azure AI Vision provides image analysis capabilities like OCR, object detection, and captioning, but it does not have native support for extracting key-value pairs from forms; it would require additional processing to structure the data. Option B is wrong because Azure AI Language focuses on text analytics, sentiment analysis, and entity recognition from written text, not from scanned forms or document layouts. Option C is wrong because Azure AI Search is a search indexing and query service that can index extracted data but does not perform the extraction itself; it relies on other services like Document Intelligence to provide the structured input.

Full explanation →

665

MCQeasy

You are prototyping a chat experience on Azure OpenAI and want the model to produce structured JSON that matches a schema your application can deserialize reliably. Which feature should you configure?

A.Enable content filtering and set the severity threshold to high.

B.Increase top_p to 1 and rely on the model's instruction-following ability.

C.Set response_format to json_schema with a strict schema definition on the chat completions call.

D.Set temperature to 0 and add the word JSON to the user prompt.

AnswerC

Structured outputs with a strict JSON schema constrain generation so the response conforms to the supplied schema, which makes deserialization dependable. This is the purpose-built mechanism for schema-constrained JSON in chat completions and is more reliable than prompt-only instructions, because the model is constrained during decoding rather than merely asked to comply.

Why this answer

Schema-constrained generation is the only option that enforces structure at decode time. Configuring response_format with a strict JSON schema makes the model emit payloads that conform to the defined fields and types, so the application can deserialize without defensive parsing. Sampling parameters, prompt hints, and content filters do not enforce schema conformance.

Exam trap

The trap here is believing that setting temperature to zero or naming JSON in the prompt guarantees valid JSON, when only schema-constrained decoding enforces the contract.

Full explanation →

666

Multi-Selecteasy

You are deploying a chat application using Azure OpenAI. The application should only answer questions based on a specific set of internal documents. Which THREE features should you use?

Select 3 answers

A.Azure AI Search index with the internal documents

B.Grounding with your data in Azure OpenAI Studio

C.Content filters to block out-of-domain questions

D.System message to limit the assistant's scope

E.Fine-tuning the model on the internal documents

AnswersA, B, D

The index provides the data source for grounding.

Why this answer

Azure AI Search indexes allow you to ingest internal documents and perform vector or hybrid search over them. When integrated with Azure OpenAI, the search results are used as grounding context for the model, ensuring responses are based solely on your data.

Exam trap

Microsoft often tests the distinction between content filtering (which handles safety) and domain restriction (which requires retrieval or prompt engineering), leading candidates to incorrectly select content filters for limiting question scope.

Full explanation →

667

MCQeasy

You need to generate an image of a cat wearing a hat using Azure OpenAI. Which model should you use?

A.Codex

B.DALL-E

C.GPT-4

D.Whisper

AnswerB

DALL-E is Azure OpenAI's dedicated image-generation model, so it produces a new picture from a text prompt such as a cat wearing a hat. GPT models output text, not images, so they cannot satisfy the image-generation requirement.

Why this answer

DALL-E is the Azure OpenAI model specifically designed for generating images from natural language descriptions. It uses a diffusion-based architecture to create high-quality, original images based on text prompts, making it the correct choice for generating an image of a cat wearing a hat.

Exam trap

The trap here is that candidates may confuse GPT-4's multimodal capabilities (which can analyze images but not generate them) with DALL-E's generative image creation, leading them to incorrectly select GPT-4 for image generation tasks.

How to eliminate wrong answers

Option A is wrong because Codex is a model specialized for generating code from natural language, not for image generation. Option C is wrong because GPT-4 is a large language model focused on text generation and reasoning, lacking native image generation capabilities. Option D is wrong because Whisper is a speech-to-text model designed for audio transcription, not image generation.

Full explanation →

668

MCQhard

You are deploying a conversational language understanding project in Azure AI Language for a banking chatbot. Testing shows the model frequently confuses the intents TransferFunds and PayBill because both utterances contain similar wording about moving money. You need to improve the model's ability to distinguish these two intents. What should you do?

A.Increase the model's confidence threshold so low-confidence predictions are rejected instead of misclassified

B.Merge the two intents into a single intent and let the bot ask a follow-up question to determine the action

C.Add more labeled utterances that are representative of each intent, including boundary examples that clarify the difference between them

D.Add a prebuilt entity such as Money to both intents so the model can use amounts to tell them apart

AnswerC

Intent confusion between semantically close classes is resolved by increasing and diversifying labeled utterances for both intents, especially examples near the decision boundary. Adding utterances that emphasize the distinguishing features, such as payee type or timing, gives the model the signal it needs to separate the classes. This directly targets the observed confusion rather than changing unrelated configuration.

Why this answer

When two intents overlap in wording, the model needs more and better-labeled examples that highlight their distinguishing characteristics. Adding representative and boundary utterances for both intents gives the training algorithm the discriminative signal required to separate them. Adjusting confidence thresholds, merging intents, or adding entities that appear in both classes does not address the underlying classification boundary.

Exam trap

The trap here is reaching for a confidence threshold change to suppress wrong predictions, when the real fix for two confusable intents is more discriminative labeled utterances.

Full explanation →

669

Multi-Selectmedium

Which THREE factors should be considered when choosing between Azure Computer Vision and Azure Custom Vision? (Choose three.)

Select 3 answers

A.The need for custom model retraining over time.

B.Whether the solution runs on edge devices.

C.The amount of labeled training data available.

D.The geographic region of the Azure subscription.

E.Whether the detection objects are generic or domain-specific.

AnswersA, C, E

Custom Vision allows retraining.

Why this answer

Azure Custom Vision is specifically designed for scenarios where you need to retrain a model over time with new labeled data, such as when the visual characteristics of objects change (e.g., new product packaging). Azure Computer Vision is a pre-trained API that cannot be retrained; it only supports fixed, generic models. Custom Vision allows iterative training with your own images, making it essential when model drift or evolving requirements demand periodic retraining.

Exam trap

Microsoft often tests the misconception that edge deployment is exclusive to Custom Vision, but in reality, both services support containerized edge deployment, so the true differentiator is the need for custom retraining and domain-specific detection.

Full explanation →

670

MCQeasy

You need to deploy a generative AI model that can be used by multiple applications within your organization. The model must support real-time inference with low latency. Which Azure service should you use?

A.Azure AI Search

B.Azure OpenAI Service

C.Azure Machine Learning real-time endpoint

D.Azure Functions

AnswerB

Azure OpenAI Service hosts generative models behind a managed REST endpoint with provisioned throughput, giving multiple applications shared, low-latency real-time inference. It satisfies the low-latency constraint that batch-oriented or self-hosted alternatives cannot meet as directly.

Why this answer

Azure OpenAI Service provides managed access to powerful generative AI models like GPT-4, which are optimized for real-time inference with low latency through provisioned throughput units (PTUs) and regional deployment options. This service is specifically designed for generative AI workloads, offering REST API endpoints that support streaming responses and sub-second latency for single-turn interactions, making it ideal for multiple applications requiring consistent, low-latency responses.

Exam trap

The trap here is that candidates often confuse Azure Machine Learning real-time endpoints (which are for custom ML models) with Azure OpenAI Service (which is purpose-built for generative AI), overlooking the fact that Azure OpenAI provides managed, low-latency inference optimized for large language models without the overhead of containerized deployments.

How to eliminate wrong answers

Option A is wrong because Azure AI Search is a retrieval service for indexing and querying vector and keyword data, not a generative AI model deployment service; it lacks native model inference capabilities. Option C is wrong because Azure Machine Learning real-time endpoints are designed for custom ML model deployment, but they introduce higher latency due to container startup times and lack the optimized inference infrastructure (e.g., PTUs) that Azure OpenAI provides for generative models. Option D is wrong because Azure Functions is a serverless compute service for event-driven code execution, not a model hosting platform; it would require manual integration with a model endpoint and cannot guarantee the low latency needed for real-time generative AI inference.

Full explanation →

671

MCQmedium

Refer to the exhibit. A developer received this response from an Azure OpenAI chat completion call. The prompt was "What is the capital of France?". The finish_reason is "stop". What does this indicate?

A.The response was truncated due to content filtering.

B.The model completed the response naturally.

C.The model stopped generating before the response was complete.

D.The response reached the max_tokens limit.

AnswerB

The finish_reason field reports why generation halted. A value of "stop" means the model emitted its natural end-of-sequence token, producing a complete answer rather than being cut off by the max_tokens limit or filtered by a content policy.

Why this answer

The finish_reason 'stop' indicates that the model completed the response naturally, meaning it generated a complete answer to the prompt and reached a logical stopping point (e.g., the end of a sentence or the end of the generated text). This is the standard behavior for a successful completion where the model did not encounter any content filter, token limit, or other interruption.

Exam trap

Microsoft often tests the distinction between finish_reason values, and the trap here is that candidates confuse 'stop' with 'length' or assume any non-error finish_reason means truncation, when in fact 'stop' explicitly signals a natural and complete generation.

How to eliminate wrong answers

Option A is wrong because 'stop' specifically means the model finished generating on its own, not that content filtering truncated the response; content filtering would return a finish_reason of 'content_filter'. Option C is wrong because 'stop' indicates the model completed the response, not that it stopped prematurely; a premature stop would be indicated by a finish_reason of 'length' (if max_tokens hit) or 'null' (if interrupted). Option D is wrong because reaching the max_tokens limit would result in a finish_reason of 'length', not 'stop'.

Full explanation →

672

MCQmedium

Your company uses Azure AI Vision to analyze images. You receive an alert that the number of 429 (Too Many Requests) errors has increased significantly. What is the most likely cause?

A.The endpoint URL is incorrect.

B.The API key has expired.

C.The service principal does not have the correct role assignment.

D.The application is exceeding the transactions-per-second (TPS) limit.

AnswerD

Exceeding the transactions-per-second limit directly triggers HTTP 429 responses, since Azure AI Vision throttles requests once the assigned TPS quota is surpassed. The stem's surge in Too Many Requests errors therefore points to throughput saturation rather than authentication or payload faults. Raising the tier or implementing retry-after backoff resolves it.

Why this answer

HTTP 429 (Too Many Requests) is a rate-limiting response from Azure AI Vision when the client exceeds the allowed transactions-per-second (TPS) for the chosen pricing tier. The alert indicates the application is sending requests faster than the service's capacity, triggering throttling to protect backend resources.

Exam trap

The trap here is confusing HTTP 429 with authentication or authorization errors (401/403), leading candidates to incorrectly select options about API keys or role assignments when the real issue is rate limiting.

How to eliminate wrong answers

Option A is wrong because an incorrect endpoint URL would produce a 404 Not Found or connection error, not a 429 rate-limit error. Option B is wrong because an expired API key results in a 401 Unauthorized or 403 Forbidden response, not a 429. Option C is wrong because an incorrect role assignment on the service principal would cause 403 Forbidden errors due to missing RBAC permissions, not a 429 throttling response.

Full explanation →

673

MCQhard

A legal compliance team needs to automatically redact personally identifiable information (PII) from legal documents before sharing them with external auditors. The documents are stored in Azure Blob Storage. The solution must use Azure AI Language to detect PII and then redact the identified entities. The redaction must be performed on the original documents, and the redacted versions must be saved to a separate container. You need to design a serverless solution with minimal latency. What should you do?

A.Use Azure Data Factory to copy documents to a processing location and call an Azure Function.

B.Use Azure Batch Service to process documents in parallel and redact PII.

C.Create an Azure Function triggered by Blob Storage events to detect PII, redact, and save to a separate container.

D.Use Azure Logic Apps with a Blob trigger to call the PII detection API and write redacted documents.

AnswerC

An Azure Function triggered by Blob Storage events is serverless, event-driven, and processes each document as it is added, making it the most efficient option with minimal latency. It can call the PII detection API and save the redacted document to another container.

Why this answer

An Azure Function triggered by Blob Storage events is the canonical serverless pattern for event-driven document processing: the blob-created event fires the function, which calls the Azure AI Language PII detection/redaction API and writes the redacted output to a separate container. This minimizes latency because processing starts immediately on upload, with no orchestration overhead and automatic scaling. It also satisfies the requirement to keep the original intact and save redacted copies elsewhere.

Exam trap

AI-102 often tests the difference between event-driven serverless (Functions with blob trigger) and orchestration services (Data Factory, Logic Apps, Batch), so candidates pick Logic Apps for 'serverless' without weighing the latency and throughput requirements.

How to eliminate wrong answers

Option A is wrong because Azure Data Factory is an orchestration/ETL service with higher startup latency and is not event-driven at the blob level without additional triggers, making it heavier than needed. Option B is wrong because Azure Batch is designed for large-scale parallel HPC workloads and requires pool management, job scheduling, and VM lifecycle handling — overkill and slower to start for per-document redaction. Option D is wrong because Logic Apps, while serverless, introduce connector overhead and higher per-execution latency than a direct blob-triggered function, and are less efficient for high-volume document processing.

Full explanation →

674

Multi-Selecteasy

Which TWO Azure AI services can be used to extract text from images as part of a knowledge mining pipeline?

Select 2 answers

A.Azure AI Language

B.Azure AI Document Intelligence

C.Azure AI Computer Vision

D.Azure AI Video Indexer

E.Azure AI Custom Vision

AnswersB, C

Document Intelligence's Read model performs OCR, returning printed and handwritten text from images and PDFs, and integrates into enrichment pipelines via the built-in Document Intelligence skill. It satisfies the requirement to extract text from images as part of knowledge mining.

Why this answer

Azure AI Document Intelligence (option B) is correct because its Read and Layout models perform OCR on documents and images, extracting printed and handwritten text along with structure for downstream knowledge mining enrichment. Azure AI Computer Vision (option C) is correct because its Read OCR feature (Image Analysis / Read API) extracts printed and handwritten text from images, a core skill in Azure AI Search cognitive skillsets. Azure AI Language (option A) is wrong because it handles text analytics such as entity recognition, sentiment, and key phrase extraction, not image OCR.

Azure AI Video Indexer (option D) is wrong because it targets video and audio insights like transcription and face tracking, not still-image text extraction. Azure AI Custom Vision (option E) is wrong because it trains image classification and object detection models, not text recognition.

Exam trap

The trap here is that candidates often confuse Azure AI Computer Vision's OCR capabilities with Azure AI Document Intelligence, but Document Intelligence is the dedicated service for structured document extraction in knowledge mining, while Computer Vision provides general-purpose image analysis and OCR without the same level of document-specific parsing.

Full explanation →

675

MCQeasy

A company wants to use Azure AI Language to automatically summarize large documents. The summarization must extract the most important sentences from each document. Which feature should they use?

A.Extractive summarization

B.Abstractive summarization

C.Key phrase extraction

D.Entity recognition

AnswerA

Extractive summarization returns the highest-scoring sentences verbatim from the source document, directly satisfying the requirement to pull the most important sentences rather than generate new phrasing. Abstractive summarization would instead rewrite content in fresh wording, which the stem explicitly excludes.

Why this answer

Extractive summarization selects the most important sentences directly from the source document to create a concise summary, preserving the original wording. This aligns with the requirement to extract key sentences without generating new text, making it the correct choice for this scenario.

Exam trap

The trap here is that candidates often confuse 'key phrase extraction' with summarization because both involve identifying important content, but key phrase extraction returns only isolated terms, not coherent sentences, which fails the requirement for a sentence-based summary.

How to eliminate wrong answers

Option B is wrong because abstractive summarization generates new sentences that paraphrase the content, rather than extracting existing sentences from the document. Option C is wrong because key phrase extraction identifies individual words or short phrases (e.g., 'machine learning', 'Azure'), not complete sentences, and does not produce a coherent summary. Option D is wrong because entity recognition identifies named entities (e.g., people, organizations, locations) within text, but does not extract or summarize sentences.

Full explanation →

Microsoft Azure AI Engineer Associate AI-102 (AI-102) — Questions 601–675