Microsoft Azure AI Engineer Associate AI-102 AI-102 Questions 676–750 | Page 10/11

676

Matchingmedium

Match each Azure Cognitive Services endpoint to its purpose.

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

Concepts

Matches

Analyze sentiment of text

Generate description of an image

Query a knowledge base

Detect faces in an image

Translate text between languages

Why these pairings

The correct matches are: Computer Vision -> image analysis, Text Analytics -> sentiment analysis, Form Recognizer -> form data extraction, Face API -> face detection and attributes. Common confusions include mixing up Computer Vision with Translator, and Text Analytics with Content Moderator.

Full explanation →

677

MCQhard

You are the Azure AI engineer for a large e-commerce company. The company uses Azure Computer Vision to automatically tag product images uploaded by sellers. The system has been running smoothly for months. However, after a recent update to the Computer Vision API, you notice that certain images of clothing items are being tagged with incorrect labels, such as 'shoe' for a shirt. The images are clear and well-lit. You have confirmed that the image format (JPEG) is supported and the size is within limits. The issue occurs consistently for clothing items with similar colors. Other product categories work fine. You suspect the issue is related to the API version. What should you do first?

A.Increase the image size limit.

B.Check the API version used in the application code and compare with the latest version.

C.Switch to a custom model trained on clothing items.

D.Reduce the confidence threshold to 50% to see if more tags appear.

AnswerB

Version-specific model changes can alter labelling behaviour, so verifying the API version in code against the latest release isolates whether the regression stems from the update. This is the first diagnostic step before retraining or altering image preprocessing.

Why this answer

The issue began after a Computer Vision API update, and the problem is specific to certain clothing images with similar colors, indicating a potential regression or behavioral change in the API version. Checking the API version used in the application code against the latest version is the first logical troubleshooting step to identify if a breaking change or bug was introduced. This aligns with Azure AI best practices: always verify API version compatibility before modifying thresholds or retraining models.

Exam trap

The trap here is that candidates may jump to retraining a custom model (Option C) or adjusting confidence thresholds (Option D) without first verifying the API version, which is the simplest and most cost-effective diagnostic step in Azure AI troubleshooting.

How to eliminate wrong answers

Option A is wrong because increasing the image size limit does not address incorrect labeling; the images are already within supported size limits and the issue is with tag accuracy, not file size. Option C is wrong because switching to a custom model is a significant investment and should only be considered after verifying that the pre-built API version is not the root cause; the problem may be a temporary API regression that a version rollback could fix. Option D is wrong because reducing the confidence threshold to 50% would increase the number of tags but not correct the mislabeling of a shirt as 'shoe'; it would likely introduce more false positives without fixing the core issue.

Full explanation →

678

MCQeasy

Your team has built a knowledge mining pipeline using Azure AI Search and Document Intelligence. After ingestion, you notice that some documents are not appearing in search results. What is the most likely cause?

A.The indexer encountered errors and marked the documents as failed

B.The index does not have a semantic configuration

C.The search service has insufficient replicas

D.The search service is throttled due to high query volume

AnswerA

Indexers record per-document status during enrichment; documents failing skill execution or field mapping are marked failed and omitted from the index, so they never surface in queries. Checking indexer execution history and error details identifies the specific failing documents.

Why this answer

When an indexer runs, it processes each document and can encounter errors such as unsupported file formats, corrupt content, or permission issues. If a document fails during indexing, the indexer records the error and does not add that document to the index, making it invisible to search queries. This is the most direct cause of missing documents after ingestion.

Other options affect search behavior but not whether documents are indexed in the first place.

Exam trap

AI-102 often tests the misconception that search service configuration (like replicas or semantic ranker) affects document indexing, when in fact indexing failures are the primary cause of missing documents.

How to eliminate wrong answers

Option B is wrong because a semantic configuration enhances ranking and relevance but is not required for documents to appear in search results; without it, documents are still indexed and searchable via full-text search. Option C is wrong because insufficient replicas affect query throughput and high availability, not whether documents are indexed; indexing is handled by indexers and the number of replicas does not determine document inclusion. Option D is wrong because throttling due to high query volume impacts query performance and may cause rate-limiting, but it does not prevent documents from being indexed; indexing and querying are separate workloads.

Full explanation →

679

MCQhard

You are deploying a Custom Vision object detection model to an Azure Container Instance for real-time inference. The model must respond within 500 ms. The default container runs on CPU. What should you do to meet the latency requirement?

A.Increase the number of CPU cores in the container instance.

B.Export the model as a Dockerfile with GPU support and deploy to a GPU-enabled ACI.

C.Deploy the model to Azure Functions with a Premium plan.

D.Use the Cognitive Services Computer Vision container instead.

AnswerB

GPU acceleration is key for low-latency object detection.

Why this answer

The default Custom Vision container runs on CPU, which is insufficient for real-time object detection inference within 500 ms. Exporting the model as a Dockerfile with GPU support and deploying to a GPU-enabled Azure Container Instance (ACI) leverages NVIDIA CUDA-accelerated inference, dramatically reducing latency to meet the sub-500 ms requirement.

Exam trap

The trap here is that candidates assume increasing CPU cores (Option A) is a valid performance fix, but Azure explicitly documents that Custom Vision object detection models require GPU acceleration for real-time latency under 500 ms, and the default CPU container is only suitable for batch or offline processing.

How to eliminate wrong answers

Option A is wrong because increasing CPU cores does not provide the parallel processing power needed for deep learning inference; object detection models like YOLO or Faster R-CNN require GPU acceleration for sub-500 ms latency. Option C is wrong because Azure Functions, even with a Premium plan, still runs on CPU and incurs cold-start latency, making it unsuitable for real-time inference under 500 ms. Option D is wrong because the Cognitive Services Computer Vision container is a pre-built container for general image analysis, not for deploying a custom-trained object detection model; it cannot be used to host your own Custom Vision model.

Full explanation →

680

MCQmedium

You are building an Azure AI solution that uses Azure AI Language to analyze customer feedback stored in an Azure Blob Storage container. The container contains 500,000 small text documents. You need to minimize the total time required to analyze all documents and minimize the number of API calls. What should you do?

A.Submit a separate synchronous request for each document by using the Analyze Text API.

B.Use Azure AI Document Intelligence to extract text from each document and then call the Azure AI Language API for each extracted document.

C.Combine all documents into a single request by using the Analyze Text API with a custom text analytics task.

D.Use the Analyze Text API asynchronously by submitting a batch job that references the documents in Azure Blob Storage.

AnswerD

Asynchronous batch processing is designed for large-scale analysis of documents in Azure Blob Storage. You submit one job that points to the container, and the service processes all 500,000 documents in parallel. This minimizes both the number of API calls (submit and poll) and the total elapsed time compared with per-document synchronous calls.

Why this answer

For large volumes of documents already stored in Azure Blob Storage, the asynchronous Analyze Text API is the intended mechanism. It accepts a batch job that references the storage container, processes documents in parallel, and returns results for retrieval. This approach reduces the number of API calls to a few (submit and poll) and shortens overall processing time compared with synchronous per-document calls.

Exam trap

The trap here is assuming that combining documents into a single synchronous request can reduce API calls, when the Analyze Text API enforces document count and payload size limits that make this impossible.

Full explanation →

681

Multi-Selectmedium

Which TWO actions should you take to optimize the performance of an Azure AI Search solution that indexes large volumes of data?

Select 2 answers

A.Use the appropriate search tier (S1, S2, etc.) based on document size

B.Use the free tier for production workloads

C.Increase the replica count for better indexing throughput

D.Batch documents in groups of up to 1000 per index operation

E.Disable scoring profiles to speed up indexing

AnswersA, D

Larger document volumes demand more storage, processing power and replica/partition capacity. Selecting a tier such as S1 or S2 sized to document volume directly addresses the stem's large-scale indexing constraint, preventing throttling and slow enrichment throughput.

Why this answer

Option A is correct because Azure AI Search tiers (Basic, S1, S2, S3, etc.) differ in storage size, document limits, partitions, and indexer throughput, so selecting a tier that matches your document size and volume is essential for indexing performance. Option D is correct because the Azure AI Search indexing API supports batching up to 1000 documents (or about 16 MB of payload) per index operation, which amortizes network and service overhead and dramatically improves indexing throughput. Option B is wrong because the Free tier is for evaluation only, with strict limits on storage, indexes, and indexers, and is not suitable for production workloads.

Option C is wrong because replicas provide high availability and increased query throughput, not indexing throughput; indexing scale is driven by partitions. Option E is wrong because scoring profiles affect query-time ranking, not indexing speed, and disabling them would not optimize indexing performance.

Exam trap

The trap here is confusing replicas (which scale query performance) with partitions (which scale indexing throughput), leading candidates to incorrectly select option C as a way to improve indexing speed.

Full explanation →

682

MCQmedium

You are building an Azure AI Search enrichment pipeline that enriches documents with key phrases detected by Azure AI Language. You need the key phrases to be returned as a structured collection that can be mapped to a Collection(Edm.String) field in the search index, and you must avoid storing the enriched text in the enrichment cache. Which output mapping configuration should you use in the indexer?

A.Define an outputFieldMapping with a source of /document/keyPhrases and a targetField of keyPhrases.

B.Define an outputFieldMapping with a source of /document/pages/*/keyPhrases and a targetField of keyPhrases.

C.Define an outputFieldMapping with a source of /document/keyPhrases and a targetField of keyPhrases, and set the indexer cache to disabled.

D.Define an outputFieldMapping with a source of /document/keyPhrases and a targetField of keyPhrases, and mark the field as retrievable and filterable.

AnswerA

The Key Phrases skill emits its result at the enrichment tree path /document/keyPhrases as a JSON array of strings. Mapping that source directly to a Collection(Edm.String) index field stores the phrases as a structured collection. Because the mapping targets an index field rather than a cache-only enrichment, the correct configuration preserves the intended behavior without persisting the enriched text in the cache.

Why this answer

The Key Phrases skill writes its output to the enrichment tree at /document/keyPhrases as an array of strings. To persist that array into a Collection(Edm.String) field, the indexer needs an outputFieldMapping whose source is that exact path and whose target is the index field. Cache behavior is configured separately on the indexer and is not part of output mapping.

Exam trap

The trap here is assuming the Key Phrases skill emits per-page results under /document/pages/*/keyPhrases, when it actually writes a single array at /document/keyPhrases for the whole document.

Full explanation →

683

MCQeasy

Refer to the exhibit. You are creating an Azure Cognitive Services account using an ARM template snippet. What type of account is being created?

A.Azure AI Language

B.Azure AI Computer Vision

C.Azure AI Services multi-service account

D.Azure OpenAI Service

AnswerC

The ARM snippet specifies `kind: "CognitiveServices"` with a multi-service SKU, provisioning one key and endpoint across Vision, Language, Speech and Translator. This satisfies the stem's requirement to identify a single Azure AI Services resource spanning multiple APIs, rather than a single-service account such as FormRecognizer or TextAnalytics.

Why this answer

The ARM template snippet uses the 'CognitiveServices' resource type and sets 'kind' to 'CognitiveServices', which provisions a multi-service account that provides access to multiple Azure AI services (e.g., Language, Computer Vision, Translator) under a single endpoint and key. This is distinct from single-service accounts, which use specific 'kind' values like 'TextAnalytics' or 'ComputerVision'.

Exam trap

The trap here is that candidates often confuse the 'CognitiveServices' kind (multi-service) with a specific single-service account, especially when the ARM template lacks explicit service-specific properties, leading them to pick a single-service option like Azure AI Language or Computer Vision.

How to eliminate wrong answers

Option A is wrong because Azure AI Language (formerly Text Analytics) is a single-service account created with 'kind': 'TextAnalytics', not 'CognitiveServices'. Option B is wrong because Azure AI Computer Vision is a single-service account created with 'kind': 'ComputerVision', not 'CognitiveServices'. Option D is wrong because Azure OpenAI Service uses a different resource type 'OpenAI' and 'kind': 'OpenAI', not the 'CognitiveServices' resource type.

Full explanation →

684

MCQeasy

You are deploying an Azure AI solution that uses multiple Azure AI services resources. You need to ensure that all resources are deployed in a consistent manner and can be managed as a single unit. You also need to be able to assign permissions to the entire group of resources at once. What should you use?

A.Management group

B.Azure Resource Manager template

C.Azure subscription

D.Resource group

AnswerD

A resource group is a logical container for Azure resources. Deploying all Azure AI services resources into the same resource group allows you to manage them as a single unit and assign role-based access control (RBAC) permissions at the resource group scope, which applies to all resources within it. This meets both the consistent deployment and centralized permission management requirements.

Why this answer

A resource group is the fundamental logical container for Azure resources. By placing all Azure AI services resources in one resource group, you can deploy them together using a template and assign RBAC permissions at the resource group scope, which cascades to all contained resources. This provides both consistent management and centralized access control without granting overly broad permissions.

Exam trap

The trap here is confusing the deployment mechanism (ARM template) with the management boundary (resource group), or choosing a broader scope like a subscription when a resource group suffices.

Full explanation →

685

MCQeasy

You deploy a custom vision model using Azure AI Custom Vision. After deployment, you notice the model has high accuracy on training data but low accuracy on new images. What is the most likely cause?

A.The training time was too short

B.The training dataset has too few images

C.The wrong domain was selected during training

D.The model is overfitted to the training data

AnswerD

Overfitting occurs when the model memorises training images rather than learning generalisable features, producing high training accuracy but poor performance on unseen images. The gap between training and new-image accuracy is the defining symptom, so more varied training data or augmentation is needed.

Why this answer

High accuracy on training data but low accuracy on new images is the classic symptom of overfitting, where the model has memorized the training examples (including noise and irrelevant patterns) rather than learning generalizable features. In Azure AI Custom Vision, this typically occurs when the training dataset is too small, too homogeneous, or lacks sufficient variation, causing the model to fail on unseen data.

Exam trap

The trap here is that candidates confuse 'too few images' (a contributing factor) with the direct diagnosis of 'overfitting,' but the question asks for the most likely cause of the described symptom, which is the overfitting itself, not its root cause.

How to eliminate wrong answers

Option A is wrong because training time in Custom Vision is automatically managed by the service; extending it does not directly cause overfitting—the model stops when convergence is reached. Option B is wrong because having too few images can contribute to overfitting, but the question asks for the 'most likely cause' given the symptom, and overfitting is the direct description of the behavior, not the root cause of small dataset size. Option C is wrong because selecting the wrong domain (e.g., 'General' vs. 'Food' or 'Landmarks') affects feature extraction and may reduce accuracy overall, but it does not specifically produce the pattern of high training accuracy and low test accuracy—that pattern is the hallmark of overfitting.

Full explanation →

686

Multi-Selecteasy

Which TWO statements about Azure OpenAI Service content filters are true?

Select 2 answers

A.They can be configured with severity levels (low, medium, high)

B.They only filter the output of the model

C.They cannot be customized for specific use cases

D.They are bypassed when using PTU deployments

E.They include categories such as hate, sexual, violence, and self-harm

AnswersA, E

Severity levels allow granular control over filtering.

Why this answer

Azure OpenAI Service content filters can be configured with severity levels (low, medium, high) to control the strictness of filtering for each content category. This allows administrators to fine-tune the filter sensitivity based on their application's risk tolerance and compliance requirements.

Exam trap

The trap here is that candidates often assume content filters only apply to model outputs (Option B) or that PTU deployments offer a way to bypass safety controls (Option D), but Azure enforces filters uniformly across all deployment types.

Full explanation →

687

MCQmedium

You are building a multilingual chatbot using Azure AI Language. For a given user utterance, you need to first detect the language, then route to the appropriate language-specific intent model. Which combination of Azure AI Language features should you use?

A.Language Detection and Conversational Language Understanding

B.Key Phrase Extraction and Conversational Language Understanding

C.Translator and Conversational Language Understanding

D.Language Detection and Custom Text Classification

AnswerA

Language Detection identifies the utterance's language, then Conversational Language Understanding applies the matching language-specific intent and entity model. This pairing satisfies the requirement to detect language first and route to the correct project for intent recognition.

Why this answer

The scenario requires first detecting the language of the user utterance (using Language Detection) and then routing to a language-specific intent model (using Conversational Language Understanding, which supports multiple languages in separate projects or deployments). This combination directly fulfills the requirement of language-aware intent routing.

Exam trap

The trap here is that candidates often confuse Translator (which changes the language) with Language Detection (which identifies the language without altering the text), leading them to choose Option C incorrectly.

How to eliminate wrong answers

Option B is wrong because Key Phrase Extraction identifies important terms in text but does not detect the language, so it cannot be used to route to a language-specific model. Option C is wrong because Translator translates text between languages, but the requirement is to detect the language and route to an existing intent model, not to translate the utterance before processing. Option D is wrong because Custom Text Classification assigns predefined labels to text but does not extract intents or entities in a conversational context, and it lacks the built-in language detection needed for routing.

Full explanation →

688

MCQmedium

You are designing a knowledge mining solution for a manufacturing company that needs to extract information from equipment maintenance manuals. The manuals are in multiple languages (English, French, German). You need to ensure that the extracted content is searchable in English only. Which approach should you use?

A.Use the Entity Recognition skill to extract entities and then index entities only.

B.Use the Language Detection skill to identify language and then index all content as-is.

C.Use the Text Translation skill to translate all content to English during indexing.

D.Use the Key Phrase Extraction skill to extract key phrases and then index them.

AnswerC

The Text Translation skill translates French and German content into English during indexing, so all extracted text is stored in English. This makes the index searchable in English only, regardless of the manuals' original languages.

Why this answer

Azure AI Search's Text Translation skill (backed by Azure AI Translator) translates non-English content into a target language during the enrichment pipeline, so all indexed content is normalized to English and searchable in English only. This directly satisfies the requirement to make multilingual manuals searchable in English. The skill runs at indexing time, so the index contains only English text.

Exam trap

AI-102 often tests the difference between enrichment skills that transform content (Translation) versus those that only extract metadata (Entity Recognition, Key Phrase Extraction), causing candidates to pick extraction skills when translation is required.

How to eliminate wrong answers

Option A is wrong because Entity Recognition extracts entities (people, places, organizations) but does not translate content, so French and German text would remain untranslated and unsearchable in English. Option B is wrong because Language Detection only tags the language; indexing content as-is leaves non-English text in the index, failing the English-only search requirement. Option D is wrong because Key Phrase Extraction pulls salient phrases but does not translate, so German/French phrases would still be indexed in their original language.

Full explanation →

689

MCQhard

You have a custom Named Entity Recognition (NER) model trained using Azure AI Language. The model is performing poorly on new data. You need to improve its accuracy. Which action should you take first?

A.Increase the training epochs.

B.Retrain the model using the same training data.

C.Review the test set results and add more labeled examples for entities with low precision/recall.

D.Reduce the number of entity types in the model.

AnswerC

Inspecting test set metrics exposes which entities have low precision or recall, so additional labelled examples target the actual weaknesses. Retraining blindly or adding unrelated data would not address the specific accuracy deficit the model exhibits on new data.

Why this answer

The first step to improve a custom NER model's accuracy is to analyze the test set results to identify which entity types have low precision or recall, then add more labeled examples for those specific entities. This targeted data augmentation addresses the root cause of poor performance—insufficient or imbalanced training data—rather than blindly adjusting hyperparameters or reducing complexity.

Exam trap

The trap here is that candidates often jump to hyperparameter tuning (epochs) or model simplification (reducing entity types) as a quick fix, when the core issue is almost always insufficient or low-quality labeled data for specific entities, which is the first diagnostic step in any custom NER workflow.

How to eliminate wrong answers

Option A is wrong because simply increasing training epochs without addressing data quality or quantity will likely lead to overfitting on the existing training data, not improve generalization to new data. Option B is wrong because retraining with the same training data will produce the same model with the same errors, offering no improvement. Option D is wrong because reducing the number of entity types may simplify the model but does not fix the underlying issue of poor labeling or insufficient examples for the entities that matter; it could also discard useful entity types that are correctly identified.

Full explanation →

690

MCQhard

You are designing a solution that must extract specific entities from customer emails, such as product names, order numbers, and dates. The solution must be able to learn from a small set of labeled examples and improve over time. Which Azure AI service should you use?

A.QnA Maker

B.Pre-built Entity Extraction

C.Text Analytics for health

D.Custom Entity Extraction

AnswerD

Custom Entity Extraction learns from a small set of labelled examples, letting you define product names, order numbers and dates as custom entities and retrain as more data arrives. This satisfies the requirement to improve over time, unlike prebuilt extraction which cannot adapt to your specific entity schema.

Why this answer

Custom Entity Extraction (D) is correct because it allows you to train a model with a small set of labeled examples to extract domain-specific entities like product names, order numbers, and dates from customer emails. This service supports iterative learning and improvement over time, making it ideal for scenarios where pre-built models lack the required specificity.

Exam trap

The trap here is that candidates often confuse pre-built entity extraction (which is out-of-the-box but inflexible) with custom entity extraction (which requires training but adapts to specific needs), leading them to choose Option B because they assume 'pre-built' means 'easier' without recognizing the requirement for custom entities.

How to eliminate wrong answers

Option A is wrong because QnA Maker is designed for building conversational question-answering bots over a knowledge base, not for extracting entities from unstructured text. Option B is wrong because Pre-built Entity Extraction provides fixed, general-purpose entity types (e.g., person, location) and cannot be trained on custom entities like product names or order numbers. Option C is wrong because Text Analytics for health is specialized for medical and healthcare entities (e.g., diagnoses, medications) and cannot be repurposed for general customer email entity extraction.

Full explanation →

691

MCQmedium

You are building a customer feedback dashboard with Azure AI Language. Analysts need to see, for each document, the overall sentiment and the sentiment expressed toward specific aspects such as shipping speed and product quality within the same request. The documents are short English reviews. Which feature and configuration should you use?

A.Call the sentiment analysis operation with opinion mining enabled and read the target and assessment pairs from the response.

B.Call the custom text classification operation with a multi-label project that has one label per aspect.

C.Call the key phrase extraction operation and infer aspect sentiment from the returned phrases.

D.Call the sentiment analysis operation with opinion mining enabled and specify the aspect targets in the request.

AnswerA

Opinion mining, also called aspect-based sentiment analysis, returns sentence-level sentiment plus target and assessment pairs that identify the aspect and the opinion expressed about it. Reading those pairs from the response gives analysts per-aspect sentiment such as shipping speed or product quality alongside the overall document sentiment in one call.

Why this answer

Aspect-based sentiment analysis, delivered through opinion mining on the sentiment analysis operation, returns the document sentiment plus target and assessment pairs that pair each aspect with its expressed opinion. That single call supplies both the overall sentiment and the per-aspect sentiment the dashboard requires, without needing separate models or manual inference.

Exam trap

The trap here is assuming you can pass the aspects you care about into the sentiment request instead of reading the targets the service detects.

Full explanation →

692

Multi-Selecteasy

Which TWO Azure services can be used to perform optical character recognition (OCR) on documents? (Select two.)

Select 2 answers

A.Azure AI Metrics Advisor

B.Azure AI Language

C.Azure AI Document Intelligence

D.Azure AI Personalizer

E.Azure AI Vision

AnswersC, E

Document Intelligence provides prebuilt and custom OCR models that read printed and handwritten text from documents, returning structured layout, key-value pairs and tables. This satisfies the scenario's requirement to perform optical character recognition on documents, going beyond simple text extraction.

Why this answer

Azure AI Document Intelligence (option C) is correct because it is the Azure service purpose-built for document processing, and its prebuilt Read and Layout models extract printed and handwritten text from documents and images using OCR. Azure AI Vision (option E) is also correct because its Image Analysis and Read capabilities include an OCR engine that extracts text from images and documents via the Read API. Azure AI Metrics Advisor (option A) is an anomaly-detection service for time-series data and performs no OCR.

Azure AI Language (option B) provides natural language processing such as sentiment analysis, key phrase extraction, and entity recognition, not optical character recognition. Azure AI Personalizer (option D) is a reinforcement-learning-based recommendation service and has no OCR functionality.

Exam trap

The trap here is that candidates often assume only Azure AI Vision (the Computer Vision service) can perform OCR, forgetting that Azure AI Document Intelligence also provides OCR as part of its document analysis capabilities, and both services are valid for OCR tasks depending on the scenario.

Full explanation →

693

Multi-Selecthard

Which THREE factors are critical to consider when designing a custom vision solution for a manufacturing quality inspection system?

Select 3 answers

A.Imbalance between defective and non-defective product samples.

B.Variation in lighting conditions across different inspection stations.

C.Inference latency requirements for real-time decisions.

D.The need for optical character recognition (OCR) of product serial numbers.

E.Multilingual support for labeling.

AnswersA, B, C

Class imbalance leads to biased models.

Why this answer

Class imbalance is a critical factor in custom vision solutions for manufacturing quality inspection. If defective samples are rare compared to non-defective ones, the model may become biased toward predicting the majority class, leading to poor recall for defects. Azure Custom Vision allows adjusting the probability threshold and using techniques like oversampling or weighted loss to mitigate this, but the imbalance must be accounted for during dataset preparation.

Exam trap

The trap here is that candidates may confuse peripheral requirements (like OCR or multilingual labels) with core design factors that directly impact model accuracy, latency, and robustness in a production vision system.

Full explanation →

694

MCQeasy

A developer is building a multilingual translation solution using Azure AI Translator. The solution must translate text to French, German, and Spanish. Which parameter should the developer set to specify the target language?

A.language

B.to

C.targetLanguage

D.from

AnswerB

The `to` parameter specifies the target language in Azure AI Translator requests, accepting one or more language codes such as fr, de, and es. This directly satisfies the stem's requirement to translate into French, German, and Spanish, since `to` defines output languages while `from` only declares the source.

Why this answer

In Azure AI Translator, the target language for translation is specified using the 'to' query parameter in the API request. This parameter accepts language codes such as 'fr' for French, 'de' for German, and 'es' for Spanish. The 'to' parameter is required and can be specified multiple times to translate into multiple target languages simultaneously.

Exam trap

The trap here is that candidates may confuse the parameter names with those from other Azure services (like 'targetLanguage' in Cognitive Services Text Analytics) or assume a generic 'language' parameter works, but Azure AI Translator specifically requires 'to' for the target language.

How to eliminate wrong answers

Option A is wrong because 'language' is not a valid parameter in the Azure AI Translator API; the correct parameter is 'to'. Option C is wrong because 'targetLanguage' is not a recognized parameter in the Azure AI Translator REST API; the API uses 'to' instead. Option D is wrong because 'from' specifies the source language, not the target language; it is optional and defaults to auto-detection if omitted.

Full explanation →

695

MCQhard

You are designing a multilingual chatbot using Azure AI Language. The chatbot must support English, Spanish, and French. You need to minimize development effort and ensure consistent intent recognition across languages. What should you do?

A.Use the Azure AI Translator to translate all utterances to English before sending to a single English-only CLU project.

B.Create a separate CLU project for each language and combine them with a routing mechanism.

C.Use a single CLU multilingual project that supports all three languages.

D.Use the Language Understanding (LUIS) service with a single app that includes language-specific utterances.

AnswerC

A single multilingual CLU project trains intents across English, Spanish and French simultaneously, so one model recognises intents in all three languages. This avoids building and maintaining three separate projects, minimising development effort while keeping recognition consistent.

Why this answer

A single CLU multilingual project in Azure AI Language natively supports multiple languages within one project, allowing you to define intents and entities once and provide utterances in English, Spanish, and French. The model learns shared semantic representations across languages, so intent recognition remains consistent without building separate models or translation pipelines. This directly minimizes development effort because you maintain one project, one deployment, and one set of intent definitions.

Exam trap

AI-102 often tests the misconception that you must either translate everything to English or build separate language models, when the correct answer is to use a single multilingual CLU project that natively handles multiple languages.

How to eliminate wrong answers

Option A is wrong because adding Azure AI Translator introduces an extra translation hop that can distort meaning, adds latency and cost, and still requires you to maintain a single English-only CLU project while handling translation failures and language detection separately. Option B is wrong because creating separate CLU projects per language multiplies training, deployment, and maintenance work, and a routing mechanism adds complexity while risking inconsistent intent behavior across models. Option D is wrong because LUIS is a legacy service being retired in favor of CLU, and a single LUIS app does not provide true multilingual training across English, Spanish, and French in the way a CLU multilingual project does.

Full explanation →

696

MCQhard

You are deploying an Azure AI solution that uses Azure OpenAI Service. The solution must be deployed in a way that minimizes latency for users in Asia. However, the company's data residency policy requires data to stay in the United States. What should you do?

A.Use Azure CDN to cache the model responses in Asia.

B.Deploy the Azure OpenAI Service in an Asian region and use Azure Front Door to route traffic.

C.Deploy the service in multiple regions globally and use Traffic Manager for routing.

D.Deploy the service in a US region and use Azure Front Door with caching to reduce latency.

AnswerD

Front Door provides low-latency access while keeping data in US.

Why this answer

It satisfies both requirements: data residency (deploying in a US region keeps data within the United States) and latency reduction for Asian users. Azure Front Door with caching stores frequently accessed model responses at edge locations closer to users in Asia, minimizing round-trip time without moving the origin data.

Exam trap

The trap here is that candidates assume caching (Option A) or global deployment (Option C) can solve latency without considering data residency, or they mistakenly think deploying in Asia (Option B) is acceptable despite the policy constraint.

How to eliminate wrong answers

Option A is wrong because Azure CDN caches static content, but Azure OpenAI Service responses are dynamic and often non-cacheable (e.g., unique prompts or streaming outputs); caching would not reduce latency for real-time inference. Option B is wrong because deploying in an Asian region violates the data residency policy requiring data to stay in the United States. Option C is wrong because deploying in multiple regions globally would require data replication outside the US, breaking the data residency constraint; Traffic Manager routes traffic but does not cache responses, so latency from a US region would remain high for Asian users.

Full explanation →

697

MCQmedium

You are building a custom entity extraction solution using Azure AI Language. You have a small dataset (50 documents) with annotated entities. You need to train a model that can extract similar entities from new documents. What is the best approach?

A.Create a custom NER project in Azure AI Language and train it with your annotated data.

B.Use the prebuilt entity recognition API to extract entities.

C.Use the Conversational PII entity extraction feature.

D.Use the built-in entity extraction skill in Azure AI Search.

AnswerA

Custom NER in Azure AI Language trains on your annotated documents, learning entity boundaries specific to your domain. With 50 labelled documents it satisfies the small-dataset constraint, extracting similar entities from new documents without prebuilt model limitations.

Why this answer

Azure AI Language's custom NER (Named Entity Recognition) feature allows you to train a model using your own annotated dataset. With 50 documents, you have enough labeled data to fine-tune a custom entity extraction model that learns the specific entity types and patterns in your domain, enabling accurate extraction from new documents.

Exam trap

The trap here is that candidates may assume prebuilt APIs or search skills can be adapted to custom entities, but Azure AI Language requires a dedicated custom NER project for training on your own annotated data.

How to eliminate wrong answers

Option B is wrong because the prebuilt entity recognition API extracts only a fixed set of common entities (e.g., person, organization, location) and cannot be customized to recognize domain-specific entities from your annotated data. Option C is wrong because Conversational PII entity extraction is designed to detect personally identifiable information (PII) in conversational text (e.g., chat logs), not for general custom entity extraction from documents. Option D is wrong because the built-in entity extraction skill in Azure AI Search is a preconfigured cognitive skill that uses prebuilt models; it does not support training on custom annotated data.

Full explanation →

698

Multi-Selectmedium

Which THREE practices should be followed to secure an Azure AI solution that uses Azure OpenAI Service and Azure AI Search?

Select 3 answers

A.Store API keys in Azure Key Vault but use them directly in application code.

B.Use managed identities to authenticate between Azure OpenAI and Azure AI Search.

C.Place all AI services in a DMZ subnet with public IP addresses.

D.Require that all client applications use HTTPS with TLS 1.2 or higher.

E.Enable firewall and private endpoints for all AI service endpoints.

AnswersB, D, E

Managed identities let Azure OpenAI authenticate to Azure AI Search without embedding keys or secrets in configuration, eliminating credential exposure and rotation overhead. This satisfies the requirement to secure service-to-service access within the AI solution.

Why this answer

Option B is correct because managed identities let Azure OpenAI and Azure AI Search authenticate to each other through Microsoft Entra ID without embedding secrets or connection strings in code, eliminating credential leakage and rotation overhead. Option D is correct because requiring HTTPS with TLS 1.2 or higher protects data in transit between client applications and the AI services, preventing interception or downgrade attacks on prompts, responses, and search queries. Option E is correct because enabling firewall rules and private endpoints on the AI service endpoints removes public internet exposure and restricts traffic to approved virtual networks, which is a core network-isolation control for Azure AI workloads.

Option A is not appropriate because using API keys directly in application code exposes secrets in source control, logs, and memory even if the keys are originally stored in Azure Key Vault. Option C is not appropriate because placing AI services in a DMZ subnet with public IP addresses increases the attack surface rather than securing the solution; private endpoints and restricted access are preferred.

Exam trap

The trap here is that candidates often think storing keys in Key Vault is sufficient for security, but the question tests whether you understand that managed identities eliminate the need to handle keys altogether, and that public endpoints (even in a DMZ) are not secure for AI services.

Full explanation →

699

MCQhard

You are using Azure AI Custom Vision to classify images of animals. The training set has 1000 images of cats and 1000 images of dogs. After training, the model performs well on the test set. However, when deployed, it misclassifies images of wolves as dogs. What is the most likely cause?

A.The training set does not include enough negative examples that look like dogs but are not.

B.The probability threshold is set too low.

C.The model is overfitted to the training data.

D.The training set has class imbalance.

AnswerA

The model learned dog features from images lacking wolf-like negatives, so it maps wolf visual traits onto the dog class. Adding negative examples resembling dogs but labelled otherwise would sharpen the decision boundary and satisfy the requirement to classify wolves correctly.

Why this answer

The model misclassifies wolves as dogs because the training set lacks negative examples that are visually similar to dogs but belong to a different class. Custom Vision learns to distinguish classes based on the features present in the training images; without images of wolf-like canines labeled as 'not dog,' the model has no basis to reject wolves. This is a classic case of insufficient hard negative mining, where the model generalizes too broadly for the 'dog' class.

Exam trap

Microsoft often tests the misconception that class imbalance is the primary cause of misclassification, but here the dataset is balanced, and the real issue is the lack of representative negative examples—a subtle but critical distinction in Custom Vision training.

How to eliminate wrong answers

Option B is wrong because the probability threshold controls the confidence required for a prediction, not the model's ability to distinguish between visually similar classes; lowering the threshold would increase false positives, not fix the underlying feature confusion. Option C is wrong because overfitting would cause poor performance on the test set, not specifically misclassify wolves as dogs; the model generalizes well to test images but fails on out-of-distribution examples like wolves. Option D is wrong because class imbalance is not present—the training set has equal numbers of cats and dogs (1000 each)—and imbalance would typically bias predictions toward the majority class, which is not the issue here.

Full explanation →

700

MCQhard

A retail company uses Azure AI Vision to analyze shelf images for inventory management. They notice that the Object Detection model sometimes misses small items. What is the most effective way to improve detection of small objects?

A.Preprocess images to remove background noise.

B.Train a custom object detection model with annotated images that include small objects.

C.Use the Background Removal API to isolate items.

D.Increase the image resolution before sending to the API.

AnswerB

Custom training with annotated images containing small objects teaches the model the specific visual features and scale variation needed, directly addressing missed detections. Generic pre-trained models lack this domain tuning, so retraining on representative shelf imagery is the effective remedy.

Why this answer

Training a custom object detection model with annotated images that include small objects directly improves the model's ability to detect them. Option A is wrong because preprocessing to remove background noise does not specifically target small object detection; the model may still miss small items. Option C is wrong because the Background Removal API is used for isolating items from the background, not for improving detection accuracy.

Option D is wrong although higher resolution can help, it is not as effective as training a custom model with properly annotated small objects, and it may increase cost and latency.

Full explanation →

701

Multi-Selecteasy

A company wants to deploy an agent using Azure Bot Service that integrates with Microsoft Teams. Which THREE steps should the team take?

Select 3 answers

A.Use the Bot Framework SDK to build the bot with Teams-specific features.

B.Create a Teams app manifest file with bot configuration.

C.Register the bot in the Azure portal and obtain a Microsoft App ID.

D.Deploy the bot code to an Azure Function.

E.Write the bot in C# using Azure SDK for .NET.

AnswersA, B, C

Building with the Bot Framework SDK provides the messaging endpoint and activity handling that Teams requires, including Teams-specific activity types and adaptive card support. It satisfies the integration constraint by producing a bot the Teams channel can register and route conversations to.

Why this answer

Option A is correct because the Bot Framework SDK provides the Teams-specific activities, Adaptive Cards, and middleware needed to build an agent that works properly inside Microsoft Teams. Option B is correct because a Teams app manifest (with the bot's ID, scopes, and commands) must be packaged and uploaded/sideloaded so Teams knows how to surface the bot to users. Option C is correct because registering the bot in Azure Bot Service yields the Microsoft App ID and password that the bot uses to authenticate with the Bot Connector and Teams channel.

Option D is not required: bot code can run on App Service, containers, or Functions, so deploying to Azure Functions is only one optional hosting choice. Option E is not required: the Bot Framework SDK supports multiple languages (C#, JavaScript, Python, Java), so writing in C# with the Azure SDK for .NET is not a mandatory step.

Exam trap

The trap here is that candidates often assume hosting (Azure Function) or language choice (C#) are mandatory steps, when in fact the three required steps are always: register the bot in Azure, build with the SDK, and create the Teams app manifest.

Full explanation →

702

MCQhard

An organization uses Azure AI Search to power an internal knowledge base. They notice that search results are returning irrelevant documents. The index includes a 'content' field with full text and a 'tags' field with metadata. Users often search for specific terms that appear in the 'tags' field. How should you configure the search index to improve relevance?

A.Add a custom scoring profile based on freshness.

B.Configure a scoring profile with a higher weight for the 'tags' field.

C.Set the 'tags' field to use the 'keyword' analyzer.

D.Enable semantic search on the 'content' field.

AnswerB

Scoring profiles apply field weights during query evaluation, so boosting the 'tags' field raises documents whose metadata matches the user's terms. This directly addresses the relevance problem, since tags carry the specific terms users search, whereas the full-text 'content' field dilutes matching scores.

Why this answer

Configuring a scoring profile with a higher weight for the 'tags' field increases the relevance score of documents where search terms match the tags, thereby prioritizing those results. Option A (freshness-based scoring) would favor newer documents but does not address matching on tags. Option C sets the 'tags' field to use the 'keyword' analyzer, which changes tokenization but does not adjust field weighting.

Option D enables semantic search on the 'content' field, which enhances understanding of natural language queries but does not specifically boost the weight of the tags field.

Full explanation →

703

MCQmedium

You are planning a solution that uses Azure AI Document Intelligence to process invoices. The solution must be deployed to a production environment with high availability. You need to use an Azure Resource Manager (ARM) template to deploy the Document Intelligence resource. The template must ensure that the resource is deployed to two Azure regions. What should you do?

A.Use the copy element in the ARM template to deploy multiple instances of the Document Intelligence resource with the same region but different names.

B.Deploy a single Document Intelligence resource and configure geo-replication through the Azure portal after deployment.

C.Create two separate Document Intelligence resources in the ARM template, each in a different region, and use a Traffic Manager profile to distribute requests.

D.Deploy the ARM template with a single Document Intelligence resource and set the location property to a variable that contains multiple regions.

AnswerC

This approach deploys two independent Document Intelligence resources in different regions, providing high availability. Traffic Manager can route traffic based on priority or performance. This is a common pattern for multi-region deployments. The ARM template can define both resources and the Traffic Manager profile, ensuring consistent deployment. This satisfies the requirement for high availability across regions.

Why this answer

To achieve high availability across regions for Azure AI Document Intelligence, you must deploy multiple resources, each in a different region. An ARM template can define both resources, and you can use Azure Traffic Manager to route requests. This ensures that if one region fails, the other can handle the load.

The other options either do not provide cross-region redundancy or rely on non-existent features.

Exam trap

The trap here is assuming that a single Azure resource can be deployed to multiple regions by specifying multiple locations in the location property.

Full explanation →

704

Multi-Selecthard

Which THREE actions can be performed using the Azure Custom Vision service?

Select 3 answers

A.Extract text from scanned receipts.

B.Export a trained model to ONNX format for offline inference.

C.Train a model to classify images of different product types.

D.Detect and locate multiple objects in an image with bounding boxes.

E.Identify specific individuals in a crowd using facial recognition.

AnswersB, C, D

Exporting a trained model to ONNX format is supported by Custom Vision, enabling offline inference on edge devices without cloud connectivity. This satisfies the scenario's requirement for local, disconnected prediction, since ONNX provides a portable, framework-agnostic representation that runs outside Azure while preserving the trained classifier's behaviour.

Why this answer

Option B is correct because Azure Custom Vision supports exporting trained models in several formats, including ONNX, TensorFlow, CoreML, and Docker, enabling offline or edge inference. Option C is correct because Custom Vision is designed for image classification, allowing you to train a model that assigns images to labeled classes such as product types. Option D is correct because Custom Vision also supports object detection, which returns bounding boxes and labels for multiple objects within an image.

Option A is not correct because extracting text from scanned receipts is an OCR task handled by Azure AI Vision (Computer Vision Read API) or Document Intelligence, not Custom Vision. Option E is not correct because identifying specific individuals via facial recognition is provided by Azure AI Face, not Custom Vision.

Exam trap

The trap here is that candidates may confuse Azure Custom Vision's capabilities with other Azure AI services, mistakenly thinking it handles OCR (like Form Recognizer) or facial recognition (like Face API), when Custom Vision is strictly for custom image classification and object detection.

Full explanation →

705

MCQmedium

You are planning to deploy an Azure AI Language resource that will be used by multiple applications. The applications must authenticate using Azure Active Directory (Azure AD) tokens. You need to assign the appropriate role to the applications' managed identities so they can call the Azure AI Language service. Which role should you assign?

A.Reader

B.Azure AI Developer

C.Cognitive Services User

D.Cognitive Services Contributor

AnswerC

The Cognitive Services User role grants access to read and write data for Azure AI services, including calling the Azure AI Language APIs. It allows the managed identity to authenticate and perform operations such as text analytics and language understanding. This role provides the necessary permissions without granting full control over the resource, following the principle of least privilege for application access.

Why this answer

For applications to call Azure AI Language using Azure AD authentication, they need a role that grants data-plane access. The Cognitive Services User role provides read and write access to the service's data plane without granting management permissions. This aligns with least privilege and is the correct choice for managed identities that only need to invoke the API.

Exam trap

The trap here is confusing the Contributor role, which manages the resource, with the User role, which allows calling the service API.

Full explanation →

706

Multi-Selectmedium

You are building an application that uses the Azure OpenAI Assistants API. The assistant must maintain conversational context across multiple user turns and use a code interpreter to analyze uploaded CSV files. Which two actions should you perform? (Choose two.)

Select 2 answers

A.Create a new assistant for each user turn to avoid context conflicts.

B.Create a thread and add user messages to it, then create a run that references the assistant and the thread.

C.Set the 'stream' parameter to true on every run to preserve conversation state between turns.

D.Use the completions endpoint with a manually maintained message array instead of the Assistants API.

E.Enable the code_interpreter tool on the assistant and upload the CSV files as files that the run can access.

AnswersB, E

The Assistants API persists conversation state in a thread. Adding messages to the thread and running the assistant against both the assistant ID and thread ID gives the model access to prior turns, satisfying the multi-turn context requirement without manually resending history in each call.

Why this answer

The Assistants API stores multi-turn context in threads, so messages are appended to a thread and runs execute against the assistant and thread. Code interpreter must be enabled as a tool, and CSV files must be uploaded and attached so the tool can read them during the run. Together these actions satisfy both the context and analysis requirements.

Exam trap

The trap here is treating streaming or per-turn assistant creation as a way to keep context, when conversation state is actually held in threads and tools must be explicitly enabled and given file access.

Full explanation →

707

MCQmedium

You need the project to support English, Spanish, and French. What change should you make to the command?

A.Change --language to "multi".

B.Change --multilingual false to --multilingual true.

C.Add --description "Multi-language support".

D.Change --project-name to "SupportBotML".

AnswerB

Setting `--multilingual true` enables the custom question answering project to train and query across multiple languages within a single project, rather than one language per project. This directly satisfies the stem's requirement for English, Spanish, and French support, since the default `false` restricts the project to a single language.

Why this answer

The command requires multilingual support for English, Spanish, and French. By default, the `--multilingual` flag is set to `false`, which restricts the project to a single language. Changing it to `true` enables the project to accept utterances in multiple languages, allowing the Conversational Language Understanding (CLU) model to process and train on intents and entities across all specified languages.

Exam trap

Azure exams often test the misconception that adding a description or changing the project name can enable multilingual support, when in fact only the explicit `--multilingual true` flag activates this feature.

How to eliminate wrong answers

Option A is wrong because `--language` specifies the primary language of the project (e.g., 'en' for English), not a multilingual mode; setting it to 'multi' is not a valid value and would cause an error. Option C is wrong because `--description` is a metadata field for human-readable notes and has no effect on language support or multilingual capabilities. Option D is wrong because `--project-name` simply renames the project and does not alter any language configuration; the project would still default to single-language mode.

Full explanation →

708

MCQeasy

You are building an agentic solution using Azure AI Agent Service. The agent needs to send an email via Microsoft Graph API. Which authentication method should you use for the action?

A.Client Certificate

B.API Key

C.OAuth 2.0

D.Basic Authentication

AnswerC

OAuth 2.0 provides delegated, scoped access tokens that Microsoft Graph requires for sending mail on a user's behalf. It satisfies the authentication constraint for Graph API actions, unlike API keys or connection strings, which Graph does not accept.

Why this answer

Microsoft Graph API requires OAuth 2.0 for authentication because it uses delegated or application permissions to access user data securely. Azure AI Agent Service can use OAuth 2.0 with a managed identity or service principal to obtain an access token for the Graph API, ensuring proper authorization and compliance with Microsoft's security model.

Exam trap

Azure certification exams often test the misconception that API keys or basic authentication can be used with modern REST APIs like Microsoft Graph, but the trap here is that candidates overlook the mandatory OAuth 2.0 requirement for Microsoft Graph API and the deprecation of basic authentication in Azure services.

How to eliminate wrong answers

Option A is wrong because client certificates are used for authentication in scenarios like mutual TLS or Azure AD app registration with certificate-based credentials, but Microsoft Graph API does not accept client certificates directly for token acquisition; OAuth 2.0 is still required to exchange the certificate for an access token. Option B is wrong because API keys are not supported by Microsoft Graph API; it relies on OAuth 2.0 tokens (Bearer tokens) for authorization, not static keys. Option D is wrong because Basic Authentication sends credentials in plaintext (Base64-encoded) and is deprecated for Microsoft Graph API; it lacks the token-based security and scoped permissions that OAuth 2.0 provides.

Full explanation →

709

Multi-Selectmedium

You are building an Azure AI Search knowledge mining solution over a repository of scanned product manuals. You need to extract structured entities such as product names and part numbers from the OCR text and store them in an index field. (Choose two.)

Select 2 answers

A.Add Microsoft.Skills.Text.LanguageDetectionSkill to the skillset.

B.Define an outputFieldMapping that writes the entity recognition output to an index field.

C.Add Microsoft.Skills.Text.KeyPhraseExtractionSkill to the skillset.

D.Configure the indexer to use a JSON parsing mode.

E.Add Microsoft.Skills.Text.EntityRecognitionSkill to the skillset and set its categories to include the needed entity types.

AnswersB, E

Skill outputs live only in the enrichment tree until they are mapped. An outputFieldMapping connects the entity recognition output node to a field defined in the index. Without this mapping the extracted entities would be discarded after enrichment, so this step is required to make the data queryable.

Why this answer

Extracting structured entities from OCR text requires a skill that performs entity recognition, and the results must be persisted through an output field mapping. The entity recognition skill with configured categories identifies the relevant named entities, while the field mapping writes those values into the index so they can be queried. Skills that only detect language or key phrases do not produce the required structured output.

Exam trap

The trap here is assuming that enriching text with any language skill automatically stores results in the index, when a field mapping is also required.

Full explanation →

710

MCQmedium

You have an Azure AI Search indexer that uses a custom skill hosted in an Azure Function to normalize product codes. The function occasionally returns HTTP 429 responses. You need the indexer to retry these calls automatically without failing the entire indexing run. What should you configure?

A.Set the batchSize property on the indexer to a smaller value.

B.Implement retry logic inside the Azure Function and return a success response after retries.

C.Add a retryPolicy to the custom skill definition in the skillset.

D.Configure the indexer's maxFailedItems and maxFailedItemsPerBatch to tolerate failures.

AnswerB

Because the throttling originates from the custom skill's downstream dependency, the function itself should handle transient failures using retry policies such as exponential backoff. Returning a successful response after internal retries prevents the indexer from seeing 429 errors. This is the supported and reliable way to make custom skills resilient to intermittent throttling.

Why this answer

Azure AI Search does not provide a retryPolicy on custom skills. When a custom skill returns transient errors such as HTTP 429, the correct pattern is to implement retry logic inside the skill implementation, for example using exponential backoff in the Azure Function, so the indexer receives a successful response once the downstream call succeeds.

Exam trap

The trap here is looking for a retryPolicy property on the custom skill, which Azure AI Search does not support; retries must be handled inside the skill code.

Full explanation →

711

MCQhard

You are building a generative AI solution using Azure AI Foundry. The solution must meet compliance requirements that require all model inputs and outputs to be auditable for a minimum of one year. What should you enable?

A.Azure Monitor alerts for unusual activity.

B.Azure Monitor metrics for the Azure AI Foundry resource.

C.Azure Monitor workbooks to visualize usage.

D.Diagnostic settings to capture request and response logs and store them in a storage account.

AnswerD

Diagnostic settings stream request and response logs to a storage account, giving durable, queryable records that satisfy the one-year audit retention requirement. This captures actual model inputs and outputs rather than metrics alone, which cannot reconstruct individual interactions.

Why this answer

Enabling diagnostic settings for the Azure AI Foundry resource allows you to capture detailed request and response logs for all model interactions. By routing these logs to a storage account, you retain the data for the required one-year audit period, meeting compliance needs for full traceability of inputs and outputs.

Exam trap

The trap here is that candidates confuse monitoring features (alerts, metrics, workbooks) with data retention capabilities, assuming any Azure Monitor feature can satisfy audit requirements without understanding that only diagnostic settings provide the raw log capture needed for compliance.

How to eliminate wrong answers

Option A is wrong because Azure Monitor alerts are designed to notify on unusual activity or anomalies, not to provide long-term audit storage of model inputs and outputs. Option B is wrong because Azure Monitor metrics capture aggregated performance data like latency or request counts, not the detailed request/response payloads needed for auditing. Option C is wrong because Azure Monitor workbooks are visualization tools for metrics and logs, not a storage mechanism for raw audit data.

Full explanation →

712

MCQeasy

You need to monitor costs for an Azure AI solution that uses multiple Azure AI services. Which Azure tool should you use to set budgets and receive alerts?

A.Azure Advisor

B.Azure Monitor

C.Azure Service Health

D.Azure Cost Management

AnswerD

Azure Cost Management aggregates spend across all Azure AI services within a subscription or resource group, letting you define budgets scoped to those resources and configure alerts when thresholds are breached. It is the native tool for cost visibility and proactive notification, matching the monitoring requirement.

Why this answer

Azure Cost Management is the dedicated Azure tool for setting budgets, defining cost thresholds, and configuring alerts when spending exceeds those limits. It provides detailed cost analysis, forecasting, and policy enforcement across all Azure services, including AI services like Cognitive Services and Azure Machine Learning.

Exam trap

The trap here is that candidates confuse Azure Advisor's cost recommendations with actual budget management, or they mistakenly think Azure Monitor's alerting capabilities extend to financial cost thresholds rather than just operational metrics.

How to eliminate wrong answers

Option A is wrong because Azure Advisor provides personalized recommendations for cost optimization, security, and performance, but it does not allow you to set budgets or configure cost alerts. Option B is wrong because Azure Monitor collects and analyzes telemetry data (metrics, logs) for application performance and health, not for financial cost tracking or budget management. Option C is wrong because Azure Service Health provides information about service outages, planned maintenance, and health advisories for Azure services, not cost monitoring or budget alerts.

Full explanation →

713

MCQeasy

A company wants to analyze customer reviews to determine whether sentiment is positive, negative, or neutral. The solution must also extract key phrases such as 'great battery life' and 'poor camera quality'. Which Azure AI feature should be used?

A.Azure AI Language - Named Entity Recognition (NER)

B.Azure AI Content Safety

C.Azure AI Language Understanding (LUIS)

D.Azure AI Language - Sentiment Analysis and Key Phrase Extraction

AnswerD

Sentiment analysis returns positive, negative, or neutral labels, while key phrase extraction pulls the salient terms such as 'great battery life'. Both capabilities sit within Azure AI Language, so one resource satisfies the stem's dual requirement without separate services.

Why this answer

Azure AI Language's Sentiment Analysis and Key Phrase Extraction are specifically designed to evaluate text for positive, negative, or neutral sentiment and to extract meaningful phrases like 'great battery life' or 'poor camera quality'. This combined capability directly matches the dual requirement of sentiment classification and key phrase extraction in a single API call, using pre-built models that require no custom training.

Exam trap

The trap here is that candidates often confuse Named Entity Recognition (NER) with key phrase extraction because both extract text, but NER targets predefined entity types (e.g., person, location) while key phrase extraction targets descriptive phrases that are not entities.

How to eliminate wrong answers

Option A is wrong because Named Entity Recognition (NER) identifies entities such as people, organizations, and locations, not sentiment or key phrases like 'great battery life'. Option B is wrong because Azure AI Content Safety detects harmful content (e.g., hate speech, self-harm) and is not designed for sentiment analysis or key phrase extraction. Option C is wrong because LUIS (Language Understanding) is a conversational AI service for intent and entity extraction from utterances, not for general-purpose sentiment analysis or key phrase extraction; it requires custom training and does not natively output sentiment scores or key phrases.

Full explanation →

714

MCQhard

Refer to the exhibit. You are using Azure AI Document Intelligence with a layout model. The pipeline returns an empty tables array even though the document contains tables. The OCR step extracts text correctly. What is the most likely issue?

A.The OCR step is not recognizing table cells.

B.The table extraction step is misconfigured.

C.The output field mapping for tables is missing.

D.The layout extraction step is not correctly identifying table structures.

AnswerD

The layout model performs its own table-structure detection, separate from OCR text extraction. Correct text with an empty tables array means the structure-detection stage is failing to recognise rows, columns and spans, so the table identification step is the faulty component.

Why this answer

The layout model in Azure AI Document Intelligence performs OCR and then uses a layout extraction step to identify structural elements like tables. If the OCR extracts text correctly but the tables array is empty, it indicates that the layout extraction step failed to detect the table boundaries or cell structure, not that OCR missed the text. Option D correctly identifies this as the most likely issue.

Exam trap

The trap here is that candidates assume OCR and table extraction are the same step, but Azure AI Document Intelligence separates text recognition from structural layout analysis, so correct OCR does not guarantee correct table detection.

How to eliminate wrong answers

Option A is wrong because the OCR step extracts text correctly, as stated in the question, so it is recognizing table cells as text; the issue is not with OCR recognition. Option B is wrong because the layout model does not have a separate 'table extraction' configuration that can be misconfigured; table extraction is an inherent part of the layout analysis, and the pipeline is using the standard layout model. Option C is wrong because output field mapping is used for custom extraction models (like prebuilt or custom neural models), not for the layout model, which returns raw structural elements like tables and cells directly in the JSON output without requiring field mapping.

Full explanation →

715

MCQmedium

A company wants to use Azure AI Vision to extract text from scanned documents that contain both printed and handwritten text in multiple languages. The documents are large and can take several minutes to process. The solution must return the extracted text asynchronously. Which Azure AI Vision API should they use?

A.Read API

B.Image Analysis API with the read feature

C.OCR API

D.Document Intelligence prebuilt-read model

AnswerA

The Read API is designed for asynchronous extraction of printed and handwritten text from documents and images. It supports multiple languages, handles large documents, and returns results via an operation that you poll. It is the correct choice for large scanned documents with mixed printed and handwritten text that require asynchronous processing and multi-language support.

Why this answer

The Read API in Azure AI Vision is designed for asynchronous extraction of printed and handwritten text from large documents and supports multiple languages. It returns results via a polling operation, which suits documents that take minutes to process. The OCR API is synchronous and limited for handwriting, while the Document Intelligence prebuilt-read model, though capable, is not an Azure AI Vision API.

Exam trap

The trap here is assuming that the OCR API can handle handwritten text and large documents asynchronously, when it is actually a synchronous, printed-text-focused API.

Full explanation →

716

MCQhard

You are building a multilingual chatbot using Azure AI Language. The chatbot must handle English, Spanish, and French. You need to configure the LUIS (Language Understanding) model to support multiple languages efficiently. What is the best practice?

A.Use the prebuilt multilingual LUIS model and fine-tune it with your intents.

B.Create a separate LUIS app for each language and use the same intents and entities structure.

C.Create a single LUIS app and add utterances in all three languages.

D.Create one LUIS app with language set to 'Multilingual' and add utterances in all languages.

AnswerB

Each language requires its own app with utterances in that language.

Why this answer

LUIS does not support a single multilingual model; each LUIS app is designed for a single primary language. To handle multiple languages efficiently, you must create a separate LUIS app per language, each with the same intent and entity schema, and route user queries to the appropriate app based on the detected language. This ensures optimal language-specific model training and accuracy.

Exam trap

The trap here is that candidates assume a 'Multilingual' setting or a single app with mixed utterances is supported, but LUIS requires separate apps per language for custom models, and the 'Multilingual' option only applies to prebuilt domains.

How to eliminate wrong answers

Option A is wrong because there is no prebuilt multilingual LUIS model; LUIS requires separate apps per language and does not offer a single fine-tunable multilingual base model. Option C is wrong because creating a single LUIS app with utterances in multiple languages violates LUIS's design, which expects all utterances to be in the app's single primary language, leading to poor performance and incorrect predictions. Option D is wrong because the 'Multilingual' setting in LUIS is a legacy feature that only enables language detection for prebuilt domains, not for custom models; it does not allow training a single app with utterances in multiple languages.

Full explanation →

717

MCQhard

A financial services company is building a computer vision solution to automatically extract data from scanned checks. The solution must recognize handwritten amounts, printed account numbers, and signature presence. The company has a large dataset of labeled check images. They need high accuracy and the ability to retrain with new data. Which Azure service should they use?

A.Azure AI Vision OCR with a custom dataset using Custom Vision

B.Azure AI Language with custom entity recognition

C.Azure AI Document Intelligence (Form Recognizer) with a custom model trained on check images

D.Azure AI Vision Image Analysis with a custom model

AnswerC

A custom Document Intelligence model trains on your labelled check images, learning the specific layouts, handwriting and field positions, and supports retraining as new data arrives. This satisfies the high-accuracy and retraining constraints for handwritten amounts, printed account numbers and signature presence.

Why this answer

Azure AI Document Intelligence (Form Recognizer) with a custom model is designed for extracting structured fields from domain-specific documents like checks, and it supports training on your labeled dataset with the ability to retrain as new data arrives. It handles handwriting, printed text, and layout, and can be trained to detect signature presence as a labeled field. This matches the accuracy and retraining requirements.

Exam trap

AI-102 often tests whether candidates confuse Document Intelligence (structured document extraction with custom training) with Vision OCR (generic text extraction) or Custom Vision (image classification), so they pick a service that cannot learn custom fields.

How to eliminate wrong answers

Option A is wrong because Custom Vision does object detection/classification, not text extraction, and Vision OCR alone cannot learn custom field schemas or detect signatures. Option B is wrong because Azure AI Language operates on text, not scanned images, so it cannot read checks. Option D is wrong because Image Analysis with a custom model is for image classification/object detection tasks, not structured document field extraction with handwriting and signature detection.

Full explanation →

718

Multi-Selecthard

Which THREE factors should you consider when selecting a model for a generative AI solution on Azure?

Select 3 answers

A.Cost per token and deployment options.

B.Model capability and modality (text, code, image).

C.Latency and throughput requirements.

D.Number of transformer layers in the model.

E.Training data source and licensing.

AnswersA, B, C

Token pricing directly determines running cost at scale, while deployment options (serverless versus provisioned throughput) govern quota and capacity planning. Both are explicit selection factors for generative AI models on Azure, letting architects balance budget against the throughput the workload demands.

Why this answer

Option A is correct because cost per token and deployment options (such as pay-as-you-go versus provisioned throughput) directly affect the total cost and scalability of a generative AI solution on Azure. Option B is correct because the model's capability and modality determine whether it can handle the required task, such as text generation, code completion, or image creation. Option C is correct because latency and throughput requirements dictate whether the chosen model and deployment type can meet the application's performance and concurrency needs.

Option D is not a primary selection factor because the number of transformer layers is an internal architectural detail that influences capability but is not a decision criterion by itself. Option E is not a primary selection factor because training data source and licensing are legal and compliance considerations, not core factors for selecting a model for a generative AI solution on Azure.

Exam trap

The trap here is that candidates confuse internal model architecture (like transformer layers) with selection criteria, when in fact Azure abstracts those details and you only need to consider cost, capability, latency, and deployment options.

Full explanation →

719

MCQeasy

You need to generate realistic synthetic data for training a machine learning model while ensuring the data does not contain personally identifiable information (PII). Which Azure service should you use?

A.Azure AI Search

B.Azure AI Document Intelligence

C.Azure OpenAI Service

D.Azure AI Language

AnswerC

Azure OpenAI Service can generate synthetic text or data through its language models, and its content filtering and data-handling controls help avoid reproducing PII. This satisfies the requirement to produce realistic training data without embedding personally identifiable information.

Why this answer

Azure OpenAI Service provides access to powerful generative AI models (e.g., GPT-4) that can create realistic synthetic data by learning patterns from training data. Crucially, these models can be configured to avoid memorizing or reproducing PII, and you can apply content filters and data masking to ensure the generated output is free of personally identifiable information.

Exam trap

The trap here is that candidates confuse Azure AI Language's text generation capabilities (e.g., summarization, question answering) with the full generative AI power of Azure OpenAI Service, but Azure AI Language does not offer the same level of flexible, high-fidelity synthetic data generation.

How to eliminate wrong answers

Option A is wrong because Azure AI Search is a search-as-a-service solution for indexing and querying data, not a generative AI service capable of creating synthetic data. Option B is wrong because Azure AI Document Intelligence (formerly Form Recognizer) is designed to extract structured information from documents (e.g., OCR, key-value pairs), not to generate new synthetic datasets. Option D is wrong because Azure AI Language provides pre-built and custom NLP capabilities (e.g., sentiment analysis, entity recognition) but does not include generative models for creating realistic synthetic data from scratch.

Full explanation →

720

MCQhard

Your company uses Azure OpenAI to generate code snippets. Developers need to ensure that the generated code does not contain security vulnerabilities. What should you implement?

A.Set usage quotas to limit the number of code generation requests

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

C.Configure Azure OpenAI content filters to block vulnerable code

D.Integrate a static code analysis tool into the CI/CD pipeline to scan generated code

AnswerD

Static analysis scans generated code for known vulnerability patterns such as injection flaws or insecure API use before merge, catching issues that prompt engineering alone cannot guarantee. Integrating it into CI/CD enforces this check consistently on every generated snippet.

Why this answer

Integrating a static code analysis tool (e.g., Microsoft Defender for DevOps, SonarQube, or Checkmarx) into the CI/CD pipeline allows automated scanning of generated code for security vulnerabilities before deployment. This approach directly addresses the requirement to ensure generated code is free of vulnerabilities, as Azure OpenAI content filters are not designed to detect code-level security flaws like SQL injection or buffer overflows.

Exam trap

The trap here is that candidates confuse Azure OpenAI content filters (which handle text-level safety) with code-level security scanning, leading them to incorrectly select Option C, while the correct approach requires a dedicated security analysis tool integrated into the development pipeline.

How to eliminate wrong answers

Option A is wrong because setting usage quotas only limits the number of requests, not the security quality of the generated code; it prevents abuse but does not scan for vulnerabilities. Option B is wrong because fine-tuning on secure code examples improves the model's output quality but does not guarantee that every generated snippet is vulnerability-free, as the model can still produce insecure patterns not present in the training data. Option C is wrong because Azure OpenAI content filters are designed to block harmful or policy-violating content (e.g., hate speech, violence), not to detect code-specific security vulnerabilities like cross-site scripting or insecure cryptographic practices.

Full explanation →

721

Multi-Selectmedium

Which THREE of the following are capabilities of Azure AI Content Safety?

Select 3 answers

A.Sexual content detection

B.Hate speech detection

C.Self-harm detection

D.Sentiment analysis

E.Personally identifiable information (PII) detection

AnswersA, B, C

Sexual content detection is a core Azure AI Content Safety capability, satisfying the stem's requirement for content moderation features. The service classifies text and images against severity levels for sexual material, alongside hate, violence and self-harm categories, using its dedicated classification models rather than general-purpose filtering.

Why this answer

Azure AI Content Safety provides built-in AI classifiers that detect harmful content across four categories: sexual, hate, violence, and self-harm, so options A (sexual content detection), B (hate speech detection), and C (self-harm detection) are all correct capabilities of the service. These categories are exposed through the Analyze Text and Analyze Image APIs, which return severity scores (0-7) for each category, enabling applications to filter or moderate harmful content. Option D (sentiment analysis) is not part of Content Safety; sentiment analysis is a feature of Azure AI Language.

Option E (PII detection) is also not part of Content Safety; PII detection is provided by Azure AI Language's Personally Identifiable Information extraction capability.

Exam trap

The trap here is that candidates confuse Azure AI Content Safety with other Azure AI services that handle sentiment analysis or PII detection, leading them to select options that belong to Azure AI Language or Azure AI Search instead of the specific content moderation service.

Full explanation →

722

MCQeasy

A company uses Azure AI Search to index customer support transcripts. They want to enable users to find relevant answers by asking natural language questions. Which feature should they enable in the search service?

A.Semantic search

B.Synonym maps

C.Cognitive skills

D.Knowledge mining

AnswerA

Semantic search adds a reranking layer over results using Microsoft's language models, matching natural-language questions to the most relevant passages in the transcripts. This directly satisfies the stem's requirement to find answers by asking questions, since keyword search alone cannot interpret query intent or rank by semantic relevance.

Why this answer

Semantic search in Azure AI Search enhances the ranking of results by using language understanding models to re-rank matches based on semantic relevance to the query, enabling users to ask natural language questions and get more relevant answers. It is the feature designed to improve relevance for natural language queries.

Exam trap

AI-102 often tests the distinction between semantic search and other Azure AI Search features — candidates may confuse semantic search with synonym maps or cognitive skills, which serve different purposes.

How to eliminate wrong answers

Option B is wrong because synonym maps expand queries with equivalent terms but do not provide semantic understanding or natural language question answering. Option C is wrong because cognitive skills are used during indexing to enrich content (e.g., entity recognition, OCR), not to enable natural language querying at search time. Option D is wrong because knowledge mining is a broader solution pattern for extracting insights from large content sets, not a specific search feature for natural language question answering.

Full explanation →

723

MCQmedium

You are building an agent for a legal firm that uses Azure OpenAI to analyze contracts. The agent must extract key clauses, identify risks, and summarize the contract. The agent uses a RAG pattern with Azure Cognitive Search as the vector database. After deployment, the agent sometimes returns irrelevant information or fails to find relevant clauses. You suspect the issue is with the chunking strategy. The contracts are large, typically 50-100 pages. Currently, you are chunking by page (each page is one chunk). You want to improve retrieval accuracy. Which action should you take?

A.Keep page-level chunking but add 50% overlap between chunks.

B.Use a different embedding model, such as text-embedding-3-large.

C.Increase the chunk size to 5 pages per chunk and reduce overlap.

D.Change chunking to use semantic boundaries: split at clause or section headings.

AnswerD

Page-based chunking splits clauses across arbitrary boundaries, so retrieved vectors mix unrelated contract text. Splitting at clause or section headings keeps each chunk semantically coherent, which is the axis that improves retrieval accuracy for the RAG pattern over 50-100 page contracts.

Why this answer

Splitting contracts at semantic boundaries (clause or section headings) preserves the natural meaning and context of each chunk, which is critical for legal document analysis. Page-level chunking often splits a clause across two pages, causing the vector search to retrieve incomplete or irrelevant information. By aligning chunks with the document's logical structure, the RAG pattern retrieves more coherent and relevant passages for the Azure OpenAI agent to process.

Exam trap

The trap here is that candidates often focus on tuning parameters like overlap or chunk size, or switching embedding models, without recognizing that the fundamental issue is the chunking strategy's failure to respect the document's logical structure.

How to eliminate wrong answers

Option A is wrong because adding 50% overlap to page-level chunking still splits clauses at arbitrary page boundaries, and the overlap only partially mitigates the issue without guaranteeing that a complete clause is captured in a single chunk. Option B is wrong because the embedding model is not the root cause; even a better model like text-embedding-3-large cannot fix retrieval accuracy if the chunking strategy destroys semantic coherence. Option C is wrong because increasing chunk size to 5 pages per chunk makes the chunks too large and reduces precision, and reducing overlap further increases the risk of missing relevant content that spans chunk boundaries.

Full explanation →

724

MCQeasy

Your team is developing a chatbot using Azure AI Bot Service. You need to ensure that the bot can handle multiple languages and respond appropriately. Which Azure AI service should you integrate to perform language detection?

A.Azure AI Language

B.Azure AI Speech

C.Azure AI Content Safety

D.Azure AI Translator

AnswerA

Azure AI Language provides the language detection feature, returning the detected language and confidence score for input text. Integrating it lets the bot identify the incoming language and route to appropriate responses, satisfying the multilingual handling requirement.

Why this answer

Azure AI Language provides pre-built language detection capabilities as part of its natural language processing (NLP) features. By integrating this service, the bot can analyze incoming text and identify the language, enabling it to route responses appropriately or trigger language-specific logic.

Exam trap

The trap here is that candidates often confuse Azure AI Translator's built-in language detection (which is a secondary capability) with the dedicated language detection service, leading them to choose Option D instead of the correct Azure AI Language.

How to eliminate wrong answers

Option B is wrong because Azure AI Speech focuses on speech-to-text, text-to-speech, and speaker recognition, not on detecting the language of text input. Option C is wrong because Azure AI Content Safety is designed to detect harmful or inappropriate content (e.g., hate speech, self-harm) in text or images, not to identify the language. Option D is wrong because Azure AI Translator is used to translate text between languages, but it does not perform standalone language detection; while Translator can sometimes infer language during translation, the dedicated language detection feature is part of Azure AI Language.

Full explanation →

725

MCQhard

A financial services firm wants to use Azure OpenAI to generate investment advice summaries. They must ensure that the model does not produce any advice that could be interpreted as personalized financial advice. What is the most effective strategy?

A.Set temperature to 0 and top_p to 0 to make outputs deterministic.

B.Use a system message that instructs the model to avoid personalized advice and apply strict content filtering.

C.Provide few-shot examples of disclaimers in the prompt.

D.Fine-tune the model on a dataset of generic financial summaries.

AnswerB

A system message sets persistent behavioural boundaries, instructing the model to decline personalised financial advice, while content filtering blocks prohibited outputs. Together they satisfy the stem's constraint that no output be interpretable as personalised investment advice.

Why this answer

Azure OpenAI's system messages allow you to set the model's behavior and constraints at the conversation level, which is the most direct and effective way to enforce a policy like avoiding personalized financial advice. Combined with Azure's content filtering (which can block harmful or restricted content), this approach provides both instruction-based and filter-based guardrails without requiring model retraining or relying solely on example-based prompting.

Exam trap

The trap here is that candidates often assume deterministic parameters (temperature=0, top_p=0) guarantee safe outputs, but they only control randomness, not content compliance—Azure's system message and content filtering are the correct tools for enforcing content policies.

How to eliminate wrong answers

Option A is wrong because setting temperature to 0 and top_p to 0 makes outputs deterministic but does not prevent the model from generating personalized financial advice; it only reduces randomness, not content compliance. Option C is wrong because few-shot examples of disclaimers in the prompt can be ignored or overridden by the model if the underlying training data biases it toward personalized responses; system messages have higher priority in the instruction hierarchy. Option D is wrong because fine-tuning on generic financial summaries requires significant labeled data and compute, and it does not guarantee the model will avoid personalized advice—it may still generate such content if the fine-tuning dataset is not carefully curated to exclude it.

Full explanation →

726

Multi-Selectmedium

Your organization needs to analyze customer call transcripts to extract key insights, including sentiment, issues, and resolution. Which THREE Azure AI Language features should you use?

Select 3 answers

A.Custom named entity recognition

B.Sentiment analysis

C.Conversation summarization

D.Key phrase extraction

E.PII detection

AnswersA, B, C

Custom named entity recognition extracts domain-specific entities—such as product names, issue categories, or resolution codes—that your labelled training data defines, satisfying the requirement to pull structured issues and resolutions from transcripts. It complements sentiment analysis and key phrase extraction, the other two features needed for the full insight set.

Why this answer

Sentiment analysis (B) is correct because Azure AI Language's sentiment analysis feature evaluates text and returns sentiment labels and confidence scores, which directly satisfies the requirement to extract sentiment from customer call transcripts. Conversation summarization (C) is correct because it is designed to summarize conversations and extract key information such as issues and resolutions from call transcripts, matching the need to identify issues and resolution outcomes. Custom named entity recognition (A) is correct because it lets you train a model to extract domain-specific entities (for example, product names, issue categories, or resolution codes) from transcripts, which supports extracting key insights beyond generic entities.

Key phrase extraction (D) is not marked correct because, while it surfaces main talking points, it does not specifically deliver sentiment, issue, or resolution extraction as required. PII detection (E) is not marked correct because it only identifies and redacts personally identifiable information and does not provide sentiment, issue, or resolution insights.

Exam trap

A common pitfall in the AI-102 exam is confusing pre-built features like key phrase extraction with customizable features like custom named entity recognition (NER). While key phrase extraction works for general keywords, custom NER is required to extract domain-specific entities tailored to the organization's needs, such as issue types and resolution steps from call transcripts.

Full explanation →

727

MCQmedium

You are developing a chat application that uses Azure OpenAI GPT-4 to answer customer questions. You need to ensure the model does not generate harmful content. Which configuration should you set?

A.Use a system prompt that instructs the model to be safe.

B.Set the temperature parameter to 0.

C.Set max_tokens to a low value.

D.Enable the content filter in Azure OpenAI Service.

AnswerD

Azure OpenAI Service content filters evaluate both prompts and completions against harm categories (hate, violence, sexual, self-harm) and block or annotate flagged content. Enabling the filter satisfies the requirement to prevent the GPT-4 model from generating harmful output.

Why this answer

Azure OpenAI Service includes a built-in content filter that actively scans both input prompts and generated completions to detect and block harmful content such as hate speech, violence, or self-harm. This filter operates at the service level, providing a robust safety layer that cannot be bypassed by model configuration alone. While system prompts can guide behavior, they are not a reliable safeguard against adversarial inputs or model misuse.

Exam trap

The trap here is that candidates assume a system prompt or parameter tuning (temperature, max_tokens) can guarantee safety, but Azure OpenAI's content filter is the only mechanism that actively blocks harmful content at the service level, regardless of model configuration.

How to eliminate wrong answers

Option A is wrong because a system prompt is merely a text instruction and can be overridden by user prompts or jailbreak attempts; it does not enforce content safety at the API or network level. Option B is wrong because setting temperature to 0 only makes the model more deterministic and less creative, but it does not prevent the generation of harmful content if the model's training data includes such patterns. Option C is wrong because max_tokens controls the length of the response, not its safety; a short response can still contain harmful content.

Full explanation →

728

Multi-Selecthard

You are planning to deploy a custom neural voice (CNV) model using Azure AI Speech. You need to ensure that the deployment meets Microsoft's responsible AI requirements. Which two actions should you take? (Choose two.)

Select 2 answers

A.Submit an application to Microsoft for access to custom neural voice and wait for approval.

B.Deploy the model to a public endpoint without any access restrictions.

C.Train the model using data from publicly available audio sources without consent.

D.Obtain explicit consent from the voice talent and store the consent statement.

E.Use a standard voice model instead of a custom neural voice to avoid the approval process.

AnswersA, D

Access to custom neural voice is restricted. You must submit an application to Microsoft that describes your use case and how you will comply with responsible AI guidelines. Microsoft reviews the application and grants access only if approved. This is a mandatory step before you can create and deploy a CNV model. It ensures that the technology is used responsibly.

Why this answer

To deploy a custom neural voice model responsibly, you must obtain explicit consent from the voice talent and submit an application to Microsoft for access. These are mandatory steps in the gated access process. Using standard voices or training without consent does not meet the requirements.

Deploying to a public endpoint without restrictions is also against policy.

Exam trap

The trap here is assuming that custom neural voice can be deployed without Microsoft approval, when in fact it requires a gated access application and consent.

Full explanation →

729

MCQhard

Your organization uses Azure AI Document Intelligence to extract data from invoices. The solution must identify custom fields not present in the prebuilt models, such as 'purchase order number' located in varying positions across documents. What should you do?

A.Use the layout model and apply manual post-processing.

B.Use Azure AI Forms Recognizer with prebuilt receipt model.

C.Use the prebuilt invoice model with field merging.

D.Train a custom extraction model using labeled sample invoices.

AnswerD

Custom extraction models learn field labels and their positional context from your own labelled invoices, so they locate fields such as purchase order number wherever they appear. Prebuilt invoice models expose only a fixed schema, which cannot capture organisation-specific fields in varying positions.

Why this answer

Azure AI Document Intelligence (formerly Form Recognizer) allows you to train a custom extraction model using labeled sample invoices. This approach enables the model to learn custom fields like 'purchase order number' that appear in varying positions, which prebuilt models cannot handle. By providing labeled examples, the model generalizes to extract the field accurately from new documents.

Exam trap

The trap here is that candidates may assume the prebuilt invoice model can be extended with custom fields via configuration or merging, but Azure AI Document Intelligence requires explicit custom model training to recognize fields not present in prebuilt schemas.

How to eliminate wrong answers

Option A is wrong because the layout model only extracts text and structure (tables, lines) without semantic field recognition; manual post-processing would be inefficient and error-prone for custom fields. Option B is wrong because the prebuilt receipt model is designed for receipts, not invoices, and cannot extract custom fields like 'purchase order number'. Option C is wrong because the prebuilt invoice model does not support field merging; it only extracts predefined fields and cannot learn new custom fields.

Full explanation →

730

Multi-Selecteasy

Which TWO monitoring metrics should you track to ensure the health and performance of an Azure AI Search service used for a customer-facing product catalog?

Select 2 answers

A.Throttled search queries count.

B.Indexer execution history and duration.

C.Storage used in GB.

D.Search latency (average and P99).

E.Number of successful search requests.

AnswersA, D

Throttled search queries count directly exposes capacity exhaustion, revealing when the service rejects requests because replica or partition limits are exceeded. For a customer-facing catalogue, this metric satisfies the availability constraint: throttling silently degrades the user experience, so tracking it triggers timely scaling before shoppers encounter failed searches.

Why this answer

Option A, throttled search queries count, is correct because throttling directly indicates that the service is hitting its query-per-second (QPS) capacity limits, which degrades the customer-facing catalog experience and signals a need to scale replicas or partitions. Option D, search latency (average and P99), is correct because latency is the primary performance indicator for a user-facing search experience; tracking both average and P99 reveals tail-latency problems that average alone would hide. Option B is not among the correct answers because indexer execution history and duration relates to data ingestion pipelines, not the query-time health of a customer-facing catalog.

Option C is not correct because storage used in GB is a capacity metric that does not reflect query performance or service health for end users. Option E is not correct because counting only successful search requests gives no insight into failures, throttling, or latency, and a rising success count can mask underlying performance degradation.

Exam trap

The trap here is that candidates often confuse operational metrics (like indexer duration or storage usage) with customer-facing performance metrics, leading them to select indexer execution history instead of search latency.

Full explanation →

731

MCQeasy

A developer is building a mobile app that uses Azure AI Vision to generate a descriptive caption for user-uploaded photos. The app must return a human-readable sentence describing the main content of each image. Which Image Analysis feature should the developer use?

A.Tags

B.Objects

C.Read

D.Caption

AnswerD

The Caption feature in Image Analysis generates a single, human-readable sentence that describes the main content of an image, such as 'a person riding a bike on a beach'. It is specifically designed for this purpose and returns a confidence score for the caption. This directly meets the requirement of producing a descriptive sentence for each photo.

Why this answer

The Caption feature of Image Analysis is purpose-built to generate a one-sentence description of an image's main content. Tags, Objects, and Read serve different purposes: tags provide keywords, objects give bounding boxes, and Read extracts text. Only Caption produces the natural language sentence needed for the mobile app.

Exam trap

The trap here is assuming that tags or objects can be concatenated into a sentence, but only the Caption feature is designed to output a fluent description.

Full explanation →

732

Multi-Selecthard

Which THREE are required when planning to use Azure OpenAI Service for a generative AI application that must comply with responsible AI principles?

Select 3 answers

A.Restrict the model to a maximum of 1000 tokens.

B.Implement content filters to block harmful outputs.

C.Design with human-in-the-loop for critical decisions.

D.Enable rate limiting to prevent abuse.

E.Establish data governance policies for training data.

AnswersB, C, E

Content filters are mandatory for Azure OpenAI deployments, screening prompts and completions to block harmful categories such as violence, hate, and self-harm. Implementing them is a required responsible AI control for any generative application handling user input and model output.

Why this answer

Option B is correct because Azure OpenAI Service provides configurable content filters (categories such as hate, violence, sexual, and self-harm, with severity thresholds) that are a core responsible AI control for blocking harmful model outputs. Option C is correct because human-in-the-loop review is a responsible AI requirement for high-impact or critical decisions, ensuring a human validates or overrides model output before it affects users. Option E is correct because data governance policies for training data address privacy, consent, provenance, and bias concerns, which are foundational to responsible AI compliance.

Option A is not required: a 1000-token cap is an arbitrary cost/latency constraint, not a responsible AI principle, and token limits are set per model/deployment as needed. Option D is not required: rate limiting is an availability and abuse-mitigation control, not a responsible AI requirement, and it does not by itself ensure fair, safe, or accountable AI behavior.

Exam trap

The trap here is that candidates confuse operational controls (like token limits or rate limiting) with responsible AI requirements, which are specifically about fairness, safety, transparency, and accountability, not performance or security.

Full explanation →

733

MCQmedium

You are configuring an Azure OpenAI deployment for a generative AI solution that summarizes long legal contracts. Users report that summaries sometimes omit clauses near the end of documents. The documents are up to 120 pages. You need to improve completeness without changing the model. What should you do?

A.Increase the max_tokens parameter of the completion request to its maximum value.

B.Enable streaming responses so the client receives partial summaries as they are generated.

C.Split each document into overlapping chunks, summarize each chunk, and then combine the summaries in a final aggregation step.

D.Raise the temperature so the model explores more of the document content.

AnswerC

Chunking with overlap ensures content near boundaries is not lost, and a map-reduce style aggregation summarizes each part before combining them. This keeps each request within the context window while covering the entire document, which directly addresses omitted clauses near the end without changing the model.

Why this answer

Long documents can exceed the model context window, causing later sections to be dropped. Chunking with overlap preserves boundary content, and summarizing chunks before aggregating covers the whole document. Output length, temperature, and streaming do not affect how much input text the model can attend to, so they cannot fix missing clauses.

Exam trap

The trap here is assuming that a larger output token limit lets the model consider more of the input document.

Full explanation →

734

MCQmedium

You are deploying an Azure AI solution that calls Azure OpenAI and Azure AI Language from an Azure Container Apps environment. The solution must authenticate to both services without using service keys, and the container app must be able to access the services even if the network is restricted to private endpoints. What should you configure?

A.Create a service principal, store its client secret in the container app environment variables, and assign the Contributor role on each Azure AI resource.

B.Use the Azure AI services multi-service account key retrieved from Azure Key Vault and pass it in the Ocp-Apim-Subscription-Key header.

C.Configure the container app to use an Azure Front Door origin with a private link and enable token-based authentication by using a shared access signature.

D.Enable a system-assigned managed identity on the container app and assign the Cognitive Services User role to the identity on each Azure AI resource.

AnswerD

A system-assigned managed identity gives the container app an identity in Microsoft Entra ID without any secret. Assigning the Cognitive Services User role on each Azure AI resource grants the data-plane permissions needed to call Azure OpenAI and Azure AI Language. This works with private endpoints because authentication uses the identity rather than keys.

Why this answer

Identity-based authentication with a managed identity and the Cognitive Services User role allows the container app to call Azure OpenAI and Azure AI Language without any keys. Because the identity is recognized by Microsoft Entra ID, the calls are authorized even when the services are reachable only through private endpoints. The other options either use keys or assign roles that do not grant data-plane access.

Exam trap

The trap here is assigning a management-plane role such as Contributor and expecting it to authorize data-plane inference calls, which require a Cognitive Services data role instead.

Full explanation →

735

MCQmedium

Refer to the exhibit. You are configuring an agent in Azure AI Foundry. The agent fails to start because the specified model is not available in the current Azure OpenAI resource. What should you do to resolve the issue?

A.Modify the system_prompt to include the model version

B.Deploy the gpt-4-0613 model in the Azure OpenAI resource

C.Change the connection_type to 'Weak'

D.Change the provider to 'AzureAI'

AnswerB

The agent requires a deployed model matching the specified name in the Azure OpenAI resource. Deploying gpt-4-0613 creates that deployment, making the model available for the agent to reference and allowing it to start successfully.

Why this answer

The agent fails to start because the specified model (likely gpt-4-0613) is not deployed in the Azure OpenAI resource. In Azure AI Foundry, agents require an existing model deployment to invoke; you cannot use a model that hasn't been deployed. Option B correctly resolves this by deploying the required model in the Azure OpenAI resource.

Exam trap

The trap here is that candidates might think modifying the system_prompt or changing a connection setting can fix a missing model deployment, but Azure OpenAI requires explicit model deployment before any resource can use it.

How to eliminate wrong answers

Option A is wrong because the system_prompt defines the agent's behavior and instructions, not the model version or deployment; modifying it cannot make an undeployed model available. Option C is wrong because connection_type is not a valid configuration for Azure OpenAI resources; 'Weak' is not a recognized connection type and does not affect model availability. Option D is wrong because the provider is already Azure (Azure OpenAI) and changing it to 'AzureAI' is not a valid provider option; the issue is the missing model deployment, not the provider.

Full explanation →

736

MCQeasy

You are building an Azure AI solution that uses Azure AI Vision to analyze images. The solution must be able to extract text from images and return the text in a structured format. You need to choose the appropriate Azure AI Vision feature. What should you use?

A.Image Analysis

B.Face API

C.Read API

D.Custom Vision

AnswerC

The Read API in Azure AI Vision is specifically designed for optical character recognition (OCR). It extracts printed and handwritten text from images and documents and returns the text in a structured format, including lines and words. This meets the requirement to extract text and return it in a structured format.

Why this answer

The Read API is the OCR component of Azure AI Vision. It extracts text from images and returns it in a structured JSON format with lines and words. This is the correct choice for extracting text from images in a structured format.

Exam trap

The trap here is confusing Image Analysis with OCR capabilities, as Image Analysis can detect text but does not provide the same structured output as the Read API.

Full explanation →

737

MCQeasy

You are using Azure AI Language Service to extract key phrases from customer reviews. You notice that for reviews containing the word 'not good', the service sometimes extracts 'good' as a key phrase. What is the most likely reason?

A.The language detection model misidentified the language

B.You need to set a confidence threshold to exclude negative phrases

C.Key phrase extraction does not consider negation

D.The service is not trained on your specific domain

AnswerC

Key phrase extraction identifies statistically significant terms without parsing negation, so 'good' is extracted as a standalone phrase from 'not good'. The model treats words independently rather than understanding that the negation inverts the sentiment.

Why this answer

Key phrase extraction in Azure AI Language Service uses a statistical model that identifies significant terms based on frequency and context, but it does not inherently understand negation. When the phrase 'not good' appears, the model may still extract 'good' as a key phrase because it recognizes 'good' as a high-value term, ignoring the negation. This is a known limitation of the feature, as it focuses on noun phrases and important terms rather than sentiment or negated constructs.

Exam trap

The trap here is that candidates often assume Azure AI Language Service handles negation across all features, but key phrase extraction explicitly does not consider negation, unlike sentiment analysis which does.

How to eliminate wrong answers

Option A is wrong because language detection is a separate step that identifies the language of the text; misidentification would cause incorrect processing but would not specifically cause 'good' to be extracted from 'not good'. Option B is wrong because confidence thresholds filter out low-confidence phrases, not negative phrases; the service does not have a built-in mechanism to exclude negated terms via threshold settings. Option D is wrong because while domain-specific training can improve accuracy, the core issue here is a fundamental limitation of the key phrase extraction model's handling of negation, not a lack of domain adaptation.

Full explanation →

738

MCQeasy

You are designing an agentic solution that uses Microsoft Copilot Studio and Azure AI Search. The agent needs to answer questions based on confidential documents. Which security measure should you implement to ensure the agent only accesses documents the user has permission to read?

A.Disable public network access on the Azure AI Search service.

B.Implement document-level security using security filters in the search index.

C.Use a managed identity for the agent to access the search index.

D.Require multi-factor authentication for all users.

AnswerB

Security filters in the Azure AI Search index apply the user's identity at query time, trimming results to documents that identity may read. This enforces document-level permissions, satisfying the constraint that the agent must only surface confidential documents the requesting user is authorised to access.

Why this answer

Azure AI Search supports document-level security through security filters, which allow you to restrict search results based on the user's identity. By storing security identifiers (e.g., group memberships or user IDs) as a field in the index and applying an OData filter at query time, the agent can ensure users only see documents they are permitted to read. This is the standard approach for implementing row-level security in Azure AI Search.

Exam trap

The trap here is confusing authentication (verifying who the user is) with authorization (determining what the user can access), leading candidates to select network controls or MFA instead of the document-level security filter mechanism.

How to eliminate wrong answers

Option A is wrong because disabling public network access on the Azure AI Search service controls network-level access to the service itself, not document-level permissions within the index; it does not differentiate between users or documents. Option C is wrong because using a managed identity for the agent authenticates the agent to the search service, but does not enforce per-document access control; the agent would have full access to all indexed documents regardless of the end user's permissions. Option D is wrong because requiring multi-factor authentication for all users strengthens authentication but does not restrict which documents a user can see after they are authenticated; it addresses identity verification, not authorization at the document level.

Full explanation →

739

MCQmedium

You are using Azure AI Search to build a knowledge base for a customer support portal. The index includes a 'sentiment' field that should be populated using the Sentiment skill. However, the sentiment scores are not being written to the index. The skillset runs successfully. What is the most likely cause?

A.The output field mapping for 'sentiment' is missing or incorrectly defined in the indexer.

B.The Sentiment skill is not correctly configured in the skillset.

C.The indexer is in a failed state and not processing documents.

D.The sentiment field in the index is of type 'Collection(Edm.String)' but the skill outputs a double.

AnswerA

Skillset execution writes enriched values into the enrichment tree, not the index. Without an output field mapping linking the sentiment skill output to the target index field, scores are discarded, so the index remains empty despite successful skillset execution.

Why this answer

The Sentiment skill outputs a 'double' value for sentiment score, but the indexer requires an explicit output field mapping to write that value into the index's 'sentiment' field. Even when a skillset runs successfully, without a correct output field mapping in the indexer definition, the skill's output is not transferred to the index. The indexer's field mappings control how enriched data flows from the skillset's output nodes to the index fields.

Exam trap

The trap here is that candidates assume a successful skillset execution guarantees data is written to the index, but Azure AI Search requires explicit output field mappings in the indexer to bridge skill outputs to index fields, and this step is often overlooked.

How to eliminate wrong answers

Option B is wrong because the question states the skillset runs successfully, meaning the Sentiment skill itself is correctly configured and executed without errors. Option C is wrong because the indexer is explicitly described as running successfully, not in a failed state, so it is processing documents. Option D is wrong because the Sentiment skill outputs a double (a numeric score between 0 and 1), and if the index field were of type 'Collection(Edm.String)', the mismatch would cause an indexer error or warning, but the question says the skillset runs successfully — the issue is the missing mapping, not a type conflict.

Full explanation →

740

MCQmedium

You are designing a chatbot using Azure AI Language. The chatbot must understand user intents and also extract entities like dates and locations. Which feature combination should you use?

A.Conversational Language Understanding (CLU) with entities

B.Sentiment analysis and entity linking

C.Custom text classification and key phrase extraction

D.Orchestration Workflow and custom text classification

AnswerA

Conversational Language Understanding provides intent classification alongside integrated entity extraction, satisfying both requirements in one Azure AI Language resource. Custom entities capture dates and locations through labelled training utterances, unlike sentiment analysis or key phrase extraction, which return no intent predictions. CLU therefore meets the stem's dual constraint without combining separate services.

Why this answer

Conversational Language Understanding (CLU) is the correct Azure AI Language feature for building a chatbot that understands user intents and extracts entities like dates and locations. CLU is specifically designed for natural language understanding (NLU) tasks, providing prebuilt and custom entity extraction alongside intent recognition, which directly matches the requirement.

Exam trap

The trap here is that candidates often confuse entity linking (which maps to external knowledge bases) with entity extraction (which pulls values directly from the utterance), leading them to choose Option B despite it lacking intent recognition.

How to eliminate wrong answers

Option B is wrong because sentiment analysis evaluates the emotional tone of text, not user intents, and entity linking maps named entities to a knowledge base (e.g., Wikipedia), not extracting arbitrary entities like dates and locations. Option C is wrong because custom text classification assigns predefined labels to entire documents, not user intents in a conversational context, and key phrase extraction identifies key terms but does not extract structured entities like dates and locations. Option D is wrong because Orchestration Workflow routes requests between different language services (e.g., CLU, QnA Maker) but does not itself perform intent recognition or entity extraction; custom text classification also does not handle entity extraction.

Full explanation →

741

MCQhard

You are deploying an Azure AI solution that uses Azure AI Document Intelligence to extract data from invoices. The solution must process documents in near real-time and must be able to handle sudden spikes in volume. You need to design the architecture to meet these requirements while minimizing cost. What should you use?

A.Azure Logic Apps with a recurrence trigger that polls the blob container every minute.

B.Azure Functions with a blob trigger that calls the Document Intelligence API and uses a consumption plan.

C.Azure Kubernetes Service (AKS) with a horizontal pod autoscaler that processes documents from a queue.

D.Azure Batch with a pool of virtual machines that processes documents from a queue.

AnswerB

Azure Functions with a blob trigger can process documents as they are uploaded, providing near real-time processing. The consumption plan automatically scales out during spikes and scales in when idle, minimizing cost. This serverless approach is ideal for unpredictable workloads and reduces infrastructure management.

Why this answer

For near real-time processing with sudden spikes and minimal cost, a serverless approach with Azure Functions on a consumption plan is best. The blob trigger ensures immediate processing when documents are uploaded, and the consumption plan scales automatically and charges only for execution time. Other options introduce latency, require infrastructure management, or do not scale to zero.

Exam trap

The trap here is assuming that Logic Apps are always cheaper, but polling and scaling limits can increase cost and delay.

Full explanation →

742

MCQmedium

A legal team wants an assistant that drafts contract summaries. Their policy requires that every generated summary include traceable references to the exact clauses used and that reviewers be able to see which source passages informed each statement. You are using Azure OpenAI with your own document index. Which approach best meets the traceability requirement?

A.Fine-tune the base model on previously approved contract summaries so its style matches legal expectations.

B.Enable the model's logprobs parameter and expose the token probabilities to reviewers as evidence of reliability.

C.Increase the model deployment's tokens-per-minute quota so longer contracts fit in a single prompt.

D.Use the On Your Data pattern with Azure AI Search, return document chunks with their IDs and titles, and instruct the model to cite the retrieved chunk identifiers in its output.

AnswerD

Grounding the completion in Azure AI Search chunks and returning citations lets the model reference the exact retrieved passages, and the service can return the citation metadata alongside the response. Reviewers can then map each statement to a chunk and its source clause, which directly satisfies the traceability policy without retraining.

Why this answer

Traceability requires linking generated text back to specific source passages. Using Azure AI Search as the grounding source with the On Your Data pattern returns citation metadata for the retrieved chunks, and prompting the model to cite those chunks produces summaries where each statement can be traced to a clause. Quota, fine-tuning, and logprobs do not create that link.

Exam trap

The trap here is confusing model confidence signals such as logprobs with source attribution, when only retrieved citation metadata can identify the clause behind a statement.

Full explanation →

743

MCQhard

You are deploying an Azure AI solution that uses Azure OpenAI Service. The solution must ensure that all API calls are logged for auditing and that the logs are retained for 90 days. You need to configure diagnostic settings. What should you do?

A.Enable diagnostic settings on the Azure OpenAI resource and send logs to a Log Analytics workspace with a 90-day retention policy.

B.Enable Azure Defender for AI and configure it to export logs to a SIEM with 90-day retention.

C.Configure Azure Monitor Application Insights to capture all API calls and set the retention to 90 days.

D.Use Azure Policy to enforce that all API calls are logged to an Azure Storage account with a 90-day lifecycle policy.

AnswerA

Azure OpenAI supports diagnostic settings that can stream logs to Log Analytics, Storage, or Event Hubs. Sending logs to a Log Analytics workspace allows you to set a retention policy of 90 days, meeting the auditing requirement. This is the native and recommended approach for logging API calls.

Why this answer

To log Azure OpenAI API calls for auditing, you must enable diagnostic settings on the Azure OpenAI resource. These logs can be sent to a Log Analytics workspace, where you can configure a 90-day retention period. Other options do not provide resource-level API logging.

Application Insights is for application telemetry, Azure Policy enforces configurations, and Defender for AI is for security alerts.

Exam trap

The trap here is confusing security alerting with audit logging; Defender for AI does not capture every API call.

Full explanation →

744

MCQmedium

You are developing an Azure AI Search solution that indexes scanned PDF invoices. The indexer must extract text from the PDFs and also recognize entities such as organization names and dates. You want to use built-in cognitive skills to minimize custom code. Which combination of skills should you include in the skillset?

A.Key Phrase Extraction skill and Language Detection skill

B.OCR skill and Entity Recognition skill

C.Text Merge skill and Text Split skill

D.Image Analysis skill and Sentiment skill

AnswerB

The OCR skill extracts text from image-based PDF content, and the Entity Recognition skill identifies entities like organizations and dates from that text. This combination directly addresses the requirement to extract text and recognize entities without writing custom code, leveraging built-in cognitive skills in Azure AI Search.

Why this answer

The OCR skill extracts text from scanned PDFs, and the Entity Recognition skill identifies organizations and dates from the extracted text. Using these built-in skills avoids custom code and satisfies both requirements efficiently. Other combinations either lack OCR or entity recognition, or provide unrelated functionality.

Exam trap

The trap here is assuming that Image Analysis performs OCR, when in Azure AI Search the dedicated OCR skill is needed for text extraction from images.

Full explanation →

745

MCQmedium

You are a solution architect at a legal firm. The firm wants to build a copilot using Microsoft Foundry that answers questions about case law documents stored in Azure Blob Storage. The copilot should use the Retrieval Augmented Generation (RAG) pattern with Azure AI Search as the vector store. The documents are in PDF format and include complex tables and footnotes. The solution must ensure that the answers are grounded in the documents and that the copilot can handle follow-up questions. You need to design the ingestion pipeline. Which approach should you take?

A.Use Azure AI Vision OCR to extract text, split by page, and use Azure AI Search keyword search

B.Use Azure AI Document Intelligence prebuilt-read model, chunk by character count, and use Azure AI Search with semantic ranking

C.Use Azure AI Document Intelligence to extract content, then chunk by headings and paragraphs, generate embeddings using Azure OpenAI, and index in Azure AI Search with vector search

D.Use Azure AI Language to extract key phrases, create a non-vector index, and use simple search

AnswerC

Document Intelligence's layout model preserves complex tables and footnotes that plain PDF text extraction loses, satisfying the grounding requirement. Heading and paragraph chunking keeps semantic units intact for retrieval, and embeddings indexed in Azure AI Search with vector search let the copilot retrieve relevant passages for follow-up questions.

Why this answer

It uses Azure AI Document Intelligence to accurately extract content from PDFs (including complex tables and footnotes), then chunks by headings and paragraphs to preserve document structure, generates embeddings via Azure OpenAI for semantic understanding, and indexes in Azure AI Search with vector search to enable RAG-based, grounded answers with follow-up support.

Exam trap

Microsoft often tests the misconception that simple OCR or keyword search is sufficient for complex documents, but the trap here is that legal documents with tables and footnotes require structure-aware extraction and vector search to support grounded, conversational RAG.

How to eliminate wrong answers

Option A is wrong because Azure AI Vision OCR is designed for image-based text extraction and lacks the ability to handle complex tables and footnotes in PDFs; splitting by page ignores document structure, and keyword search alone cannot support semantic understanding or follow-up questions. Option B is wrong because the prebuilt-read model extracts raw text without preserving table/footnote structure, chunking by character count breaks logical content boundaries, and semantic ranking on keyword search does not provide the vector-based retrieval needed for RAG. Option D is wrong because key phrase extraction loses document context and structure, a non-vector index cannot support semantic similarity search, and simple search cannot ground answers in document content or handle follow-up questions effectively.

Full explanation →

746

MCQeasy

A developer is tasked with integrating Azure OpenAI Service into an application that generates product descriptions. The developer needs to ensure that the generated content does not contain offensive language. Which Azure AI service should be used in addition to Azure OpenAI?

A.Azure AI Search

B.Azure AI Vision

C.Azure AI Language

D.Azure AI Content Safety

AnswerD

Azure AI Content Safety provides dedicated hate, violence, sexual and self-harm classifiers that screen generated text after Azure OpenAI produces it, satisfying the requirement to block offensive language that the model's own filters may not catch.

Why this answer

Azure AI Content Safety (D) is the correct service because it provides built-in content moderation capabilities that can detect and filter offensive, inappropriate, or harmful language in text and images. By integrating Azure AI Content Safety with Azure OpenAI, the developer can automatically screen generated product descriptions for profanity, hate speech, or other offensive content before they are displayed to users, ensuring compliance with content policies.

Exam trap

The trap here is that candidates may confuse Azure AI Language's text analytics features (like sentiment analysis) with content moderation, but Azure AI Language does not include dedicated offensive language filtering, which is a distinct capability of Azure AI Content Safety.

How to eliminate wrong answers

Option A is wrong because Azure AI Search is a cognitive search service used for indexing and retrieving data, not for content moderation or filtering offensive language. Option B is wrong because Azure AI Vision is designed for image analysis tasks such as object detection, OCR, and facial recognition, and does not include text-based content safety features. Option C is wrong because Azure AI Language provides natural language processing capabilities like sentiment analysis, key phrase extraction, and language understanding, but it does not offer dedicated content moderation or offensive language detection; that functionality is specifically handled by Azure AI Content Safety.

Full explanation →

747

Multi-Selecteasy

You are using Azure AI Language to analyze social media comments. You need to identify the language of each comment and then extract key phrases. Which TWO features should you use? (Select TWO.)

Select 2 answers

A.Sentiment analysis

B.Summarization

C.Language detection

D.Entity recognition

E.Key phrase extraction

AnswersC, E

Language detection returns the detected language name and ISO code for each comment, which is needed before any language-specific processing. It directly satisfies the first requirement of identifying the language of each social media comment.

Why this answer

Language detection is the correct feature because it identifies the language of each comment, which is a prerequisite for further analysis. Key phrase extraction is the second correct feature because it extracts important terms from the text, directly addressing the requirement to 'extract key phrases' after language identification.

Exam trap

Microsoft Azure AI Language often tests the distinction between features that analyze content (sentiment, entities, key phrases) versus those that identify metadata (language), and the trap here is that candidates might confuse 'key phrase extraction' with 'entity recognition' because both extract terms, but key phrases are broader and not limited to named entities.

Full explanation →

748

MCQmedium

Your organization is using Azure AI Search to index a large collection of PDF documents stored in Azure Blob Storage. The index currently returns search results, but users complain that the results are not relevant when they search using natural language phrases. You need to improve the relevance of search results without rewriting the application. What should you do?

A.Increase the number of replicas for the search service to improve query performance.

B.Create a new index with a blob indexer that uses the 'content' field only.

C.Enable semantic search on the index and configure a semantic configuration.

D.Configure a custom analyzer on the index to handle stop words and synonyms.

AnswerC

Enabling semantic search and defining a semantic configuration adds L2 reranking and captions over the existing index, so natural-language phrase queries return more relevant results. No application rewrite is needed because the same query endpoint is used, just with the semantic query type.

Why this answer

Semantic search in Azure AI Search uses advanced language models to understand the intent behind natural language queries, re-ranking results based on semantic relevance rather than just keyword matching. Enabling semantic search and configuring a semantic configuration directly addresses the user complaint about poor relevance for natural language phrases without requiring application changes.

Exam trap

The trap here is that candidates often confuse improving query performance (replicas) or basic text processing (custom analyzers) with the semantic understanding needed for natural language queries, leading them to pick options that address performance or tokenization rather than relevance.

How to eliminate wrong answers

Option A is wrong because increasing replicas only improves query throughput and availability, not the relevance or semantic understanding of search results. Option B is wrong because creating a new index with only the 'content' field would reduce the available data for matching, likely worsening relevance rather than improving it. Option D is wrong because custom analyzers handle tokenization, stop words, and synonyms at indexing time, but they do not provide the deep semantic understanding needed to interpret natural language phrases; semantic search is required for that.

Full explanation →

749

Multi-Selectmedium

Which TWO Azure AI services can be used to build a multilingual question-answering bot that retrieves answers from a knowledge base of documents?

Select 2 answers

A.Azure AI Language Understanding (LUIS)

B.Azure OpenAI Service with a RAG pattern

C.Azure AI Translator

D.Azure AI Document Intelligence

E.Azure AI Language - Custom Question Answering

AnswersB, E

Azure OpenAI Service with retrieval-augmented generation embeds documents in a search index, retrieves relevant chunks per query, and prompts the model to answer in the user's language. This satisfies the multilingual requirement because the underlying model handles translation and generation natively.

Why this answer

Azure OpenAI Service with a RAG (Retrieval-Augmented Generation) pattern is correct because it combines a large language model with a retrieval layer that fetches relevant passages from a document knowledge base (e.g., via Azure AI Search) and generates grounded, multilingual answers. Azure AI Language - Custom Question Answering is correct because it is purpose-built to create a knowledge base from documents and FAQs and return precise answers to natural-language questions, including multilingual support. Azure AI Language Understanding (LUIS) is not correct because it only performs intent and entity extraction, not document retrieval or answer generation.

Azure AI Translator is not correct because it only translates text between languages and does not retrieve or answer from a knowledge base. Azure AI Document Intelligence is not correct because it only extracts structured data (text, tables, key-value pairs) from documents and does not provide question-answering retrieval.

Exam trap

The AI-102 exam often tests the distinction between services that process language (like LUIS or Translator) versus services that combine retrieval with generation (like Azure OpenAI with RAG) to answer questions from documents, leading candidates to mistakenly choose LUIS or Translator for a task that requires document-based Q&A.

Full explanation →

750

MCQeasy

You are building a solution that must summarize long documents in real time as they are uploaded to Azure Blob Storage. The summaries must be concise and capture the main points. You want to use Azure AI Language. Which feature should you use?

A.Extractive summarization

B.Key phrase extraction

C.Named entity recognition (NER)

D.Abstractive summarization

AnswerD

Abstractive summarization generates new, concise sentences that capture the main ideas, making it ideal for producing brief summaries of long documents. It is available in Azure AI Language and can process documents in real time via the API. This meets the requirement for concise, main-point summaries.

Why this answer

Abstractive summarization in Azure AI Language generates new, concise sentences that capture the main ideas of a document. It is designed for producing brief summaries and is available as a real-time API. Extractive summarization selects existing sentences, which may be less concise.

Key phrase extraction and NER provide different types of analysis and do not generate summaries.

Exam trap

The trap here is confusing summarization with other text analytics features like key phrase extraction or NER, which do not produce a summary.

Full explanation →

Microsoft Azure AI Engineer Associate AI-102 (AI-102) — Questions 676–750