Microsoft Azure AI Engineer Associate AI-102 AI-102 Questions 151–225 | Page 3/11

151

MCQmedium

Your organization has a large corpus of legal documents stored in Azure Blob Storage. You need to build a solution that allows lawyers to ask natural language questions and get answers directly from the documents, without moving data out of Azure. Which service should you use?

A.Azure AI Document Intelligence

B.Azure AI Search with semantic search

C.Azure AI Language Service with custom question answering

D.Azure AI Computer Vision

AnswerB

Azure AI Search with semantic search reranks results using language understanding models, returning relevant passages from indexed legal documents. It satisfies the requirement to answer natural-language questions directly from documents while keeping data within Azure.

Why this answer

Azure AI Search with semantic search is the correct choice because it allows you to index legal documents stored in Azure Blob Storage, then query them using natural language questions. The semantic search capability re-ranks results based on contextual understanding, enabling the system to extract precise answers from the document corpus without moving data out of Azure.

Exam trap

The trap here is that candidates often confuse Azure AI Document Intelligence (which extracts structured data from forms) with a search-based Q&A solution, failing to recognize that Azure AI Search with semantic search is the correct service for querying unstructured text corpora with natural language.

How to eliminate wrong answers

Option A is wrong because Azure AI Document Intelligence is designed for extracting structured data (e.g., key-value pairs, tables) from scanned documents, not for answering natural language questions across a large corpus. Option C is wrong because Azure AI Language Service with custom question answering requires a predefined FAQ or QnA pair structure and does not natively index and search unstructured document content like legal documents. Option D is wrong because Azure AI Computer Vision is focused on image analysis and optical character recognition (OCR), not on text-based question answering or semantic search.

Full explanation →

152

MCQeasy

A company wants to build a customer support agent using Microsoft Copilot Studio. The agent needs to understand natural language and handle complex queries beyond simple keyword matching. The agent should be able to escalate to a human agent when it cannot resolve the issue. Which feature should the agent use to understand natural language?

A.Create a Power Automate flow to process queries.

B.Enable generative answers and configure a knowledge source.

C.Create topics with trigger phrases.

D.Use entities to extract key information.

AnswerB

Generative answers with a configured knowledge source let the agent use large language models to interpret intent and retrieve relevant content, handling complex queries beyond keyword matching. This satisfies the natural language understanding requirement, and escalation can be added separately.

Why this answer

Generative answers in Microsoft Copilot Studio use large language models (LLMs) to interpret natural language queries and generate responses based on configured knowledge sources (e.g., SharePoint, websites, or custom data). This enables the agent to handle complex, conversational queries beyond simple keyword matching, and it can escalate to a human agent when confidence is low or the issue cannot be resolved.

Exam trap

The trap here is that candidates often confuse entity extraction (Option D) or topic triggers (Option C) with true natural language understanding, not realizing that generative answers powered by LLMs are required for handling complex, unconstrained queries beyond simple keyword matching.

How to eliminate wrong answers

Option A is wrong because Power Automate flows are for automating workflows and integrating systems, not for understanding natural language or handling complex queries in a conversational agent. Option C is wrong because topics with trigger phrases rely on keyword-based pattern matching to route conversations, which cannot handle the nuanced, multi-turn understanding required for complex queries. Option D is wrong because entities extract specific data points (e.g., dates, product names) from user input but do not provide the broad natural language comprehension needed to interpret and respond to complex queries.

Full explanation →

153

MCQmedium

Refer to the exhibit. You are deploying an Azure AI Services resource using an ARM template. After deployment, you cannot access the resource from any client, including the Azure portal. What is the most likely cause?

A.The SKU S0 does not support network restrictions.

B.The cognitiveServiceName conflicts with an existing resource.

C.The networkAcls defaultAction is set to Deny with no allowed IP or virtual network rules.

D.The virtualNetworkRules array is empty.

AnswerC

Setting networkAcls defaultAction to Deny without any allow rules blocks all traffic, including the Azure portal, since no IP or subnet is permitted. This explains why every client fails to reach the Azure AI Services resource after deployment.

Why this answer

The networkAcls defaultAction set to Deny with no allowed IP or virtual network rules blocks all traffic to the Azure AI Services resource, including requests from the Azure portal and any client. This is because the default network access control list (ACL) denies all traffic unless explicitly permitted by IP rules or virtual network rules, rendering the resource inaccessible even for management operations.

Exam trap

The trap here is that candidates often assume an empty virtualNetworkRules array means no restrictions, but the defaultAction property controls the default behavior, and setting it to Deny with no allow rules blocks all traffic regardless of empty arrays.

How to eliminate wrong answers

Option A is wrong because the S0 SKU does support network restrictions; network ACLs are available for all paid SKUs (S0 and above), and the issue is not related to SKU limitations. Option B is wrong because a name conflict would cause a deployment error, not a scenario where the resource is deployed but inaccessible; the ARM template would fail with a conflict error during deployment. Option D is wrong because an empty virtualNetworkRules array is irrelevant when the defaultAction is Deny; the resource would still be blocked since no IP rules are specified to allow traffic, and an empty array does not implicitly permit any traffic.

Full explanation →

154

MCQeasy

You are using Azure AI Language to analyze customer reviews. You need to determine whether each review expresses a positive, negative, or neutral sentiment. Which API should you call?

A.Language detection API.

B.Entity recognition API.

C.Sentiment analysis API.

D.Key phrase extraction API.

AnswerC

The sentiment analysis API returns per-document and per-sentence scores across positive, negative and neutral labels, directly satisfying the requirement to classify each review's polarity. Other Azure AI Language features, such as key phrase extraction or entity recognition, do not produce sentiment classifications.

Why this answer

The Sentiment Analysis API is specifically designed to evaluate text and return sentiment labels (positive, negative, neutral) along with confidence scores. This directly matches the requirement to determine whether each customer review expresses positive, negative, or neutral sentiment.

Exam trap

The trap here is that candidates confuse 'sentiment' with 'key phrases' or 'entities,' assuming that extracting important words or names can imply sentiment, but only the Sentiment Analysis API directly evaluates emotional tone.

How to eliminate wrong answers

Option A is wrong because the Language Detection API identifies the language of the text (e.g., English, Spanish), not the sentiment. Option B is wrong because the Entity Recognition API extracts named entities such as people, places, and organizations, not sentiment. Option D is wrong because the Key Phrase Extraction API identifies important phrases and topics in the text, but does not evaluate sentiment.

Full explanation →

155

MCQmedium

You are testing an Azure OpenAI Service chat completion with function calling. The assistant returned null content and a tool call. What does this indicate?

A.The model is unable to answer the question and returned an error

B.The system message is preventing the model from responding

C.The model is using a tool to answer the question, so content is null

D.The function call syntax is invalid and the model skipped it

AnswerC

With function calling, the model returns a tool call instead of a natural-language reply, so the content field is null. Your code must execute the named function and send its result back for a final completion.

Why this answer

In Azure OpenAI Service function calling, when the model determines that it needs to invoke a tool (e.g., an external API or database query) to fulfill the user's request, it returns a `tool_calls` object with `content` set to `null`. This is the expected behavior—the model is signaling that it is delegating the response to the tool, not that it has failed or is blocked.

Exam trap

The trap here is that candidates often misinterpret a null content as a model failure or error, when in fact it is a standard indicator that the model is actively using a tool to fulfill the request.

How to eliminate wrong answers

Option A is wrong because a null content with a tool call is not an error; it is a deliberate design pattern in function calling where the model defers to the tool for the answer. Option B is wrong because the system message does not prevent the model from responding; the model actively chooses to return a tool call instead of text content. Option D is wrong because if the function call syntax were invalid, the model would either ignore the function definitions entirely or return an error, not a valid tool call with null content.

Full explanation →

156

MCQmedium

A company uses the Computer Vision Image Analysis API to generate captions for images. The captions are often too generic. How can they improve the descriptiveness of captions?

A.Use the Object Detection API instead.

B.Train a custom model with Custom Vision.

C.Increase the confidence threshold for captions.

D.Use the Dense Captioning feature.

AnswerD

Dense Captioning generates multiple descriptions for distinct regions within an image, not one caption for the whole frame. This satisfies the stem's requirement for richer, more descriptive output, since generic captions stem from whole-image summarisation. It also returns bounding boxes, adding spatial detail the standard caption endpoint cannot provide.

Why this answer

The Dense Captioning feature in the Computer Vision Image Analysis API generates one-sentence descriptions for each of up to 10 regions detected in an image, providing more specific and detailed captions than the single generic caption. This directly addresses the problem of captions being too generic by breaking the image into meaningful areas and describing each one individually.

Exam trap

The trap here is that candidates confuse increasing the confidence threshold with improving descriptiveness, when in reality it only reduces the number of captions returned without adding detail.

How to eliminate wrong answers

Option A is wrong because the Object Detection API identifies and locates objects with bounding boxes but does not generate descriptive captions or improve caption descriptiveness. Option B is wrong because training a custom model with Custom Vision requires labeled images and is designed for classification or object detection, not for generating richer captions from the existing Image Analysis API. Option C is wrong because increasing the confidence threshold for captions only filters out lower-confidence results, making captions less frequent or more conservative, not more descriptive.

Full explanation →

157

MCQeasy

You are developing a chatbot that uses Azure AI Language to understand user intents. The chatbot must handle multiple languages and direct users to the appropriate support team based on the detected intent. Which Azure AI Language feature should you use?

A.Text Analytics

B.Conversational language understanding (CLU)

C.QnA Maker

D.Translator

AnswerB

Conversational language understanding (CLU) extracts intents and entities from utterances, and its multilingual training lets one project recognise the same intent across several languages. That satisfies the stem's requirement to detect intent and route users to the correct support team, which plain language detection or translation alone cannot do.

Why this answer

Conversational language understanding (CLU) is the correct feature because it is specifically designed to extract intents and entities from user utterances in a multi-language conversational context. CLU enables the chatbot to detect the user's intent (e.g., 'billing', 'technical support') and route them to the appropriate support team, while also supporting multiple languages through its language-agnostic model training and per-language project configurations.

Exam trap

The trap here is that candidates often confuse 'Text Analytics' (pre-built NLP) with 'Conversational language understanding' (custom intent/entity extraction), mistakenly thinking that Text Analytics can be trained to classify intents, when in fact it only provides pre-built capabilities like sentiment and key phrases.

How to eliminate wrong answers

Option A is wrong because Text Analytics (now part of Azure AI Language) provides pre-built sentiment analysis, key phrase extraction, and entity recognition, but it does not offer custom intent classification or entity extraction for conversational flows — it is not designed for building a chatbot's intent routing logic. Option C is wrong because QnA Maker (now Azure AI Language's custom question answering) is optimized for providing direct answers from a knowledge base of FAQ-style Q&A pairs, not for detecting user intents and routing to different support teams; it lacks the intent classification engine required for multi-intent conversational scenarios. Option D is wrong because Translator is a pure machine translation service that translates text between languages without any capability to detect intents or entities — it cannot interpret the user's goal or route them to a support team.

Full explanation →

158

MCQeasy

You are building a conversational AI system using Azure OpenAI Service. The system must maintain context across multiple user turns. Which parameter determines how many previous messages are considered for the next response?

A.max_tokens

B.temperature

C.top_p

D.The length of the messages array in the API call

AnswerD

The messages array carries the full conversation history sent to the model, so its length directly bounds how many prior turns inform the next completion. Truncating or extending that array controls retained context; no separate memory parameter exists in the Chat Completions API.

Why this answer

The Azure OpenAI Service API uses a `messages` array in the request body to represent the conversation history. Each entry in this array corresponds to a previous turn (with roles like 'user', 'assistant', or 'system'), and the length of this array directly determines how many prior messages are considered when generating the next response. By including more messages, you extend the context window; by truncating the array, you limit it.

Exam trap

The trap here is that candidates often confuse parameters that control output generation (like `max_tokens`, `temperature`, or `top_p`) with the mechanism for maintaining conversation history, which is explicitly managed by the structure of the API call's `messages` array.

How to eliminate wrong answers

Option A is wrong because `max_tokens` controls the maximum number of tokens (words/subwords) in the generated response, not the number of previous messages considered for context. Option B is wrong because `temperature` is a sampling parameter that influences the randomness or creativity of the output, not the conversation history length. Option C is wrong because `top_p` (nucleus sampling) sets a probability threshold for token selection, affecting output diversity, not the number of prior turns used as context.

Full explanation →

159

MCQeasy

You need to build a solution that detects whether a person is wearing a hard hat in a construction site image. Which Azure AI service should you use?

A.Azure AI Video Indexer

B.Azure AI Face

C.Azure AI Custom Vision

D.Azure AI Document Intelligence

AnswerC

Custom Vision trains an image classification or object detection model on your own labelled hard-hat images, satisfying the stem's specific detection requirement. Prebuilt Computer Vision models lack a hard-hat class, so custom training is required.

Why this answer

Azure AI Custom Vision is the correct service because it allows you to train a custom object detection model to identify specific objects—such as a hard hat—in images. Unlike pre-built services, Custom Vision lets you upload labeled images of workers with and without hard hats, train a model, and then use it to detect hard hat presence in construction site photos. This tailored approach is necessary because hard hat detection is a specialized use case not covered by generic vision APIs.

Exam trap

The trap here is that candidates often confuse Azure AI Custom Vision with pre-built vision services like Azure AI Face or Video Indexer, assuming they can be repurposed for custom object detection, but only Custom Vision allows training on your own labeled images for specific items like hard hats.

How to eliminate wrong answers

Option A is wrong because Azure AI Video Indexer is designed for analyzing video content (e.g., extracting transcripts, faces, and scenes) and does not support custom object detection for specific items like hard hats in static images. Option B is wrong because Azure AI Face is specialized for detecting and analyzing human faces (e.g., attributes like age, emotion, or identity) and cannot identify objects such as hard hats. Option D is wrong because Azure AI Document Intelligence (formerly Form Recognizer) is built for extracting text, tables, and key-value pairs from documents, not for object detection in images.

Full explanation →

160

MCQeasy

You need to enforce that only users from your Microsoft Entra ID tenant can call your Azure AI Language API endpoint. Which security mechanism should you configure?

A.Microsoft Entra ID authentication

B.API key authentication

C.IP whitelist

D.Azure Firewall

AnswerA

Microsoft Entra ID authentication validates bearer tokens issued by your tenant, so only principals within that tenant can call the endpoint. This enforces the tenant restriction directly, unlike key-based access which any key holder can use.

Why this answer

Microsoft Entra ID authentication (formerly Azure AD) allows you to enforce that only users and applications from your specific Entra ID tenant can call the Azure AI Language API. This is achieved by configuring a managed identity or service principal and assigning it the Cognitive Services User role, which ensures that tokens issued by your tenant are required for access. API keys and IP whitelists do not provide tenant-level identity enforcement, and Azure Firewall is a network-level control that does not authenticate individual users or applications.

Exam trap

The trap here is that candidates often confuse network-level controls (IP whitelist, Azure Firewall) with identity-level controls, mistakenly thinking that restricting by IP address or network firewall is sufficient to enforce tenant-specific access, when in fact only Entra ID authentication can validate the caller's tenant membership.

How to eliminate wrong answers

Option B is wrong because API key authentication uses a static key that can be shared or leaked, and it does not verify the caller's identity or tenant membership, so any user with the key can access the endpoint regardless of their Entra ID tenant. Option C is wrong because an IP whitelist only restricts access based on source IP addresses, which does not authenticate the user or ensure they belong to a specific Entra ID tenant; an attacker could spoof an IP or use a VPN from an allowed range. Option D is wrong because Azure Firewall is a network security service that filters traffic at the network layer, not at the application or identity layer, and it cannot enforce tenant-specific authentication for API calls.

Full explanation →

161

Multi-Selecteasy

Which THREE Azure AI services can be used to extract text from images?

Select 3 answers

A.Azure AI Speech

B.Azure AI Search

C.Azure AI Document Intelligence layout model

D.Azure AI Vision OCR

E.Azure AI Language custom NER

AnswersB, C, D

Azure AI Search supports OCR enrichment through its built-in cognitive skills, extracting text from image content during indexing. This satisfies the requirement for a service that pulls text out of images, alongside Vision and Document Intelligence.

Why this answer

Option B (Azure AI Search) is correct because it includes AI enrichment with the OCR cognitive skill, which extracts text from image files (e.g., JPEG, PNG) during the indexing pipeline, making the text searchable. Option C (Azure AI Document Intelligence layout model) is correct because the layout model performs OCR on documents and images, extracting printed and handwritten text along with tables and structure from forms and files. Option D (Azure AI Vision OCR) is correct because the Read/OCR API in Azure AI Vision extracts printed and handwritten text from images and documents.

Option A (Azure AI Speech) is incorrect because it handles speech-to-text, text-to-speech, and translation of audio, not text extraction from images. Option E (Azure AI Language custom NER) is incorrect because it extracts named entities from existing text, not text from images.

Exam trap

The trap is that candidates may overlook Azure AI Search as a text extraction service because it is not a dedicated OCR service, but it can indeed extract text from images when configured with an OCR skill.

Full explanation →

162

Multi-Selectmedium

Which TWO actions should you take to ensure that a generative AI model deployed on Azure Machine Learning is compliant with data privacy regulations?

Select 2 answers

A.Implement data masking during preprocessing.

B.Store all training data in a separate Azure region.

C.Use differential privacy during model training.

D.Encrypt the model at rest using Azure Key Vault.

E.Log all raw input data to Azure Monitor for auditing.

AnswersA, C

Masking replaces sensitive field values such as names or card numbers with obfuscated substitutes before the data reaches training, so personally identifiable information never enters the model. This directly satisfies the regulation's requirement that personal data be de-identified at the preprocessing stage.

Why this answer

Options A and C are correct. A: Data masking during preprocessing helps anonymize sensitive data before it is used, reducing privacy risks. C: Differential privacy adds noise to the training process, preventing the model from memorizing individual data points.

B is wrong because storing data in a separate Azure region does not inherently address privacy compliance; it may help with data residency but not the core privacy protection. D is wrong because encrypting the model at rest with Azure Key Vault is a security measure, not a privacy compliance measure. E is wrong because logging raw input data to Azure Monitor could expose sensitive information, violating privacy.

Full explanation →

163

Multi-Selecteasy

You are using the Azure AI Language service to process customer reviews. You need to extract the following insights: overall sentiment, key phrases, and entity types (such as product names). Which THREE operations should you call?

Select 3 answers

A.Key Phrase Extraction

B.Language Detection

C.Sentiment Analysis

D.PII Detection

E.Entity Recognition

AnswersA, C, E

Key Phrase Extraction returns the salient terms within review text, satisfying the requirement to surface key phrases. It is one of three separate Azure AI Language operations, alongside Sentiment Analysis and Named Entity Recognition, so it cannot alone deliver sentiment scores or entity types.

Why this answer

Sentiment Analysis (C) is correct because it returns the overall sentiment of the review (positive, negative, neutral, or mixed) along with confidence scores, which is exactly the first insight required. Key Phrase Extraction (A) is correct because it identifies the main talking points in the review text, satisfying the key phrases requirement. Entity Recognition (E) is correct because it extracts and classifies named entities such as product names, locations, and organizations, which covers the entity types requirement.

Language Detection (B) is not needed because the scenario does not ask for identifying the review's language, and PII Detection (D) is not needed because the scenario does not require detecting or redacting personal information.

Exam trap

The trap here is that candidates often confuse Language Detection or PII Detection with the required insights, mistakenly thinking language identification or privacy data extraction fulfills the need for sentiment, key phrases, and entity types, when in fact they serve entirely different purposes.

Full explanation →

164

MCQhard

You executed the Azure CLI command to list Azure OpenAI resources. You need to programmatically access the endpoint of the resource named 'gpt4' in a script. What is the most reliable way to extract the endpoint?

A.Use Azure PowerShell Get-AzCognitiveServicesAccount cmdlet

B.Parse the table output using string manipulation

C.Use az cognitiveservices account show with --query to filter by name

D.Deploy a new Azure OpenAI resource with a known endpoint

AnswerC

Using `az cognitiveservices account show` with `--query` retrieves the endpoint directly from the resource's properties, satisfying the need for programmatic extraction in a script. Unlike parsing list output, querying the named account returns `properties.endpoint` deterministically, avoiding ambiguity when multiple Azure OpenAI resources exist in the subscription.

Why this answer

`az cognitiveservices account show` with the `--query` parameter allows you to retrieve the specific endpoint property for a named resource using JMESPath filtering. This approach is reliable, scriptable, and avoids parsing unstructured output, which is error-prone. The Azure CLI returns structured JSON, and `--query` extracts the exact value without manual string manipulation.

Exam trap

The trap here is that candidates may default to parsing the default table output (Option B) because it looks human-readable, but the exam tests understanding that structured JSON queries are the reliable, production-grade method for programmatic access.

How to eliminate wrong answers

Option A is wrong because `Get-AzCognitiveServicesAccount` retrieves all cognitive services accounts, but does not directly filter by resource name in a single cmdlet; you would need additional piping or filtering, and it returns the entire object, not just the endpoint. Option B is wrong because parsing table output with string manipulation is fragile, depends on column alignment, and breaks if the CLI output format changes or if the resource name contains special characters. Option D is wrong because deploying a new resource is unnecessary, wasteful, and does not solve the requirement to access an existing resource's endpoint.

Full explanation →

165

MCQeasy

You are using Azure AI Content Safety to moderate user-generated content in a social media app. The solution must detect and block hate speech and self-harm content in real time. Which Content Safety feature should you use?

A.Custom category management

B.Severity analysis

C.Image moderation

D.Text moderation

AnswerD

Text moderation applies Azure AI Content Safety's classification models to written user content, returning severity scores for hate and self-harm categories so the app can block them in real time. It directly satisfies the stated requirement to detect and block those two harm types.

Why this answer

Text moderation is the correct choice because Azure AI Content Safety's text moderation API is specifically designed to detect and block hate speech and self-harm content in real time. It analyzes text for severity levels across multiple categories, including hate and self-harm, and returns a severity score to trigger blocking actions. This directly meets the requirement for real-time detection of these specific content types in user-generated text.

Exam trap

The trap here is that candidates may confuse severity analysis as a standalone feature, but it is actually a sub-component of text moderation; the question asks for the feature that performs the detection, not the analysis of the detection results.

How to eliminate wrong answers

Option A is wrong because custom category management allows you to define your own content categories (e.g., brand-specific terms), but it does not replace the built-in hate speech and self-harm detection; the question requires using existing Azure AI Content Safety features, not custom definitions. Option B is wrong because severity analysis is a component of the moderation process (it assigns a severity score to detected content), not a standalone feature; you must use text moderation to first detect the content before severity analysis can be applied. Option C is wrong because image moderation is used to analyze visual content (images) for inappropriate content, but the question specifically mentions user-generated content that includes hate speech and self-harm, which are primarily text-based; image moderation does not analyze text within images by default.

Full explanation →

166

MCQhard

Refer to the exhibit. You are defining a custom entity recognition model in Azure AI Language. The exhibit shows a partial configuration. What is the relationship between 'Laptop' and 'Electronics'?

A.Laptop is a type of Electronics.

B.There is no defined relationship.

C.Electronics is a type of Laptop.

D.Laptop is a part of Electronics.

AnswerA

The exhibit configures 'Electronics' as a parent entity and 'Laptop' as its child, so the model treats Laptop as a subtype of Electronics. This hierarchical relationship enables extraction of both the specific item and its broader category from text.

Why this answer

In Azure AI Language custom entity recognition, you define entity types and subtypes using a hierarchical structure. The exhibit shows 'Laptop' as a child of 'Electronics', meaning Laptop is a subtype or specific type of the broader Electronics category. This allows the model to recognize that any Laptop entity is also an instance of Electronics, enabling more granular classification and downstream processing.

Exam trap

The trap here is that candidates confuse hierarchical 'type-of' relationships with 'part-of' relationships, leading them to incorrectly select Option D, or they assume no relationship exists (Option B) because they overlook the visual hierarchy in the exhibit.

How to eliminate wrong answers

Option B is wrong because the exhibit explicitly shows a parent-child relationship between 'Electronics' and 'Laptop', so there is a defined relationship. Option C is wrong because it reverses the hierarchy: 'Electronics' is the parent category, not a subtype of 'Laptop'. Option D is wrong because 'part of' implies a meronymic relationship (e.g., a keyboard is part of a laptop), but the exhibit uses a type-of (hyponymic) relationship, not a part-whole relationship.

Full explanation →

167

MCQhard

Your team is using the Azure OpenAI Batch API to process 200,000 product-description rewrites overnight. The job must complete within a fixed window and cost as little as possible, and results are not needed interactively. A developer reports that the submitted batch job failed with an error about the input file. What is the most likely cause?

A.The batch job used a deployment with a lower tokens-per-minute quota than the interactive deployment.

B.The input file was uploaded as a single large JSON array containing all 200,000 requests.

C.The input file was uploaded as a JSONL file with one request object per line and a unique custom_id for each line.

D.The batch job was submitted without specifying a completion window, so it defaulted to the maximum allowed.

AnswerB

Batch API input must be a JSONL file where each line is one request object with its own custom_id. A single JSON array is not the accepted format, so the job fails validation with an input-file error. Converting the payload to line-delimited request objects and uploading it to the expected storage location resolves the failure.

Why this answer

The Batch API expects a JSONL input file in which every line is an independent request object with a unique custom_id. Submitting one large JSON array violates that contract, so the job is rejected during input validation before any processing begins.

Exam trap

The trap here is assuming any valid JSON is acceptable, when the Batch API specifically requires line-delimited JSONL with one request per line.

Full explanation →

168

MCQmedium

A company uses Azure AI Language's custom text classification to categorize support tickets. The model was trained with 5000 labeled examples and achieves 90% accuracy. However, for a specific category (e.g., 'billing'), the model frequently misclassifies tickets that contain both billing and technical issues. Which action should you take to improve classification for this category?

A.Reduce the number of categories to simplify the classification.

B.Add more labeled examples for the 'billing' category, especially those that are mixed with other categories.

C.Increase the number of training epochs to further train the model.

D.Use a different classification algorithm, such as a neural network.

AnswerB

Mixed billing-and-technical tickets are underrepresented, so the model lacks boundary examples distinguishing overlapping categories. Adding labelled examples that explicitly cover these ambiguous, multi-intent tickets gives the classifier the discriminative signal needed to separate billing from technical content.

Why this answer

Adding more labeled examples for the 'billing' category, especially those that are mixed with other categories, will help the model learn to distinguish them better. Option A is wrong because reducing the number of categories may not address the specific confusion. Option C is wrong because increasing the training epochs may lead to overfitting.

Option D is wrong because using a different algorithm is not an option in Azure AI Language's custom text classification.

Full explanation →

169

MCQmedium

You are planning an Azure AI solution that uses Azure OpenAI. You need to ensure that the solution can be deployed to multiple regions for high availability. The solution must automatically route requests to the nearest available region and fail over if a region becomes unavailable. What should you use?

A.Azure Front Door with latency-based routing

B.Azure Load Balancer with availability zones

C.Azure Application Gateway with multi-site listeners

D.Azure Traffic Manager with priority routing

AnswerA

Azure Front Door is a global HTTP load balancer that supports latency-based routing, which directs requests to the lowest-latency endpoint. It also provides automatic failover if an endpoint becomes unhealthy. This meets the requirements for multi-region deployment, nearest-region routing, and failover for Azure OpenAI.

Why this answer

Azure Front Door is the appropriate service because it provides global HTTP load balancing with latency-based routing and automatic failover. It can route requests to the nearest available Azure OpenAI endpoint across multiple regions and detect unhealthy endpoints to redirect traffic. The other options are either regional load balancers or do not support latency-based global routing.

Exam trap

The trap here is selecting a regional load balancer like Azure Load Balancer or Application Gateway for a multi-region scenario, or choosing Traffic Manager with priority routing when latency-based routing is required.

Full explanation →

170

MCQeasy

You need to monitor the performance of an Azure AI Language service custom entity recognition model. Which metric should you track to evaluate the model's ability to correctly identify entities?

A.Throughput

B.Response latency

C.F1 score

D.Accuracy

AnswerC

F1 score is the harmonic mean of precision and recall, so it captures both missed entities and false positives. That balance directly measures how correctly the custom entity recognition model identifies entities, unlike accuracy or latency metrics.

Why this answer

The F1 score is the standard metric for evaluating custom entity recognition models in Azure AI Language, as it balances precision (correctly identified entities) and recall (missed entities). Unlike accuracy, which can be misleading due to class imbalance in entity labeling, F1 provides a harmonic mean that reflects the model's ability to correctly identify entities without bias toward the majority class.

Exam trap

The trap here is that candidates confuse accuracy (a common metric in classification) with the specialized F1 score required for entity recognition, where class imbalance makes accuracy a poor indicator of model performance.

How to eliminate wrong answers

Option A is wrong because throughput measures the number of requests processed per second, not the quality of entity identification. Option B is wrong because response latency measures the time taken to return a prediction, not the correctness of entity predictions. Option D is wrong because accuracy (ratio of correct predictions to total predictions) is misleading for entity recognition tasks where the number of non-entity tokens vastly outnumbers entity tokens, leading to inflated accuracy even if the model fails to identify entities.

Full explanation →

171

MCQeasy

You are a solution architect at a media company. The company uses Azure AI Speech to generate subtitles for videos. The current solution uses the batch transcription API and takes several hours to process a 1-hour video. The business requires near-real-time subtitles for live streaming events. You need to design a new solution that provides low-latency transcription. You have the following options: Option A: Use the batch transcription API with a higher priority queue. Option B: Use the Speech-to-text REST API for real-time streaming with the Speech SDK. Option C: Use the Azure AI Language API to transcribe audio from a file. Option D: Use Azure AI Video Indexer to generate subtitles.

A.Option C

B.Option B

C.Option D

D.Option A

AnswerB

Speech SDK with real-time streaming provides low-latency transcription.

Why this answer

The Speech-to-text REST API with the Speech SDK supports real-time streaming transcription, which provides low-latency results suitable for live streaming events. Unlike the batch transcription API, which processes audio asynchronously and can take hours, the streaming API processes audio chunks in near-real-time, returning partial and final results with sub-second latency.

Exam trap

The trap here is that candidates may confuse the batch transcription API's priority queues with real-time performance, or mistakenly think the Azure AI Language API can handle speech-to-text tasks, when it is strictly a text-based NLP service.

How to eliminate wrong answers

Option A is wrong because the batch transcription API is designed for asynchronous, high-latency processing; even with a higher priority queue, it still processes audio in batches and cannot achieve the sub-second latency required for live streaming. Option C is wrong because the Azure AI Language API is for text analytics (e.g., sentiment, key phrases), not for transcribing audio from a file; it does not include speech-to-text capabilities. Option D is wrong because Azure AI Video Indexer is optimized for indexing and analyzing pre-recorded videos, not for real-time streaming transcription; it introduces significant latency due to its indexing pipeline.

Full explanation →

172

MCQmedium

You are building an Azure AI Search enrichment pipeline that must extract text and layout information from scanned PDFs stored in Azure Blob Storage. The extracted content must include bounding boxes for each text line so that a downstream custom skill can associate key-value pairs spatially. You need to add a built-in skill to the skillset to perform this extraction. Which skill should you add?

A.DocumentExtractionSkill

B.TextMergeSkill

C.OcrSkill

D.DocumentIntelligenceLayoutSkill

AnswerD

DocumentIntelligenceLayoutSkill is a built-in Azure AI Search skill that leverages Azure AI Document Intelligence to extract text, layout, and structure from documents, including scanned PDFs. It returns bounding boxes for text lines and other layout elements, enabling spatial association in downstream skills. This skill is specifically designed for document layout analysis, making it the correct choice to meet the requirement for bounding box coordinates.

Why this answer

The requirement is to extract text and layout information, including bounding boxes, from scanned PDFs. DocumentIntelligenceLayoutSkill is a built-in Azure AI Search skill that uses Azure AI Document Intelligence to analyze document structure and return bounding boxes for text lines. This enables downstream spatial association.

Other skills either do not provide layout data or are not designed for direct PDF analysis, so they fail to meet the spatial requirement.

Exam trap

The trap here is assuming that any text extraction skill returns layout coordinates, when in fact only DocumentIntelligenceLayoutSkill provides bounding boxes for scanned PDFs.

Full explanation →

173

MCQmedium

A company is using Azure OpenAI to generate customer support responses. They want to ensure the model does not use any personally identifiable information (PII) in its outputs. What should they implement?

A.Fine-tune the model on anonymized data.

B.Use prompt engineering to instruct the model to redact PII.

C.Use Azure AI Content Safety to filter PII from the output.

D.Use a system message instructing the model to avoid PII.

AnswerC

Azure AI Content Safety can detect and block PII.

Why this answer

Azure AI Content Safety provides built-in PII detection and redaction capabilities that can automatically scan and filter sensitive information from model outputs. This is the most reliable approach because it operates as a post-processing filter, catching PII that the model might generate despite instructions. Fine-tuning, prompt engineering, and system messages are all fallible because they rely on the model's compliance rather than enforced filtering.

Exam trap

The trap here is that candidates confuse 'instruction-based approaches' (prompts, system messages) with 'enforcement-based approaches' (content safety filters), assuming that telling the model not to do something is as effective as actively filtering the output.

How to eliminate wrong answers

Option A is wrong because fine-tuning on anonymized data does not prevent the model from generating PII during inference; it only reduces the likelihood based on training data, and the model can still hallucinate or leak PII from its pretrained knowledge. Option B is wrong because prompt engineering is a soft instruction that the model may ignore or fail to apply consistently, especially with edge cases or adversarial inputs, and it does not provide guaranteed redaction. Option D is wrong because a system message is merely a directive to the model, not a technical enforcement mechanism; the model can still output PII if it misinterprets or overrides the instruction.

Full explanation →

174

MCQeasy

You are using Azure OpenAI Service to summarize customer emails. The summaries must be concise and contain only key information. Which prompt engineering technique should you apply?

A.Use few-shot prompting with examples of desired summaries

B.Use chain-of-thought prompting

C.Use zero-shot prompting with a one-sentence instruction

D.Use negative prompting to avoid verbose output

AnswerA

Few-shot prompting supplies the model with paired examples of emails and their concise summaries, directly demonstrating the desired length and content selection. This satisfies the stem's constraint that summaries contain only key information, since the exemplars establish the pattern the model imitates rather than relying on vague instructions alone.

Why this answer

Few-shot prompting is the correct technique because it provides the model with explicit examples of desired input-output pairs (e.g., a verbose email and its concise summary). This guides the model to learn the exact format, tone, and level of detail required for the summaries, which is critical for consistency in a production summarization pipeline. Without examples, the model may default to its training distribution and produce overly verbose or irrelevant output.

Exam trap

The trap here is that candidates often assume a simple instruction (zero-shot) is sufficient for summarization, underestimating how much the model relies on explicit examples to enforce output structure and conciseness, especially when the task requires domain-specific key information extraction.

How to eliminate wrong answers

Option B is wrong because chain-of-thought prompting is designed for multi-step reasoning tasks (e.g., math word problems, logical deduction) where intermediate steps are needed, not for summarization where the goal is direct extraction of key information. Option C is wrong because zero-shot prompting with a one-sentence instruction lacks the concrete examples needed to constrain the model's output style and length, often resulting in summaries that are too long or miss critical details. Option D is wrong because negative prompting (e.g., 'do not be verbose') is unreliable; the model may misinterpret the negation or still produce verbose output because it lacks positive examples of the desired concise format.

Full explanation →

175

MCQhard

Your Azure AI solution uses multiple AI services including Computer Vision and Language. To reduce costs, you want to share a single key and endpoint across services. Which Azure resource type should you deploy?

A.Separate single-service resources for each AI service

B.Azure Key Vault for storing keys

C.Azure API Management gateway

D.Azure AI services multi-service resource

AnswerD

A multi-service resource provisions one key and endpoint spanning Computer Vision, Language and other Azure AI services, so every call authenticates against the same credentials. This directly satisfies the cost-reduction constraint of sharing a single key and endpoint rather than deploying separate per-service resources.

Why this answer

D is correct because Azure AI services multi-service resource provides a single endpoint and key that can be used across multiple AI services (e.g., Computer Vision, Language, Face, etc.), reducing management overhead and cost by consolidating billing. This resource type is designed specifically for scenarios where you want to share credentials across services without deploying separate single-service instances.

Exam trap

The trap here is that candidates may confuse Azure API Management (Option C) as a way to share endpoints, but it is a gateway for API management, not a shared AI resource; the correct answer is the multi-service resource that natively provides a single key and endpoint for multiple AI services.

How to eliminate wrong answers

Option A is wrong because deploying separate single-service resources for each AI service would require managing multiple keys and endpoints, increasing complexity and cost, which contradicts the goal of reducing costs through sharing. Option B is wrong because Azure Key Vault is a service for securely storing and managing secrets (like keys), not for providing a shared endpoint or key for AI services; it does not replace the need for an AI resource. Option C is wrong because Azure API Management gateway is used to create, publish, and manage APIs, not to provide a shared key and endpoint for Azure AI services; it adds an extra layer of abstraction and cost, not a direct shared resource.

Full explanation →

176

MCQeasy

You are developing a generative AI application that uses Azure OpenAI Service to summarize large documents. The application experiences high latency when processing requests. You need to reduce the latency without changing the model. What should you do?

A.Increase the temperature parameter

B.Increase the top_p parameter

C.Reduce the max_tokens parameter in the API request

D.Increase the max_tokens parameter

AnswerC

max_tokens caps generated output length, so lowering it shortens generation time and reduces latency without altering the deployed model. The stem forbids changing the model, and this parameter is a per-request setting, making it the appropriate lever for summarisation workloads.

Why this answer

Reducing the max_tokens parameter limits the length of the generated response, which directly reduces the processing time required by the Azure OpenAI Service to produce the output. Since latency is caused by the model generating a long sequence of tokens, capping the output tokens decreases the number of autoregressive decoding steps, thereby lowering response time without altering the underlying model.

Exam trap

The trap here is that candidates often confuse parameters that affect output length (max_tokens) with those that affect output diversity (temperature, top_p), mistakenly believing that adjusting randomness can speed up generation.

How to eliminate wrong answers

Option A is wrong because increasing the temperature parameter controls the randomness of the output, not the length or speed of generation; it has no direct impact on latency. Option B is wrong because increasing the top_p parameter (nucleus sampling) affects the diversity of token selection but does not reduce the number of tokens generated or the processing time. Option D is wrong because increasing the max_tokens parameter would allow longer responses, which would increase the number of decoding steps and worsen latency, the opposite of the desired outcome.

Full explanation →

177

MCQhard

A company uses Azure AI Language Service for custom text classification. The model is trained to classify support tickets into categories. After deployment, the model performs well on the test set but poorly on new incoming tickets. Which action should be taken to improve generalization?

A.Switch to a prebuilt text classification model

B.Increase the number of training epochs

C.Reduce the confidence threshold for classification

D.Add more labeled data from actual production tickets

AnswerD

Adding labelled examples drawn from real production tickets exposes the model to the actual vocabulary, phrasing and class distribution of live traffic, which the test set failed to represent. This directly addresses the poor generalisation caused by training data that did not match incoming ticket characteristics.

Why this answer

Adding more labeled data from actual production tickets helps the model learn the true distribution of real-world inputs, reducing overfitting to the test set. The model's poor performance on new tickets indicates it memorized patterns specific to the training data rather than generalizing. Incorporating production data directly addresses the distribution shift between the test set and live traffic.

Exam trap

Candidates often mistakenly tune hyperparameters (epochs, confidence threshold) or switch to a prebuilt model, but the real issue is distribution shift between test data and production data. Adding representative labeled data from production is the correct solution.

How to eliminate wrong answers

Option A is wrong because switching to a prebuilt text classification model would not solve the generalization issue; prebuilt models are generic and unlikely to match the custom categories or domain-specific language of support tickets, potentially worsening performance. Option B is wrong because increasing the number of training epochs can lead to overfitting, especially if the model already performs well on the test set; more epochs do not improve generalization and may exacerbate memorization. Option C is wrong because reducing the confidence threshold for classification would lower the bar for predictions, causing more false positives and misclassifications, not improving the model's ability to generalize to new data.

Full explanation →

178

Multi-Selectmedium

You are designing a solution that uses Azure AI Vision to analyze images uploaded by users. You need to extract text and also generate a descriptive caption for each image. You want to use the Image Analysis API. Which two capabilities should you enable? (Choose two.)

Select 2 answers

A.Read

B.Tags

C.Brands

D.Caption

E.Objects

AnswersA, D

The Read capability in Image Analysis extracts printed and handwritten text from images and returns it as structured lines and words. This directly satisfies the requirement to extract text from user-uploaded images. It is the correct choice because it provides OCR functionality within the same API call, avoiding the need to call a separate OCR endpoint.

Why this answer

The Read capability extracts text, and the Caption capability generates a descriptive sentence for the image. Together they satisfy both requirements. Tags, Objects, and Brands provide different types of analysis that are not needed here.

Enabling Read and Caption allows a single API call to return both OCR results and a caption, streamlining the solution.

Exam trap

The trap here is confusing Tags with Caption; Tags provide keywords, whereas Caption produces a full descriptive sentence, which is what the scenario requires.

Full explanation →

179

MCQeasy

You need to analyze customer feedback to determine whether the sentiment is positive, negative, or neutral. Which Azure AI service should you use?

A.Azure AI Language - Key Phrase Extraction

B.Azure AI Language - Named Entity Recognition

C.Azure AI Language - Sentiment Analysis

D.Azure AI Language - Language Detection

AnswerC

Azure AI Language Sentiment Analysis returns per-document scores and labels of positive, negative or neutral, directly matching the requirement to classify customer feedback polarity. It is purpose-built for opinion mining rather than translation, OCR or entity extraction.

Why this answer

Azure AI Language's Sentiment Analysis is the correct service because it is specifically designed to evaluate text and return sentiment labels (positive, negative, neutral) along with confidence scores. This directly matches the requirement to determine whether customer feedback sentiment is positive, negative, or neutral.

Exam trap

The trap here is that candidates often confuse Key Phrase Extraction or Named Entity Recognition with Sentiment Analysis, because they all involve analyzing text, but only Sentiment Analysis directly outputs positive/negative/neutral labels.

How to eliminate wrong answers

Option A is wrong because Key Phrase Extraction identifies important words or phrases in text but does not evaluate sentiment. Option B is wrong because Named Entity Recognition extracts entities like people, places, or organizations, not sentiment. Option D is wrong because Language Detection identifies the language of the text (e.g., English, Spanish), not the sentiment expressed.

Full explanation →

180

Multi-Selectmedium

Which TWO actions should you take to optimize a custom text classification model in Azure Cognitive Service for Language?

Select 2 answers

A.Ensure that training examples for different labels do not have overlapping content.

B.Use a stratified split of training and testing data.

C.Oversample the minority classes to balance the dataset.

D.Remove all stop words from the training data.

E.Remove examples with neutral sentiment to focus on positive and negative classes.

AnswersA, B

Overlapping content confuses the model.

Why this answer

Overlapping content between labels (e.g., the same text appearing in both 'positive' and 'negative' training examples) confuses the custom text classification model, leading to poor decision boundaries. Azure Cognitive Service for Language uses a multi-class or multi-label classifier that learns distinct patterns for each label; overlapping content introduces ambiguity, reducing precision and recall. Ensuring distinct, non-overlapping training examples per label helps the model learn clear, separable features.

Exam trap

The trap here is that candidates often confuse general data preprocessing techniques (like oversampling or stop word removal) with the specific optimization requirements of Azure Cognitive Service for Language's custom text classification, where the service's internal architecture already handles many of these concerns, and the key optimization is ensuring label distinctness and proper data splitting.

Full explanation →

181

MCQeasy

You are planning to deploy an Azure AI Content Safety solution. What is the primary requirement for using the service?

A.Data must be stored in the same region as the service

B.All users must have Microsoft Entra ID P2 licenses

C.An active Azure subscription

D.The application must be written in C#

AnswerC

Content Safety is an Azure resource, so provisioning it requires an active Azure subscription with billing enabled. Without a subscription you cannot create the resource, obtain keys or endpoint, or call the REST API, making this the fundamental prerequisite.

Why this answer

An active Azure subscription is the primary requirement because Azure AI Content Safety is a cloud-based service that requires a valid subscription for resource provisioning, API access, and billing. Without an Azure subscription, you cannot create the Content Safety resource or authenticate API calls, regardless of other configurations.

Exam trap

The trap here is that candidates often confuse 'primary requirement' with optional or advanced features like data residency (A), licensing (B), or specific programming languages (D), but the fundamental prerequisite is always an active Azure subscription for any Azure AI service.

How to eliminate wrong answers

Option A is wrong because data does not need to be stored in the same region as the service; Azure AI Content Safety processes content in the region where the resource is deployed, but data residency requirements are separate and not a primary prerequisite. Option B is wrong because Microsoft Entra ID P2 licenses are not required; Azure AI Content Safety uses standard Azure AD authentication (free tier) or API keys, and P2 licenses are only relevant for advanced identity protection features unrelated to this service. Option D is wrong because the application does not need to be written in C#; the service is language-agnostic and can be accessed via REST APIs, SDKs in Python, Java, JavaScript, .NET, and other languages.

Full explanation →

182

MCQhard

You are planning an Azure AI solution that will process documents containing personal data. The solution uses Azure AI Document Intelligence and Azure OpenAI in the same subscription. Corporate governance requires that data processed by these services never leave a specified geographic region and that you can audit which operations were performed on the data. You need to select a deployment approach that meets these requirements with the least administrative effort. What should you do?

A.Deploy all resources in the required region and enable diagnostic settings to send logs to a Log Analytics workspace in the same region.

B.Deploy resources in the required region and configure customer-managed keys stored in a key vault in a different region.

C.Deploy resources in the required region and enable soft delete on all storage accounts used by the solution.

D.Deploy resources in multiple regions and use Azure Traffic Manager to route requests to the closest region.

AnswerA

Deploying every resource in the required region keeps data processing within that geography, and diagnostic settings capture resource logs and metrics for auditing. This is a built-in capability that requires minimal ongoing administration while satisfying both the data residency and auditability requirements.

Why this answer

Placing all resources in the required region ensures data is processed within that geography, and diagnostic settings stream resource logs and metrics to Log Analytics for auditing. This uses native platform features with little ongoing effort. Multi-region routing, cross-region key vaults, and soft delete do not satisfy both residency and auditability.

Exam trap

The trap here is confusing data durability features such as soft delete or key management with data residency and audit logging, which are governed by deployment location and diagnostic settings.

Full explanation →

183

MCQeasy

You are building an agent using Microsoft Copilot Studio to handle customer returns. The agent must collect the order ID, reason for return, and then provide a return shipping label. The process requires the user to provide information step-by-step. Which type of conversation flow should you implement?

A.Use an adaptive card to collect all inputs in one step.

B.Use multiple question nodes in a sequential flow.

C.Use a single question node to collect all information at once.

D.Use a generative answers node to parse the user's intent.

AnswerB

Sequential question nodes present one prompt at a time and wait for the user's answer before advancing, which matches the requirement to collect order ID and return reason step-by-step. Each node captures a single value, ensuring ordered data collection before the shipping label is generated.

Why this answer

Microsoft Copilot Studio uses 'Question' nodes to collect user input one piece at a time in a sequential flow, which matches the step-by-step requirement for order ID, reason, and shipping label. This approach ensures each piece of data is validated before moving to the next, maintaining a guided conversation.

Exam trap

The trap here is that candidates might confuse the flexibility of generative answers or adaptive cards with the structured, sequential data collection needed for transactional workflows, overlooking that Copilot Studio's Question nodes are purpose-built for step-by-step input gathering.

How to eliminate wrong answers

Option A is wrong because an adaptive card collects all inputs in one step, which violates the requirement for step-by-step collection and can overwhelm users or miss validation per field. Option C is wrong because a single question node cannot collect multiple distinct pieces of information at once; it only handles one input per node. Option D is wrong because a generative answers node is designed for open-ended Q&A using AI, not for structured data collection with specific fields like order ID and reason.

Full explanation →

184

MCQhard

You are building a generative AI solution with Azure OpenAI Service that uses the GPT-4 model. The solution must process user requests and call external APIs to retrieve real-time data. You need to ensure the model can invoke the correct API based on user intent and return the results in a structured format. Which feature should you implement?

A.Prompt engineering with few-shot examples of API calls.

B.Using the Assistants API with code interpreter enabled.

C.Fine-tuning the model on a dataset of API calls and responses.

D.Function calling with the 'tools' parameter in the Chat Completions API.

AnswerD

Function calling allows the model to output a JSON object specifying which function to call and with what arguments. You define functions in the 'tools' parameter, and the model decides when to invoke them. This enables integration with external APIs and structured data retrieval, directly meeting the requirement.

Why this answer

Function calling is the native feature in Azure OpenAI that allows the model to request invocation of external functions with structured arguments. By defining tools, the model can determine when to call an API and with what parameters. This provides reliable integration with external systems and returns structured responses, fulfilling the requirement.

Exam trap

The trap here is confusing function calling with code interpreter or fine-tuning, which do not provide dynamic, structured API invocation.

Full explanation →

185

Multi-Selectmedium

Which THREE factors should be considered when choosing a region for deploying Azure AI services?

Select 3 answers

A.Number of Azure data centers in the region.

B.Service availability and feature support.

C.Compliance and data residency requirements.

D.Latency to end users.

E.Cost of the services in each region.

AnswersB, C, D

Not all services are available in all regions.

Why this answer

Service availability and feature support (Option B) is a critical factor because not all Azure AI services are available in every region; for example, certain Cognitive Services like Azure OpenAI or Computer Vision OCR may be in preview or fully supported only in specific regions. Choosing a region without the required service or feature would prevent deployment or limit functionality, directly impacting solution design.

Exam trap

Microsoft often tests the misconception that the number of data centers or cost are primary region selection factors, but the exam emphasizes that service availability, compliance, and latency are the three key considerations for Azure AI services.

Full explanation →

186

MCQmedium

You are building an Azure AI Vision solution that analyzes live video from a camera mounted on a delivery truck. The solution must read street signs in real time and return the recognized text with bounding box coordinates. You need to minimize latency and cost. Which Azure AI Vision feature should you use?

A.Azure AI Document Intelligence prebuilt-read model

B.Custom Vision object detection model

C.Read API with asynchronous processing

D.Optical character recognition (OCR) synchronous API

AnswerD

The synchronous OCR API is designed for near real-time, single-image text extraction and returns lines and words with bounding box coordinates. It is ideal for live video frames because you can call it per frame with low latency and pay only for the images you submit. It supports printed text in many languages and gives the positional data required to overlay results on the video.

Why this answer

The synchronous OCR API is the correct choice because it extracts printed text with bounding boxes in a single call, which suits low-latency, real-time video frame analysis. The Read API and Document Intelligence are better for documents and asynchronous processing, and Custom Vision detects objects rather than reading text. Using synchronous OCR minimizes latency and cost for live street sign recognition.

Exam trap

The trap here is assuming that the Read API is always the best OCR choice, when its asynchronous, document-oriented design makes it unsuitable for real-time video frames.

Full explanation →

187

Multi-Selecteasy

Which TWO Azure AI services can be used to build a conversational chatbot that uses generative AI? (Choose two.)

Select 2 answers

A.Azure AI Search

B.Azure AI Bot Service

C.Azure AI Translator

D.Azure AI Language

E.Azure OpenAI Service

AnswersB, E

Azure AI Bot Service provides the conversational orchestration layer, handling channels, turn management and dialog state, while integrating generative AI models through its SDK. It satisfies the stem's requirement for building a chatbot, since the service supplies the bot framework and hosting that generative responses plug into.

Why this answer

Azure AI Bot Service (B) provides a dedicated platform for building, deploying, and managing conversational chatbots, integrating with channels like Microsoft Teams and Slack. Azure OpenAI Service (E) enables generative AI capabilities by providing access to large language models (e.g., GPT-4) that can generate human-like responses, making it ideal for powering the conversational intelligence of a chatbot.

Exam trap

The trap here is that candidates may confuse Azure AI Language (which includes conversational language understanding for intent recognition) with a full chatbot builder, but it lacks the generative AI and channel integration that Azure AI Bot Service and Azure OpenAI Service provide together.

Full explanation →

188

MCQeasy

You are creating an Azure AI Search solution that must extract named entities such as people, organizations, and locations from text documents. You want to use a built-in cognitive skill to perform this extraction during indexing. Which skill should you add to the skillset?

A.KeyPhraseExtractionSkill

B.LanguageDetectionSkill

C.EntityRecognitionSkill

D.TextTranslationSkill

AnswerC

EntityRecognitionSkill is a built-in cognitive skill in Azure AI Search that extracts entities such as persons, organizations, and locations from text. It is specifically designed for this purpose and integrates directly into the enrichment pipeline. Adding it to the skillset will produce enriched output that can be mapped to index fields, meeting the requirement without custom code.

Why this answer

EntityRecognitionSkill is the built-in Azure AI Search skill that extracts named entities and classifies them into categories such as persons, organizations, and locations. It is part of the cognitive skills library and is designed to enrich documents during indexing. The other skills perform different natural language processing tasks such as key phrase extraction, language detection, or translation, none of which provide categorized entity extraction.

Exam trap

The trap here is confusing entity extraction with key phrase extraction, which also analyzes text but does not categorize entities into types like people or organizations.

Full explanation →

189

MCQhard

A logistics company uses Azure AI Vision to analyze images of packages on conveyor belts. They need to detect damaged packages and read tracking numbers. The solution must process high throughput (1000 images per minute) with low latency (<500ms per image). The images are captured by fixed cameras. Which approach should you recommend?

A.Use Azure AI Document Intelligence to process package labels

B.Train a single Custom Vision model that detects damage and reads tracking numbers using OCR

C.Use Azure AI Video Indexer to analyze the video stream from cameras

D.Use Azure AI Vision OCR Read API for tracking numbers and a separate Custom Vision model for damage detection

AnswerD

This approach splits the tasks: the OCR Read API reads tracking numbers, and a Custom Vision model detects damage. Both can run in parallel or be deployed at the edge to meet latency and throughput. This is the only technically feasible option.

Why this answer

It combines Azure AI Vision OCR Read API (for text extraction) with a separate Custom Vision model (for damage detection). This leverages the specialized strengths of each service: the OCR Read API is optimized for text reading with high accuracy, while Custom Vision excels at object detection. The latency requirement (<500ms per image) can be met through parallel processing or edge deployment, and throughput can be achieved by scaling API calls.

Option B is incorrect because Custom Vision does not include built-in OCR capability; it cannot read tracking numbers. Options A and C are less suitable: Document Intelligence is designed for structured documents, not real-time package analysis, and Video Indexer is for video streams, not still images.

Exam trap

The trap is believing that Custom Vision can perform OCR natively. Custom Vision only handles image classification and object detection; text reading requires a dedicated OCR service. Candidates may think a single model is simpler, but it's not technically possible.

How to eliminate wrong answers

Option A is wrong because Azure AI Document Intelligence is designed for structured document extraction (e.g., invoices, forms), not for real-time damage detection on conveyor belt images, and its latency is typically higher than 500ms per image. Option C is wrong because Azure AI Video Indexer is optimized for analyzing pre-recorded video streams with indexing and metadata extraction, not for real-time, low-latency processing of individual images at 1000 images per minute. Option D is wrong because using two separate services (Azure AI Vision OCR Read API for tracking numbers and a separate Custom Vision model for damage detection) introduces additional network round-trips and processing overhead, likely exceeding the 500ms latency budget per image.

Full explanation →

190

MCQhard

You are designing an Azure AI solution that uses Azure AI Search. The solution must ensure that only authorized users can query the search index, and that users can only access documents they are permitted to see. You need to implement document-level access control. What should you do?

A.Enable Azure Private Link for the search service and restrict access to the corporate network.

B.Implement security filters in queries by using the user's identity and a field in the index that maps to allowed groups.

C.Use customer-managed keys (CMK) to encrypt the index and control access to the keys.

D.Use Azure role-based access control (RBAC) to assign the Search Index Data Reader role to users.

AnswerB

Azure AI Search supports document-level security through security filters. You can include a field in the index that contains the allowed groups or users for each document, and at query time, filter results based on the authenticated user's group memberships. This ensures users only see documents they are authorized to access.

Why this answer

To enforce document-level access control in Azure AI Search, you must implement security filters. This involves adding a field to the index that specifies which users or groups can access each document, and then applying a filter in queries based on the authenticated user's identity. This ensures that users only retrieve documents they are permitted to see.

Exam trap

The trap here is assuming that Azure RBAC or network security can provide document-level security; they control service access, not per-document access.

Full explanation →

191

MCQmedium

Your Azure AI Document Intelligence model is failing to extract tables from scanned PDFs. The PDFs are low-quality images. What should you do first?

A.Verify that the Read OCR step is extracting text correctly.

B.Use Azure AI Computer Vision to enhance the image.

C.Retrain the model with more table examples.

D.Use a higher resolution scanner for input PDFs.

AnswerA

Verifying the Read OCR output comes first because Document Intelligence's layout and table extraction depend entirely on that OCR layer. Low-quality scans degrade character and word recognition, so tables cannot be reconstructed reliably. Confirming whether the Read model extracts text correctly isolates the OCR constraint before tuning table extraction.

Why this answer

The Read OCR step is the foundational layer for table extraction in Azure AI Document Intelligence. If the OCR cannot accurately recognize text from low-quality images, subsequent table extraction models will fail regardless of training or image enhancement. Verifying OCR output first isolates whether the issue is at the text recognition stage or the table parsing stage, following a systematic troubleshooting approach.

Exam trap

The trap here is that candidates often jump to retraining or image enhancement without realizing that Document Intelligence's table extraction is entirely dependent on the quality of the OCR output, and the first diagnostic step must be to check that foundational layer.

How to eliminate wrong answers

Option B is wrong because Azure AI Computer Vision image enhancement does not improve OCR accuracy for Document Intelligence; the service already applies its own preprocessing, and external enhancement may introduce artifacts. Option C is wrong because retraining the model with more table examples will not fix the root cause if the OCR step cannot correctly extract text from low-quality images; the model relies on accurate OCR input. Option D is wrong because using a higher resolution scanner is a hardware solution that may not be feasible for existing PDFs and does not address the immediate diagnostic need; the first step should be software-based verification of OCR output.

Full explanation →

192

MCQmedium

A research organization uses Azure AI Language to process large volumes of scientific papers. They need to extract specific entities such as gene names, protein names, and chemical compounds. The entity types are highly specialized and not covered by prebuilt models. The organization has a labeled dataset of 10,000 documents. You need to recommend the most efficient approach to build the entity extraction solution. What should you do?

A.Use the prebuilt NER model and map the recognized entities to the required types.

B.Train a Custom Named Entity Recognition (NER) model using the labeled dataset in Azure AI Language.

C.Use Azure Logic Apps to call the Text Analytics API and post-process the results.

D.Train a custom NER model for genes and use prebuilt NER for chemicals.

AnswerB

Custom NER trains on your labelled dataset to recognise domain-specific entity types such as gene, protein and chemical names that prebuilt models do not cover. With 10,000 labelled documents, it is the most efficient fit for specialised extraction.

Why this answer

Custom Named Entity Recognition (NER) in Azure AI Language allows you to train a model on your own labeled dataset (10,000 documents) to extract highly specialized entity types like gene names, protein names, and chemical compounds that are not covered by prebuilt models. This approach is the most efficient as it leverages the labeled data directly, avoiding the need for complex post-processing or hybrid solutions.

Exam trap

The trap here is that candidates may assume prebuilt NER can be adapted via mapping or post-processing, but Azure AI Language's prebuilt models are fixed and cannot recognize custom entity types without training a custom model.

How to eliminate wrong answers

Option A is wrong because prebuilt NER models only recognize general entity types (e.g., person, location, organization) and cannot be remapped to extract highly specialized scientific entities like gene or protein names without additional training. Option C is wrong because Azure Logic Apps calling the Text Analytics API would still rely on prebuilt NER capabilities, which cannot extract the specialized entities required, and post-processing would be inefficient and error-prone. Option D is wrong because training a custom NER model for genes while using prebuilt NER for chemicals is inconsistent—prebuilt NER does not recognize chemical compounds in a specialized scientific context, and this hybrid approach would require separate handling and likely reduce accuracy.

Full explanation →

193

MCQeasy

You are developing a solution that uses Azure OpenAI to generate customer support responses. You want to prevent the model from repeating the same phrases. Which parameter should you adjust?

A.top_p

B.presence_penalty

C.temperature

D.frequency_penalty

AnswerD

Frequency_penalty applies a proportional penalty to tokens each time they appear, directly reducing verbatim repetition across the generated response. It satisfies the stem's constraint of preventing repeated phrases, unlike presence_penalty, which penalises only first occurrence and encourages new topics rather than curbing recurrence.

Why this answer

The frequency_penalty parameter (option D) is correct because it directly reduces the likelihood of the model repeating the same phrases by penalizing tokens that have already appeared in the generated text. A higher frequency_penalty value (e.g., 0.5 to 1.0) decreases the probability of reusing tokens, making the output more diverse and less repetitive. This is specifically designed to address repetition in generative AI responses.

Exam trap

The trap here is that candidates often confuse presence_penalty with frequency_penalty, but presence_penalty only penalizes tokens that have appeared at least once (regardless of count), while frequency_penalty penalizes based on the actual frequency of occurrence, making it the correct choice for preventing repeated phrases.

How to eliminate wrong answers

Option A is wrong because top_p (nucleus sampling) controls the cumulative probability threshold for token selection, influencing randomness and diversity of output, but it does not specifically penalize repeated phrases. Option B is wrong because presence_penalty penalizes tokens that have appeared at least once in the text, encouraging the model to talk about new topics, but it does not target the frequency of repetition of the same phrases. Option C is wrong because temperature controls the randomness of token selection by scaling the logits before softmax, affecting creativity and variability, but it has no direct mechanism to prevent repetition of phrases.

Full explanation →

194

MCQeasy

Refer to the exhibit. You are calling the Azure AI Language NER API. The response returns no entities. What is the most likely reason?

A.The text does not contain any recognized entities

B.The API version is incorrect

C.The document language should be 'es' for Spanish

D.The endpoint URL is for the wrong region

AnswerA

The NER API returns entities only when it recognises spans matching its trained entity categories. Empty results mean the submitted text contained no such recognisable entities, not a request failure, so the most likely cause is simply that no entities were present.

Why this answer

The NER API returns entities only if the input text contains recognized entity types (e.g., Person, Location, Organization, DateTime, etc.). If no entities are found, the API returns an empty entities array. This is the most straightforward and common reason for a zero-entity response, assuming the request is otherwise valid.

Exam trap

Azure often tests the misconception that a missing or incorrect parameter (like API version, language, or region) would silently return empty results, when in reality those errors manifest as HTTP status codes or error messages, not a successful empty response.

How to eliminate wrong answers

Option B is wrong because an incorrect API version would typically result in an HTTP 400 Bad Request or 404 Not Found error, not a successful response with zero entities. Option C is wrong because the document language parameter is optional; if omitted, the API auto-detects the language, and even if set to 'es', Spanish text would still return entities if present. Option D is wrong because an incorrect region endpoint would cause a connection or authentication failure (e.g., 401 Unauthorized or 403 Forbidden), not a valid response with no entities.

Full explanation →

195

MCQhard

You have an Azure AI Vision resource named MyVisionService. You run the above Azure CLI command and get the keys. Your application uses key1 for authentication. You need to rotate the keys without downtime. What should you do?

A.Delete and recreate the Cognitive Services resource

B.Regenerate key1 immediately and update the application to use the new key1

C.Regenerate both keys at the same time

D.Update the application to use key2, then regenerate key1

AnswerD

Both keys are valid simultaneously, so switching the application to key2 first keeps authentication working while key1 is regenerated. This ordering avoids downtime, whereas regenerating key1 while still in use would immediately break authentication for the running application.

Why this answer

It enables key rotation without downtime. By first updating the application to use key2 (the secondary key), you ensure that authentication continues to work while key1 is being regenerated. After key1 is regenerated, you can optionally update the application back to key1 at a later time.

This pattern is standard for Azure Cognitive Services to maintain continuous access.

Exam trap

The trap here is that candidates may think regenerating the key currently in use is acceptable if done quickly, but Azure explicitly requires using the secondary key to avoid any period of invalid credentials.

How to eliminate wrong answers

Option A is wrong because deleting and recreating the Cognitive Services resource would cause a complete loss of service and all associated configuration, resulting in significant downtime. Option B is wrong because regenerating key1 immediately would invalidate the key currently used by the application, causing authentication failures and downtime until the application is updated with the new key1. Option C is wrong because regenerating both keys at the same time would invalidate all active keys, leaving no valid key for the application to use, causing immediate downtime.

Full explanation →

196

MCQeasy

You are creating an Azure AI Search index for a knowledge mining solution. The index must support searching for documents by a required category field that can have one of five predefined values, and you want to enable faceted navigation on that field. Which index field configuration should you use?

A.Edm.String with searchable set to true and filterable set to false.

B.Edm.String with retrievable set to false and filterable set to true.

C.Edm.String with filterable and facetable set to true, and searchable set to false.

D.Edm.Int32 with filterable set to true and facetable set to false.

AnswerC

For a field used for filtering and faceting on exact values, setting it as Edm.String with filterable and facetable true and searchable false is optimal. This configuration allows exact-match filtering and facet counts without incurring the overhead of full-text analysis, which is unnecessary for predefined category values.

Why this answer

A category field with predefined values is best modeled as a filterable and facetable string field that is not searchable, because full-text analysis is unnecessary and can interfere with exact matching. Making it searchable could tokenize values, and omitting facetable would break faceted navigation. The retrievable attribute should remain true if the value needs to be displayed.

Exam trap

The trap here is enabling searchable on a category field, which tokenizes values and breaks exact filtering and faceting.

Full explanation →

197

MCQmedium

You are building an Azure AI Search solution that enriches documents by detecting the language of each document and then routing content to language-specific analyzers. You add a LanguageDetectionSkill to the skillset and want the detected language code to be available to downstream skills and to be stored in the index. The detected language must be mapped to a field named 'languageCode' in the index. What should you do?

A.Configure the LanguageDetectionSkill with a 'defaultLanguageCode' parameter and set the index field 'languageCode' to use the 'fr.lucene' analyzer.

B.Create a custom skill that calls the Azure AI Language service and writes the detected language directly into the index using the Azure AI Search REST API.

C.Set the 'languageCode' field in the index to be retrievable and filterable, and rely on the indexer to automatically populate it from the skill output.

D.Add an outputFieldMapping in the indexer that maps the '/document/languageCode' enrichment node to the 'languageCode' index field.

AnswerD

Output field mappings in the indexer explicitly connect enriched document nodes to index fields. The LanguageDetectionSkill emits a 'languageCode' value under /document, and an outputFieldMapping with sourceFieldName '/document/languageCode' and targetFieldName 'languageCode' persists it. This is the supported mechanism for projecting skill output into the search index.

Why this answer

The LanguageDetectionSkill outputs a language code under the enriched document, but that value only reaches the index if the indexer maps it. Output field mappings declare which enrichment nodes become index field values. Configuring analyzers or field attributes changes how data is stored or queried, not whether it is stored.

A custom skill is unnecessary because the built-in skill already emits the required value.

Exam trap

The trap here is assuming that enabling a skill and adding a matching index field is enough, when the indexer still needs an explicit output field mapping to persist the enriched value.

Full explanation →

198

Multi-Selectmedium

Which TWO options are valid ways to index content from Azure SQL Database into Azure AI Search? (Select TWO.)

Select 2 answers

A.Use the Push API to send data directly to the search index.

B.Use Azure Data Factory to copy data to Blob Storage, then index from Blob.

C.Use Azure AI Document Intelligence to extract data and push to index.

D.Use Azure Event Hubs to stream data into the search index.

E.Use the Azure AI Search SQL Server indexer.

AnswersA, E

The Push API lets your code serialise SQL rows into JSON documents and POST them to the index, giving full control over transformation and scheduling. It works without an indexer, satisfying the requirement for a valid SQL-to-search ingestion path.

Why this answer

Option A is correct because the Push API (the REST/SDK endpoint that accepts JSON documents directly into an index) lets an application read rows from Azure SQL Database and send them straight to Azure AI Search, bypassing any intermediate storage. Option E is correct because Azure AI Search provides a built-in Azure SQL indexer (created via the portal, REST, or the Azure.Search.Documents SDK) that connects to Azure SQL Database using a change-tracking column (or high-water mark) to pull and index rows on a schedule. Option B is not a direct indexing method for Azure SQL Database; it inserts an unnecessary Data Factory copy to Blob Storage and then relies on a Blob indexer, which is a different data source.

Option C is wrong because Azure AI Document Intelligence extracts text from documents (PDFs, images, forms), not from relational Azure SQL tables. Option D is wrong because Event Hubs is a streaming ingestion service and Azure AI Search has no Event Hubs indexer; streaming into the index would still require the Push API.

Full explanation →

199

MCQmedium

A company plans to deploy a Copilot Studio agent to Microsoft Teams. The agent should be available to all employees in the company. The security team requires that only authenticated users from the company's Microsoft Entra ID tenant can access the agent. Which channel configuration should be used?

A.Publish the agent to the Direct Line channel and embed it in a Teams tab.

B.Publish the agent to the Web channel and share the link in Teams.

C.Publish the agent to the Teams channel and turn off authentication.

D.Publish the agent to the Teams channel and configure authentication to require Microsoft Entra ID with the company's tenant ID.

AnswerD

Publishing to the Teams channel with authentication set to Microsoft Entra ID and the company's tenant ID restricts access to authenticated employees within that tenant, satisfying the security team's requirement that only tenant users reach the agent.

Why this answer

Publishing the Copilot Studio agent to the Teams channel and configuring authentication to require Microsoft Entra ID with the company's tenant ID ensures that only authenticated users from that specific tenant can access the agent. This meets the security requirement by restricting access to the company's Entra ID tenant, while the Teams channel provides native integration for all employees.

Exam trap

The trap here is that candidates may think the Teams channel inherently restricts access to the company's tenant, but without explicitly configuring authentication to require the specific tenant ID, the agent could be accessible to external guests or users from other tenants.

How to eliminate wrong answers

Option A is wrong because the Direct Line channel is designed for custom application integration, not for native Teams distribution, and embedding it in a Teams tab would not enforce the required Entra ID authentication at the channel level. Option B is wrong because the Web channel uses anonymous or generic authentication by default, and sharing a link in Teams does not restrict access to the company's Entra ID tenant. Option C is wrong because turning off authentication on the Teams channel would allow any user, including unauthenticated or external users, to access the agent, violating the security requirement.

Full explanation →

200

Multi-Selectmedium

A company is building a computer vision solution using Azure AI Vision to analyze images of retail shelves. The solution must detect product presence and read expiration dates. Which TWO Azure AI Vision features should be used?

Select 2 answers

A.Face detection

B.Brand detection

C.Object detection

D.Optical Character Recognition (OCR)

E.Image captioning

AnswersC, D

Object detection returns bounding boxes with labels for each product on the shelf, directly satisfying the product-presence requirement. Unlike image classification, which assigns a single label per image, detection localises multiple distinct items, so the solution can confirm which products are present and where before reading their expiration dates.

Why this answer

Object detection (C) is correct because it locates and classifies multiple objects within an image, which is exactly what is needed to determine whether specific products are present on retail shelves. Optical Character Recognition (D) is correct because OCR extracts printed or handwritten text from images, enabling the solution to read expiration dates printed on product packaging. Face detection (A) only identifies human faces and their attributes, which is irrelevant to detecting products or reading dates.

Brand detection (B) identifies known company logos but does not determine product presence or read expiration dates. Image captioning (E) generates a natural-language description of an image and cannot reliably detect specific products or extract date text.

Exam trap

Microsoft Azure often tests the distinction between object detection and image classification or captioning, where candidates mistakenly choose image captioning for product presence instead of object detection, which provides precise localization and identification.

Full explanation →

201

MCQmedium

You are configuring an Azure AI Search indexer to process documents from Azure Blob Storage. The documents include PDFs and Microsoft Word files. You need to extract both text and metadata such as author and creation date. Which indexer configuration should you use?

A.Set the parsingMode to json

B.Set the parsingMode to default

C.Set the parsingMode to delimitedText

D.Set the parsingMode to text

AnswerB

The default parsingMode uses the built-in document cracking capabilities to extract text and metadata from various file formats, including PDF and Microsoft Office files. It automatically detects the file type and uses the appropriate extractor. This mode is designed for exactly this scenario, where you need to process multiple document types and extract both content and metadata fields like author and creation date.

Why this answer

The default parsingMode in Azure AI Search indexers is designed to handle a variety of document formats, including PDF and Microsoft Office files. It uses built-in document cracking to extract text and metadata, such as author and creation date, which are then available for mapping to index fields. Other parsing modes are specialized for JSON, delimited text, or plain text and do not support rich document extraction.

Exam trap

The trap here is assuming that a specific parsing mode like text is needed for text extraction, when the default mode already handles common document formats with metadata.

Full explanation →

202

MCQhard

You are reviewing an ARM template for deploying Azure OpenAI Service. The template includes a deployment for gpt-35-turbo with a capacity of 100. You need to ensure that the deployment uses provisioned throughput instead of standard. What should you modify?

A.Change the sku name to 'ProvisionedManaged'.

B.Remove the raiPolicyName property.

C.Increase the capacity to 200.

D.Change the model format to 'GPT-4'.

AnswerA

Provisioned throughput requires the deployment's SKU name to be 'ProvisionedManaged' rather than 'Standard'; capacity then represents provisioned throughput units. Changing the sku name satisfies the stem's requirement to switch from standard to provisioned throughput deployment.

Why this answer

To use provisioned throughput (PTU) with Azure OpenAI Service, you must set the SKU name to 'ProvisionedManaged' in the ARM template. The default SKU is 'Standard', which uses pay-per-token consumption. Changing the SKU name to 'ProvisionedManaged' tells the resource provider to allocate dedicated throughput capacity for the deployment, ensuring consistent latency and throughput regardless of other workloads.

Exam trap

The trap here is that candidates often think increasing capacity or changing the model version enables provisioned throughput, but the exam tests the specific SKU name 'ProvisionedManaged' as the only way to switch from standard to provisioned throughput in an ARM template.

How to eliminate wrong answers

Option B is wrong because removing the raiPolicyName property does not affect throughput provisioning; it only removes content filtering or responsible AI policies, which are unrelated to capacity allocation. Option C is wrong because increasing capacity to 200 only scales the number of tokens per minute under the current SKU (Standard), but does not change the SKU to provisioned throughput; PTU requires the SKU name change, not just a higher capacity value. Option D is wrong because changing the model format to 'GPT-4' does not enable provisioned throughput; PTU is a SKU-level setting independent of the model version, and GPT-4 can also be deployed with Standard SKU.

Full explanation →

203

MCQhard

Refer to the exhibit. A developer is configuring a QnA Maker skill for a bot. The skill fails to respond to queries. What is the most likely issue?

A.The kbId is not published.

B.The skill is not deployed.

C.The endpointKey is incorrect.

D.The modelUrl points to the authoring API instead of the runtime endpoint.

AnswerD

The URL should be for the runtime (e.g., https://westus.api.cognitive.microsoft.com/qnamaker/v4.0) but the correct runtime endpoint is typically 'https://<your-resource-name>.azurewebsites.net/qnamaker' or similar. The v4.0 authoring API is not used for querying.

Why this answer

The modelUrl uses the v4.0 API, but the endpoint key and kbId are hardcoded and may be invalid or expired. Additionally, the URL should be for the runtime endpoint, not the authoring API.

Full explanation →

204

MCQmedium

Your organization is using Azure AI Search with semantic ranking. Users report that search results are not showing relevant documents at the top. You need to improve relevance. What should you configure?

A.Add synonyms to the index

B.Define a custom scoring profile

C.Enable semantic search configuration on the index

D.Change the index analyzer to a different language analyzer

AnswerC

Semantic ranking only reorders results when a semantic configuration is defined on the index and referenced in the query. Without it, scoring falls back to BM25 keyword relevance, so enabling the semantic configuration is what lifts relevant documents to the top.

Why this answer

Semantic search configuration is required to enable semantic ranking, which uses deep learning models to re-rank search results based on contextual relevance rather than just keyword matching. Without this configuration, the index cannot leverage semantic ranking even if the service tier supports it, so enabling it directly addresses the user's complaint about irrelevant documents appearing at the top.

Exam trap

Microsoft often tests the misconception that enabling semantic ranking is automatic with the service tier, but candidates must explicitly configure a semantic configuration on the index and specify it in the query request to activate the feature.

How to eliminate wrong answers

Option A is wrong because adding synonyms expands query matching but does not re-rank results based on semantic understanding; it only broadens recall, not precision. Option B is wrong because custom scoring profiles operate on lexical term frequency and field weights, not on the deep neural network models that semantic ranking uses to understand query intent. Option D is wrong because changing the index analyzer affects tokenization and language-specific stemming, not the semantic re-ranking stage that determines which documents are most contextually relevant.

Full explanation →

205

MCQeasy

You are building a question answering solution using Azure AI Language. You have a set of frequently asked questions (FAQs) in a Word document. You need to import the FAQs into a project. Which approach should you use?

A.Use Azure AI Document Intelligence to extract QnA pairs.

B.Create a custom question answering project and import the Word document as a source.

C.Use the prebuilt question answering API to parse the document.

D.Use conversational language understanding (CLU) to extract intents and entities.

AnswerB

Custom question answering ingests FAQ documents directly, extracting question-and-answer pairs automatically, so the Word file becomes a project source without manual reformatting. This satisfies the requirement to import existing FAQs into the project rather than authoring them by hand.

Why this answer

Azure AI Language's custom question answering feature allows you to create a project and import a Word document directly as a source, automatically extracting question-and-answer pairs. This is the intended approach for ingesting FAQ documents into a knowledge base.

Exam trap

AI-102 often tests the confusion between Document Intelligence (for form and document extraction) and custom question answering (for FAQ knowledge bases), leading candidates to pick Document Intelligence for QnA import.

How to eliminate wrong answers

Option A is wrong because Azure AI Document Intelligence is for extracting text and structure from documents, not for directly creating QnA pairs in a question answering project. Option C is wrong because the prebuilt question answering API is for querying an existing knowledge base, not for parsing and importing documents. Option D is wrong because Conversational Language Understanding (CLU) is for intent and entity extraction in conversational apps, not for importing FAQ documents into a QnA project.

Full explanation →

206

MCQeasy

You are creating an Azure AI Search index that will store documents enriched with key phrases and sentiment scores. You need to define the index fields to store these enriched values. The key phrases should be searchable and retrievable, and the sentiment score should be filterable and sortable. Which field definitions should you use?

A.keyPhrases: Collection(Edm.String) with searchable=true, filterable=true; sentimentScore: Edm.Double with filterable=true, sortable=true.

B.keyPhrases: Collection(Edm.String) with searchable=true, retrievable=true; sentimentScore: Edm.Double with filterable=true, sortable=true, retrievable=true.

C.keyPhrases: Edm.String with searchable=true, retrievable=true; sentimentScore: Edm.Int32 with filterable=true, sortable=true, retrievable=true.

D.keyPhrases: Collection(Edm.String) with filterable=true, retrievable=true; sentimentScore: Edm.Double with searchable=true, retrievable=true.

AnswerB

Key phrases are typically a collection of strings, so Collection(Edm.String) is appropriate. Marking it searchable and retrievable allows full-text search and retrieval in results. Sentiment scores are numeric, so Edm.Double works, and marking it filterable and sortable enables filtering and sorting. This definition meets all requirements.

Why this answer

Key phrases are a collection of strings, so Collection(Edm.String) is correct. They must be searchable and retrievable to enable full-text search and to return them in results. Sentiment score is a numeric value, so Edm.Double is suitable.

It must be filterable and sortable, and retrievable to be included in results. The combination of these attributes satisfies the scenario's requirements.

Exam trap

The trap here is overlooking the retrievable attribute for key phrases or choosing an incorrect data type for sentiment score.

Full explanation →

207

MCQmedium

A company is using Azure AI Vision to analyze images from a manufacturing line. The solution must detect defects in real-time. The team discovers that the model's accuracy drops significantly when images are captured under different lighting conditions. What is the best approach to improve the model's robustness?

A.Apply image pre-processing to normalize lighting before sending to the model.

B.Increase the number of training images without varying lighting conditions.

C.Retrain the model using images captured under various lighting conditions, using data augmentation.

D.Use a pre-built model from Azure AI Vision instead of a custom model.

AnswerC

Accuracy drops because the model has not learned invariance to illumination, a covariate shift between training and inference data. Retraining with images spanning varied lighting, plus augmentation that synthesises those variations, exposes the model to the real-world distribution it must handle, directly restoring robustness under the differing lighting conditions the stem describes.

Why this answer

The accuracy drop is caused by a domain shift: the model was trained on images with limited lighting variation but is deployed under diverse lighting. Retraining with images captured under various lighting conditions and applying data augmentation (brightness, contrast, exposure jitter) teaches the model to be invariant to those variations, directly improving robustness. This addresses the root cause rather than masking it at inference time.

Exam trap

The trap is choosing inference-time preprocessing (normalize lighting) as a quick fix instead of addressing the training data distribution, which is the actual cause of the robustness gap.

How to eliminate wrong answers

Option A is wrong because pre-processing to normalize lighting is a partial mitigation that can lose defect-relevant detail and does not improve the model's learned invariance; it also adds a fragile hand-tuned step. Option B is wrong because adding more images without varying lighting reinforces the same bias and will not improve performance under new lighting conditions. Option D is wrong because a pre-built Azure AI Vision model is generic and not trained on the company's specific defect classes, so it cannot outperform a properly retrained custom model for this task.

Full explanation →

208

MCQmedium

You are developing a solution to detect defects on a manufacturing assembly line using computer vision. The solution must classify images as 'defective' or 'non-defective'. You have a limited set of labeled images (500 per class). Which approach should you recommend?

A.Use Azure AI Vision Image Analysis with a pre-built model

B.Use Azure AI Custom Vision with image classification

C.Use Azure AI Custom Vision with object detection

D.Train a deep learning model from scratch using Azure Machine Learning

AnswerB

Azure AI Custom Vision image classification handles small labelled datasets through transfer learning, fine-tuning a pre-trained model on your 500 images per class. This satisfies the limited-data constraint directly, unlike object detection, which needs bounding boxes, or training from scratch, which would overfit.

Why this answer

Azure AI Custom Vision with image classification is the best choice because it allows you to fine-tune a pre-trained deep learning model on your limited dataset (500 images per class) to classify images as 'defective' or 'non-defective'. This approach requires minimal data and expertise compared to training from scratch, and it is specifically designed for custom classification tasks with small datasets.

Exam trap

The trap here is that candidates may confuse image classification (assigning a single label to the whole image) with object detection (locating objects), or assume that a pre-built model can be retrained for custom classes, when in fact Azure AI Custom Vision is the correct service for custom classification with limited data.

How to eliminate wrong answers

Option A is wrong because Azure AI Vision Image Analysis pre-built models are designed for general-purpose tasks (e.g., describing images, detecting common objects) and cannot be retrained on custom classes like 'defective' vs 'non-defective'. Option C is wrong because object detection identifies and locates multiple objects within an image, which is overkill for a simple binary classification task where only the presence of a defect matters, not its location. Option D is wrong because training a deep learning model from scratch with only 500 images per class would likely result in poor generalization and overfitting, requiring significantly more data and computational resources.

Full explanation →

209

MCQhard

Your organization uses Azure AI Language for custom text classification. You have deployed a model to a dedicated endpoint. After updating the training data, you retrain and redeploy the model. Users report that the endpoint still returns predictions from the old model. What is the most likely cause?

A.The training data changes are not saved

B.The project needs to be rebuilt from scratch

C.The endpoint has a caching issue

D.The new model is not yet deployed; you must deploy it to the endpoint

AnswerD

Retraining creates a new model version but does not automatically replace the one assigned to a deployed endpoint. The endpoint continues serving its previously assigned deployment until you explicitly deploy the updated model to it, which explains why users still receive old predictions.

Why this answer

In Azure AI Language, retraining a custom text classification model does not automatically update the deployed endpoint. After training, you must explicitly deploy the new model to the endpoint using the 'Deploy model' action. Until that step is completed, the endpoint continues to serve predictions from the previously deployed model.

Exam trap

The trap here is that candidates assume retraining automatically updates the endpoint, but Azure AI Language requires an explicit deployment step to bind the new model to the endpoint.

How to eliminate wrong answers

Option A is wrong because training data changes are automatically saved when you edit the dataset in Azure AI Language; the issue is not about saving but about deployment. Option B is wrong because rebuilding the project from scratch is unnecessary; you can retrain and redeploy the same project without recreating it. Option C is wrong because Azure AI Language endpoints do not have a client-side or server-side caching mechanism that would serve stale model predictions; the endpoint simply returns results from whichever model is currently deployed.

Full explanation →

210

MCQmedium

Your team is building a custom question-answering solution using Azure AI Language. The solution must be able to answer questions based on a set of PDF documents. You need to import the documents and create a knowledge base. What should you do first?

A.Use Azure AI Foundry to create a project and upload the documents

B.Create an index in Azure AI Search and upload the documents

C.Use the Azure AI Language service with the custom question answering feature and import the documents

D.Deploy an Azure AI Bot Service and connect it to the documents

AnswerC

Custom question answering ingests PDF documents as sources and builds the knowledge base from them. Importing the documents into the Azure AI Language custom question answering project is the prerequisite step before training and publishing the knowledge base.

Why this answer

The custom question answering feature of Azure AI Language is specifically designed to ingest documents (including PDFs) and build a knowledge base that can be used for question-answering. This feature provides a built-in pipeline to extract question-answer pairs from documents, create a knowledge base, and deploy it as a service without needing additional search indexing or bot orchestration.

Exam trap

The trap here is that candidates often confuse Azure AI Search (a general-purpose search service) with the custom question answering feature, which is purpose-built for extracting and managing QnA pairs from documents, leading them to choose Option B incorrectly.

How to eliminate wrong answers

Option A is wrong because Azure AI Foundry is a development environment for building and managing AI models, but it does not directly import documents into a question-answering knowledge base; the custom question answering feature is the correct service for this task. Option B is wrong because creating an index in Azure AI Search is used for full-text or vector search, not for the structured question-answer pair extraction and management that custom question answering provides. Option D is wrong because Azure AI Bot Service is a framework for building conversational bots, not a tool for importing documents and creating a knowledge base; the knowledge base must be created first using the custom question answering feature before a bot can consume it.

Full explanation →

211

Multi-Selecthard

Which THREE factors should be considered when choosing between Azure AI Language's pre-built sentiment analysis and custom sentiment analysis for a specialized domain?

Select 3 answers

A.Custom models require a large set of labeled training data.

B.Custom models always have faster response times.

C.The pre-built model may not accurately handle domain-specific jargon.

D.Pre-built models cannot be used in containers.

E.Pre-built models offer multilingual support out-of-the-box.

AnswersA, C, E

Custom sentiment models are trained via Azure AI Language's labelled classification workflow, so a substantial volume of domain-tagged utterances is a prerequisite. This labelled-data overhead is the practical cost that distinguishes custom training from simply calling the pre-built endpoint.

Why this answer

Option A is correct because custom sentiment analysis in Azure AI Language is a fine-tuned model that requires you to provide a substantial set of labeled training data (typically hundreds of labeled utterances per class) so the model can learn domain-specific patterns. Option C is correct because the pre-built sentiment analysis model is trained on general-purpose text, so specialized jargon, acronyms, or industry-specific phrasing in a niche domain may be misclassified, which is a key reason to consider a custom model. Option E is correct because Azure AI Language's pre-built sentiment analysis supports multiple languages out-of-the-box, which is a significant advantage when your data spans several languages and you want to avoid building separate custom models per language.

Option B is not correct because custom models are not inherently faster; latency depends on deployment, and custom models can add overhead compared to the pre-built service. Option D is not correct because pre-built Azure AI Language models can be deployed in containers (for example, via Docker with the Language container images) for on-premises or disconnected scenarios.

Exam trap

The trap here is that candidates may assume custom models are always superior or faster, overlooking the critical requirement for labeled training data and the fact that pre-built models already offer robust multilingual support and container deployment options.

Full explanation →

212

MCQmedium

You are building a generative AI application with Azure OpenAI Service. The application must use your own product catalog data stored in an Azure AI Search index to ground model responses. You need to configure the model deployment so that retrieved documents are automatically injected into the prompt without you manually assembling the context in application code. What should you configure on the model deployment?

A.Add a 'data source' configuration (your Azure AI Search index) to the model deployment by using the Azure OpenAI 'On Your Data' feature.

B.Increase the deployment's tokens-per-minute quota so that the entire product catalog can be sent in each request.

C.Enable content filtering on the deployment to restrict responses to catalog topics.

D.Set the model temperature to 0 so the model always returns deterministic answers from the catalog.

AnswerA

Configuring the model deployment with a data source such as an Azure AI Search index uses the Azure OpenAI On Your Data capability. The service automatically retrieves relevant chunks and injects them into the prompt, and can return citations, so the application does not need to manually assemble context. This matches the requirement for grounding on the product catalog index without custom retrieval code.

Why this answer

The Azure OpenAI On Your Data feature lets you attach a supported data source, such as an Azure AI Search index, directly to a model deployment. At inference time the service performs retrieval, injects the relevant chunks into the prompt, and can return citations. This avoids writing custom retrieval and prompt-assembly logic while grounding responses in the product catalog.

Exam trap

The trap here is assuming that generation settings such as temperature or quota changes can provide grounding, when grounding requires an explicit data source configuration on the deployment.

Full explanation →

213

MCQmedium

You are developing a generative AI application using Azure OpenAI Service. The application must generate summaries of customer emails and then extract action items. You want to minimize the number of API calls and ensure the model outputs structured JSON. What should you do?

A.Fine-tune the model to output JSON with both summary and action items.

B.Use the Azure OpenAI function calling feature to define a function that returns both summary and action items.

C.Use a single prompt that asks the model to summarize the email and extract action items, and specify the output format as JSON in the prompt.

D.Use two separate prompts: one for summarization and one for action item extraction, and combine the results in code.

AnswerC

Combining both tasks into one prompt reduces API calls and latency. By instructing the model to output JSON, you get structured data that is easy to parse. This approach leverages the model's ability to handle multiple instructions in one request, which is efficient and meets the requirement for structured output.

Why this answer

The most efficient method is to use a single prompt that instructs the model to both summarize the email and extract action items, with the output specified as JSON. This reduces API calls to one, lowers latency, and provides structured data for easy parsing. It leverages the model's ability to handle multiple instructions and format outputs as requested.

Exam trap

The trap here is assuming that function calling or fine-tuning is needed for structured output, when prompt engineering with JSON instructions suffices.

Full explanation →

214

MCQeasy

A company wants to build a solution that summarizes long support call transcripts into concise paragraphs. The transcripts are stored as plain text in Azure Blob Storage. You plan to use the summarization feature in Azure AI Language. Which summarization type should you use to generate a short paragraph that captures the main points of each transcript?

A.Extractive summarization

B.Key phrase extraction

C.Named entity recognition

D.Abstractive summarization

AnswerD

Abstractive summarization generates new, concise text that captures the main ideas of the input, rather than extracting existing sentences. This matches the requirement to produce a short paragraph summarizing the call transcripts. The Azure AI Language abstractive summarization feature is designed for exactly this scenario, producing coherent summaries in natural language.

Why this answer

Abstractive summarization in Azure AI Language generates new, concise text that captures the essence of the input, making it suitable for creating short paragraphs from long transcripts. Extractive summarization only selects existing sentences, key phrase extraction returns terms, and named entity recognition identifies entities. Only abstractive summarization produces the rephrased, paragraph-style summary required.

Exam trap

The trap here is assuming that extractive summarization produces a rephrased paragraph, when it actually returns a subset of the original sentences without generating new text.

Full explanation →

215

Multi-Selectmedium

Which THREE actions should an engineer take when deploying a custom question answering project in Azure Cognitive Service for Language?

Select 3 answers

A.Integrate LUIS for intent detection.

B.Set up a multi-turn extraction policy for follow-up questions.

C.Enable active learning to improve answer suggestions.

D.Add chit-chat to handle common conversational phrases.

E.Configure a single-turn extraction policy.

AnswersB, C, D

Multi-turn extraction is needed for conversation flow.

Why this answer

Multi-turn extraction is a core feature of custom question answering that allows the system to handle follow-up questions by maintaining context across turns. This is essential for conversational flows where a user's subsequent query depends on the previous answer, and it is configured via the project settings in Language Studio.

Exam trap

The trap here is that candidates often confuse the need for LUIS integration (Option A) with question answering, not realizing that custom question answering is a standalone service that does not require intent detection from LUIS.

Full explanation →

216

MCQhard

Refer to the exhibit. You deployed a custom model for Language service. Which command should you run to check if the deployment is ready to accept inference requests?

A.az cognitiveservices account deployment list --resource-group myRG --name myLangService

B.az cognitiveservices account deployment delete --resource-group myRG --name myLangService --deployment-name myDeployment

C.az cognitiveservices account deployment show --resource-group myRG --name myLangService --deployment-name myDeployment

D.az cognitiveservices account deployment create --resource-group myRG --name myLangService --deployment-name myDeployment

AnswerC

The deployment show command returns the provisioning state and status of the named custom model deployment. A succeeded provisioningState confirms the deployment is ready to accept inference requests, which is exactly what the check requires.

Why this answer

The `az cognitiveservices account deployment show` command retrieves the current state of a specific deployment, including its provisioning status (e.g., 'Succeeded'). Only when the status is 'Succeeded' can the deployment accept inference requests. This is the correct command to verify readiness before sending any prediction calls.

Exam trap

Azure often tests the distinction between commands that manage resources (create, delete, list) versus those that inspect state (show), and the trap here is that candidates might confuse 'list' (which shows all deployments but not readiness) with 'show' (which gives the specific deployment's status).

How to eliminate wrong answers

Option A is wrong because `az cognitiveservices account deployment list` returns all deployments in the account, not the status of a specific deployment, and does not directly indicate readiness for inference. Option B is wrong because `az cognitiveservices account deployment delete` removes the deployment entirely, which is destructive and unrelated to checking readiness. Option D is wrong because `az cognitiveservices account deployment create` initiates a new deployment or updates an existing one, but it does not check the current state; it is used to create or modify, not to verify.

Full explanation →

217

MCQhard

You are designing an Azure AI solution that uses multiple Azure AI services, including Azure AI Vision and Azure AI Language. You need to ensure that the solution can be deployed in a way that minimizes latency between services and provides a single endpoint for management. What should you use?

A.Deploy a single Azure AI Services resource for each service and use private endpoints to connect them.

B.Deploy each service as a separate resource in the same region and use Azure API Management to aggregate them.

C.Deploy a single Azure AI Services multi-service resource that includes both Vision and Language capabilities.

D.Deploy each service as a separate resource in different regions and use Azure Front Door to route requests.

AnswerC

An Azure AI Services multi-service resource allows you to access multiple AI services, such as Vision and Language, through a single endpoint and key. Deploying this resource in one region ensures that all services are co-located, minimizing network latency between them. It also simplifies management by providing a single resource for billing and access control.

Why this answer

A multi-service Azure AI Services resource bundles multiple AI capabilities into a single Azure resource with one endpoint and key. Deploying it in one region ensures all services are co-located, reducing latency. It also simplifies management by consolidating billing and access control, meeting both requirements.

Exam trap

The trap here is assuming that using API Management or private endpoints alone can achieve a single endpoint and minimal latency, when actually a multi-service resource is the native solution for co-located services with unified management.

Full explanation →

218

MCQmedium

Your team develops a document translation solution using Azure AI Translator. The solution must translate documents while preserving formatting and layout. Which feature should you use?

A.Azure AI Translator Custom Translator

B.Azure AI Translator Document Translation

C.Azure AI Document Intelligence

D.Azure AI Translator Text Translation

AnswerB

Document Translation is the Translator feature that translates whole documents while preserving their original formatting and layout, returning translated files in the source format. Plain text translation APIs discard structure, so they cannot satisfy the layout-preservation constraint.

Why this answer

Azure AI Translator Document Translation is specifically designed to translate entire documents while preserving the original formatting, structure, and layout. Unlike Text Translation, which handles only plain text strings, Document Translation processes files (e.g., PDF, Word, Excel) and returns a translated version with the same formatting, making it the correct choice for this requirement.

Exam trap

The trap here is that candidates often confuse Document Translation with Text Translation, assuming that any translation feature can handle documents, but Text Translation only processes plain text strings and cannot preserve formatting or layout.

How to eliminate wrong answers

Option A is wrong because Custom Translator is a feature for building custom translation models tailored to specific domain terminology, not for preserving document formatting or layout. Option C is wrong because Azure AI Document Intelligence (formerly Form Recognizer) is used for extracting text, key-value pairs, and tables from documents, not for translating them. Option D is wrong because Text Translation only handles plain text strings and cannot preserve the formatting or layout of an entire document.

Full explanation →

219

MCQmedium

You are designing a knowledge mining solution for a medical research organization. The solution must extract relationships between drugs, diseases, and genes from scientific articles. The data will be stored in a knowledge graph for querying. Which Azure AI service should you use for the extraction?

A.Azure AI Search with semantic ranking

B.Azure AI Translator with dictionary lookup

C.Azure AI Document Intelligence custom extraction model

D.Azure AI Language healthcare entity recognition and relation extraction

AnswerD

Healthcare entity recognition extracts drugs, diseases and genes as typed entities, and relation extraction links them, producing exactly the drug-disease-gene relationships the knowledge graph requires. General entity extraction cannot capture these biomedical relation types, so it fails the graph-querying requirement.

Why this answer

Azure AI Language's healthcare entity recognition and relation extraction is specifically designed to extract medical entities (drugs, diseases, genes) and their relationships from unstructured text, making it ideal for building a knowledge graph. This pre-built model uses deep learning trained on biomedical literature, directly supporting the required extraction without custom training.

Exam trap

The trap here is that candidates confuse general-purpose text extraction (Azure AI Document Intelligence) or search (Azure AI Search) with domain-specific biomedical entity and relation extraction, which requires a specialized healthcare NLP model like Azure AI Language's healthcare feature.

How to eliminate wrong answers

Option A is wrong because Azure AI Search with semantic ranking is a search and ranking service, not an extraction service; it cannot extract entities or relationships from text. Option B is wrong because Azure AI Translator with dictionary lookup performs language translation and word-level lookup, not structured entity or relation extraction from scientific articles. Option C is wrong because Azure AI Document Intelligence custom extraction model is designed for extracting fields from forms and documents (e.g., invoices, receipts), not for complex biomedical entity and relation extraction from unstructured narrative text.

Full explanation →

220

MCQeasy

You are building a solution that analyzes customer feedback in real-time using Azure AI Language. The feedback is streamed from a web app and must be processed within 1 second. You need to extract sentiment and key phrases. Which service should you use?

A.Azure AI Language synchronous API

B.Azure AI Language asynchronous API

C.Azure AI Language container

D.Azure AI Language batch processing with Azure Blob Storage

AnswerA

The synchronous API for Azure AI Language is designed for real-time processing of small text inputs, returning results immediately. It supports sentiment analysis and key phrase extraction, and can meet the 1-second latency requirement for short texts. This is the appropriate choice for low-latency, interactive scenarios.

Why this answer

The synchronous API is built for immediate analysis of short texts, making it ideal for real-time sentiment and key phrase extraction from streaming feedback. Asynchronous and batch options introduce delays, and containers are not optimized for low-latency cloud-based streaming.

Exam trap

The trap here is assuming that any Azure AI Language endpoint can handle real-time streaming, when only the synchronous API is designed for low-latency individual requests.

Full explanation →

221

MCQmedium

You are designing an agent that uses Azure AI Search as a knowledge store. The agent must handle multiple languages. Which feature should you configure in Azure AI Search to ensure the agent retrieves relevant results for queries in different languages?

A.Scoring profiles

B.Language analyzers

C.Semantic search

D.Synonym maps

AnswerB

Language analyzers apply language-specific tokenisation, stemming and stop-word rules per field, so indexed content and queries in each language are processed consistently. This lexical matching satisfies the stem's requirement that the agent retrieve relevant results across multiple languages.

Why this answer

Language analyzers in Azure AI Search are specifically designed to handle linguistic variations across different languages, such as stemming, stop word removal, and tokenization rules. By configuring the appropriate language analyzer (e.g., 'en.microsoft' for English or 'fr.microsoft' for French) on a searchable field, the agent can retrieve relevant results for queries in multiple languages because the analyzer processes both the indexed content and the query string using the same language-specific rules.

Exam trap

The trap here is that candidates often confuse semantic search (which improves relevance via AI) with language-specific text processing, assuming semantic search alone can handle multilingual queries, but semantic search still relies on the underlying analyzer for tokenization and cannot perform language-specific stemming or stop word removal.

How to eliminate wrong answers

Option A is wrong because scoring profiles influence the ranking of search results based on fields, functions, or weights, but they do not alter how text is tokenized or stemmed for different languages; they cannot ensure cross-lingual retrieval relevance. Option C is wrong because semantic search improves result relevance by understanding query intent and context using deep learning models, but it does not provide language-specific tokenization or stemming; it works on top of existing analyzers and is not a substitute for language analyzers. Option D is wrong because synonym maps expand queries with equivalent terms but do not handle language-specific linguistic rules like stemming or diacritic normalization; they are language-agnostic and cannot adapt to different languages' morphological structures.

Full explanation →

222

MCQeasy

You are using Microsoft Copilot Studio to create an agent that handles customer support. The agent needs to understand the user's intent from free-text input. Which feature should you use to map user utterances to specific topics?

A.Configure variables to capture user input

B.Add actions to process the input

C.Create custom entities to extract key phrases

D.Define trigger phrases for each topic

AnswerD

Trigger phrases map free-text utterances to specific topics by matching user input against defined example phrases, letting the agent route intent correctly. This is Copilot Studio's native mechanism for intent recognition without custom language models.

Why this answer

In Microsoft Copilot Studio, trigger phrases are the primary mechanism for mapping user utterances to specific topics. When a user types a free-text input, the agent's natural language understanding (NLU) engine compares the input against the defined trigger phrases for each topic. The topic with the highest confidence score based on semantic similarity is triggered, enabling intent recognition without requiring exact keyword matches.

Exam trap

The trap here is that candidates often confuse entity extraction (Option C) with intent recognition, assuming that extracting key phrases is sufficient to understand the user's intent, whereas in Copilot Studio, trigger phrases are the dedicated feature for mapping utterances to topics.

How to eliminate wrong answers

Option A is wrong because configuring variables captures and stores user input after it has been processed, but does not perform intent recognition or map utterances to topics. Option B is wrong because actions (such as calling Power Automate flows or APIs) are used to execute logic after a topic is triggered, not to understand the user's intent from free-text input. Option C is wrong because custom entities extract specific data points (like product names or dates) from utterances, but they do not map the entire utterance to a topic; entities are used within a topic to refine understanding, not to trigger the topic itself.

Full explanation →

223

Drag & Dropmedium

Drag and drop the steps to configure an Azure AI Search index with a custom skill into the correct order.

Drag or tap steps into the slots.

Steps

Order

1Step 1

2Step 2

3Step 3

4Step 4

Why this order

Start by creating the search service, define the index, then create the custom skill, set up the indexer with the skillset, and finally run it.

Full explanation →

224

MCQhard

You are designing an enterprise search solution using Azure AI Search. The solution must index data from multiple sources: SQL Database, SharePoint Online, and custom REST APIs. The search index must support faceted navigation and filtering by metadata such as department and document type. You also need to ensure that updates to source data are reflected in the index within 5 minutes. Which approach should you use?

A.Use the push API to index all data from a custom application that polls all sources.

B.Create a single indexer that reads from all three sources using a data source definition.

C.Use only indexers for all sources by creating a custom indexer for the REST API.

D.Configure indexers for SQL and SharePoint, and use the push API for the REST API. Schedule indexers to run every 5 minutes.

AnswerD

Native indexers cover SQL Database and SharePoint Online, while the push API handles the custom REST source that has no indexer. Scheduling indexers every 5 minutes meets the freshness requirement, and index fields marked filterable and facetable support faceted navigation.

Why this answer

It combines the strengths of indexers (for SQL Database and SharePoint Online, which have native connectors) with the push API for custom REST APIs, which lack a built-in indexer. Scheduling the indexers to run every 5 minutes ensures that updates are reflected within the required latency window, while the push API can be triggered on demand or via a polling mechanism to meet the same 5-minute SLA.

Exam trap

The trap here is that candidates assume a single indexer can handle multiple data sources or that a custom indexer can be built for any source, when in reality each indexer is tied to one specific data source type and custom REST APIs require the push API.

How to eliminate wrong answers

Option A is wrong because using the push API exclusively requires building a custom application to poll all sources, which is unnecessary overhead for SQL and SharePoint when native indexers exist, and it does not leverage Azure AI Search's built-in change tracking and scheduling capabilities. Option B is wrong because a single indexer cannot read from multiple heterogeneous data sources; each indexer is bound to one data source definition, and you must create separate indexers for SQL, SharePoint, and REST APIs. Option C is wrong because Azure AI Search does not support creating custom indexers for REST APIs; the only way to index data from a custom REST API is via the push API, not an indexer.

Full explanation →

225

Multi-Selecteasy

Which THREE components are required to build a custom skill for Azure AI Search enrichment?

Select 3 answers

A.A database to store intermediate results.

B.A Power Automate flow to orchestrate the skill.

C.A web API endpoint that accepts JSON input and returns JSON output.

D.An HTTPS endpoint for the API.

E.A JSON schema defining inputs and outputs.

AnswersC, D, E

A custom skill is invoked by the enrichment pipeline as a web API call, so it must expose an endpoint accepting a JSON request body and returning a JSON response containing the enriched values the skillset consumes.

Why this answer

To build a custom skill for Azure AI Search enrichment, you must expose your logic through a web API endpoint that accepts JSON input and returns JSON output (C), because the skillset invokes the skill via an HTTP POST with a JSON payload and expects a JSON response. The endpoint must be secured with HTTPS (D), since Azure AI Search requires custom skill connections to use HTTPS for secure transport. You also need a JSON schema defining inputs and outputs (E), which describes the expected input fields and output fields so the skillset can map enriched document content correctly.

A database for intermediate results (A) is not required, as the skillset pipeline passes data between skills in memory, and a Power Automate flow (B) is not a supported orchestration mechanism for custom skills.

Exam trap

The trap here is that candidates often think a custom skill requires an orchestration tool like Power Automate or a persistent storage layer, but Azure AI Search's enrichment pipeline handles orchestration natively and only needs a stateless HTTPS endpoint with a defined JSON schema.

Full explanation →

Microsoft Azure AI Engineer Associate AI-102 (AI-102) — Questions 151–225