Microsoft Azure AI Engineer Associate AI-102 AI-102 Questions 76–150 | Page 2/11

MCQmedium

You are deploying a generative AI solution that uses Azure OpenAI Service with your own data stored in Azure AI Search. Users report that answers are sometimes pulled from documents the user is not permitted to see, because the retrieval step searches the entire index. You need to ensure each user only receives answers grounded in documents they are authorized to access, without creating a separate index per user. What should you do?

A.Add a security filter field to the Azure AI Search index and pass the user's group or tenant identifier as an OData filter on each query.

B.Enable Azure OpenAI content filtering at High severity for both prompts and completions.

C.Create a separate Azure OpenAI deployment for each department and route users to the deployment matching their department.

D.Increase the chunk size of the indexed documents so that fewer documents are returned per query.

AnswerA

Azure AI Search supports filterable fields that can be combined with the search query, so indexing an access-control field (for example, allowed groups) and passing the caller's identity as an OData filter restricts retrieval to documents that user may see. This keeps a single shared index while enforcing per-user authorization at query time, which is exactly the requirement. The model then only receives permitted grounding content, so answers cannot cite restricted documents.

Why this answer

Authorization must be enforced where the documents are selected, not where the model generates text. Adding a filterable access-control field to the Azure AI Search index and applying the caller's identity as an OData filter causes the retrieval step to return only permitted chunks, so the model can never ground an answer in a restricted document. This preserves one shared index while honoring per-user permissions.

Exam trap

The trap here is assuming that Azure OpenAI content filtering or a separate model deployment provides document-level authorization, when access control must actually be applied as a filter in the retrieval index.

Full explanation →

MCQmedium

A retail company uses Azure Computer Vision to analyze in-store camera feeds. They recently added a new product line and updated the object detection model. However, the model fails to detect the new products. What should the company do first?

A.Use the pre-built 'products' model from Computer Vision.

B.Increase the confidence threshold in the API call.

C.Retrain the custom object detection model with images of the new products.

D.Recreate the Computer Vision resource in a different region.

AnswerC

Custom object detection models only recognise classes present in their training data, so the newly added product line is invisible to the current model. Retraining with labelled images of the new products teaches the model the new classes, directly resolving the detection failure.

Why this answer

The model fails to detect new products because it was never trained on them. Retraining the custom object detection model with labeled images of the new products is the correct first step, as it updates the model's knowledge to recognize the new product line. Pre-built models or threshold adjustments cannot add new object classes.

Exam trap

The trap here is that candidates may assume a pre-built model or a simple threshold tweak can handle new object classes, when in fact custom object detection requires retraining with labeled examples of the new items.

How to eliminate wrong answers

Option A is wrong because the pre-built 'products' model from Computer Vision is a fixed, general-purpose model that cannot be extended to recognize custom or newly introduced product lines. Option B is wrong because increasing the confidence threshold would only filter out low-confidence detections, not enable detection of entirely new object classes that the model was never trained to recognize. Option D is wrong because recreating the Computer Vision resource in a different region has no impact on the model's ability to detect new products; region selection affects data residency and latency, not model capabilities.

Full explanation →

Multi-Selecteasy

Which TWO capabilities are provided by Azure AI Language's pre-built entity recognition?

Select 2 answers

A.Identifying domain-specific medical terms

B.Identifying names of people

C.Extracting key phrases from text

D.Identifying organization names

E.Determining overall sentiment of the text

AnswersB, D

Pre-built NER in Azure AI Language includes a Person category that returns personal names such as individuals mentioned in text. This is a native pre-built entity type, requiring no custom training, so it directly satisfies the question's requirement for a provided entity recognition capability.

Why this answer

Azure AI Language's pre-built NER (Named Entity Recognition) model extracts general-purpose entities from unstructured text, and its entity categories explicitly include Person, so option B (identifying names of people) is correct. The same pre-built model also includes an Organization category, so option D (identifying organization names) is correct. Both are part of the standard entity taxonomy returned by the NER endpoint without any custom training.

Option A is wrong because domain-specific medical terms require the separate Text Analytics for Health feature, not the general pre-built NER model. Option C is wrong because key phrase extraction is a distinct Azure AI Language capability (the Key Phrase Extraction feature), not entity recognition. Option E is wrong because sentiment analysis is its own Azure AI Language feature, separate from NER.

Exam trap

The trap here is that candidates often confuse the distinct capabilities within Azure AI Language—entity recognition, key phrase extraction, and sentiment analysis—and assume they are all part of the same pre-built entity recognition feature.

Full explanation →

MCQeasy

A team is developing a solution to automatically summarize long documents using Azure AI Language. Which feature should they use?

A.Sentiment analysis.

B.Key phrase extraction.

C.Extractive summarization.

D.Entity recognition.

AnswerC

Extractive summarization selects and returns the most salient existing sentences from the source document verbatim, which suits condensing long documents without generating new wording. Abstractive summarization would paraphrase instead, risking factual drift from the original text.

Why this answer

Extractive summarization is the correct feature because it specifically identifies and extracts the most important sentences from a document to create a concise summary. Azure AI Language's extractive summarization uses a ranking model to score sentences based on relevance and informativeness, directly addressing the requirement to automatically summarize long documents.

Exam trap

The trap here is that candidates often confuse key phrase extraction (which finds important words) with extractive summarization (which extracts entire sentences), leading them to choose Option B instead of the correct feature for document summarization.

How to eliminate wrong answers

Option A is wrong because sentiment analysis determines the overall positive, negative, or neutral sentiment of text, not the extraction of key content for summarization. Option B is wrong because key phrase extraction identifies individual words or short phrases that are important, but it does not produce a coherent summary of sentences or paragraphs. Option D is wrong because entity recognition identifies named entities like people, places, and organizations, but it does not extract or rank sentences to form a summary.

Full explanation →

MCQhard

Based on the exhibit, which entity should you focus on improving by adding more labeled examples?

A.Date

B.OrderNumber

C.All entities need improvement.

D.ProductName

AnswerD

ProductName shows the weakest per-entity precision and recall in the exhibit, so adding labelled examples for it yields the greatest accuracy gain. Improving entities already scoring highly would waste labelling effort without lifting overall model performance.

Why this answer

The exhibit shows that ProductName has a recall of 0.65, which is lower than the recall for Date (0.98) and OrderNumber (0.99). Low recall indicates that the model is missing many true instances of ProductName. Adding more labeled examples specifically for ProductName will help the model learn its patterns better, improving recall and overall performance.

This aligns with the practice of iterative model improvement in custom entity extraction within Azure AI Language.

Exam trap

The trap is that candidates may choose 'All entities need improvement' (Option C) because they overlook the recall scores shown in the exhibit. While ProductName has low recall (0.65), Date and OrderNumber have very high recall (0.98 and 0.99), indicating they are already performing well. The pitfall is failing to compare the scores and identify the one entity with significantly lower recall.

How to eliminate wrong answers

Option A is wrong because Date likely has a high confidence score (as dates follow predictable formats), so adding more labeled examples would yield minimal improvement. Option B is wrong because OrderNumber, like dates, typically follows a structured pattern (e.g., alphanumeric codes), so the model already performs well on it. Option C is wrong because not all entities need improvement; only the entity with the lowest confidence (ProductName) should be prioritized for additional labeling to optimize effort and resources.

Full explanation →

MCQhard

You work for a manufacturing company that uses Azure AI services to automate quality inspection on a production line. You have a Custom Vision object detection model that identifies defects on metal parts. The model was trained on images captured under ideal lighting conditions. However, when deployed in the factory, the model's accuracy drops significantly due to inconsistent lighting and glare. You need to improve the model's robustness without collecting new images from the factory floor. What should you do?

A.Increase the number of training iterations to force the model to learn more features.

B.Apply data augmentation techniques such as brightness, contrast, and blur adjustments to the existing training images.

C.Use higher resolution images for training.

D.Change the model type from object detection to classification.

AnswerB

Augmentation synthesises lighting variation—brightness, contrast and blur—directly from the existing dataset, so the detector learns glare-invariant features without any new factory-floor capture. This satisfies the stem's constraint of improving robustness under inconsistent lighting while collecting no additional images.

Why this answer

Using data augmentation techniques like brightness and contrast adjustments, rotation, and noise injection can simulate various lighting conditions and improve robustness. Option A is wrong because increasing training iterations may overfit to the existing data. Option C is wrong because higher resolution does not address lighting variation.

Option D is wrong because changing the model type does not address the data issue.

Full explanation →

MCQmedium

You need to analyze images to detect objects and read text from documents using a single Azure AI service. Which service should you use?

A.Azure AI Vision

B.Azure AI Document Intelligence (Form Recognizer)

C.Azure AI Custom Vision

D.Azure AI Language Service

AnswerA

Azure AI Vision bundles both capabilities: object detection and OCR via the Image Analysis and Read features. A single resource therefore satisfies the requirement to detect objects and read document text without provisioning separate services.

Why this answer

Azure AI Vision is the correct choice because it provides both image analysis (object detection) and optical character recognition (OCR) for reading text from documents within a single service. Its Read API and Analyze Image API cover both requirements without needing separate services.

Exam trap

The trap here is that candidates often confuse Azure AI Document Intelligence (Form Recognizer) as the only service for text extraction, overlooking that Azure AI Vision also provides OCR for general document text reading.

How to eliminate wrong answers

Option B is wrong because Azure AI Document Intelligence (Form Recognizer) is specialized for extracting structured data from forms and documents, not general object detection in images. Option C is wrong because Azure AI Custom Vision is designed for training custom image classification and object detection models, not for reading text from documents. Option D is wrong because Azure AI Language Service focuses on natural language processing tasks like sentiment analysis and key phrase extraction, not image analysis or OCR.

Full explanation →

Multi-Selectmedium

Which TWO actions should you take to ensure that an Azure AI Language Service custom entity recognition model complies with data privacy regulations?

Select 2 answers

A.Use prebuilt entity recognition models instead of custom

B.Increase the number of training epochs

C.Enable diagnostic logging for audit trails

D.Configure data retention policies to delete data after processing

E.Anonymize or remove PII from training data

AnswersD, E

Configuring retention policies that delete utterance data after processing directly satisfies the data minimisation and storage limitation principles. Azure AI Language stores training utterances and their labelled entities in the project's storage account; purging them post-processing removes personal data from the training pipeline, limiting exposure without affecting the deployed model's inference capability.

Why this answer

Option D is correct because configuring data retention policies to delete data after processing ensures that sensitive text used by the Azure AI Language custom entity recognition model is not stored longer than necessary, directly supporting data minimization and storage-limitation requirements of privacy regulations such as GDPR. Option E is correct because anonymizing or removing PII from training data prevents the model from learning and potentially exposing personal data, reducing privacy risk at the source and aligning with principles of data protection by design. The other options do not address privacy compliance: A (using prebuilt models) changes model type but does not itself enforce privacy controls, B (increasing training epochs) only affects model accuracy/training time, and C (enabling diagnostic logging) may aid auditing but can also create additional personal data retention unless carefully governed, so it is not one of the two required actions here.

Exam trap

The trap here is that candidates confuse operational features like logging or model tuning with data privacy controls, mistakenly thinking audit trails or increased epochs satisfy compliance requirements.

Full explanation →

MCQeasy

You are designing a solution to extract structured data from a large number of handwritten forms. The forms are scanned and stored as images. Which Azure AI feature should you use?

A.Azure AI Vision's image analysis

B.Azure AI Speech to text

C.Azure Bot Service

D.Azure AI Document Intelligence's OCR capability

AnswerD

Azure AI Document Intelligence's OCR capability is built for handwritten text extraction, using models trained on handwriting rather than printed typefaces. It returns structured fields from scanned images, satisfying the stem's requirement to extract structured data from a large volume of handwritten forms without custom model training.

Why this answer

Azure AI Document Intelligence's OCR capability is specifically designed to extract structured data from scanned documents, including handwritten forms. It uses advanced optical character recognition (OCR) and layout analysis to identify text, tables, and key-value pairs, making it the correct choice for this scenario.

Exam trap

The trap here is that candidates often confuse Azure AI Vision's general image analysis with Document Intelligence's specialized OCR, but the key differentiator is that Document Intelligence is purpose-built for extracting structured data from forms and documents, including handwriting.

How to eliminate wrong answers

Option A is wrong because Azure AI Vision's image analysis focuses on describing images, detecting objects, and generating captions, not on extracting structured data from handwritten text. Option B is wrong because Azure AI Speech to text converts spoken audio into text, not written or handwritten content from images. Option C is wrong because Azure Bot Service is a framework for building conversational agents, not a tool for OCR or document data extraction.

Full explanation →

MCQmedium

You are developing an application that processes images of handwritten forms. The forms contain checkboxes that may be checked or unchecked. Which Azure AI service should you use to detect the state of the checkboxes?

A.Azure AI Custom Vision

B.Azure AI Language

C.Azure AI Document Intelligence

D.Azure AI Computer Vision

AnswerC

Azure AI Document Intelligence includes a prebuilt read and custom form models that detect checkbox selection state via the selectionMark field, returning checked or unchecked. This directly satisfies the requirement to determine checkbox state on handwritten forms.

Why this answer

Azure AI Document Intelligence (formerly Form Recognizer) is the correct service because it is specifically designed to extract structured data from documents, including detecting the state of checkboxes (checked or unchecked) in forms. It uses prebuilt models like the 'prebuilt-document' or custom extraction models to analyze form fields and checkbox selections, making it the optimal choice for this task.

Exam trap

The trap here is that candidates often confuse Azure AI Computer Vision's OCR capabilities with Document Intelligence's form-specific extraction, leading them to choose Computer Vision even though it cannot reliably detect checkbox states without additional custom logic.

How to eliminate wrong answers

Option A is wrong because Azure AI Custom Vision is used for training custom image classification and object detection models, not for extracting structured data like checkbox states from forms. Option B is wrong because Azure AI Language focuses on natural language processing tasks such as sentiment analysis, key phrase extraction, and entity recognition, not on visual document analysis or checkbox detection. Option D is wrong because Azure AI Computer Vision provides general image analysis capabilities like OCR and object detection, but it lacks the specialized form understanding and field extraction features needed to reliably detect checkbox states in structured documents.

Full explanation →

MCQhard

You operate a RAG assistant on Azure OpenAI Service that answers questions over a product catalog. The index contains thousands of items, and users complain that answers mix details from unrelated products. You need retrieval to consider semantic meaning and handle paraphrased queries while returning only the most relevant items. What should you implement?

A.Faceted navigation that filters results by product category

B.Scoring profiles that boost documents containing a specific field value

C.Vector search using embeddings generated by an Azure OpenAI embedding model

D.Keyword search using the Azure AI Search simple query syntax

AnswerC

Vector search compares the embedding of the query against embeddings of the catalog items, so semantically similar content ranks highly even when wording differs. This handles paraphrased queries and sharpens relevance, which reduces the chance of pulling details from unrelated products. It is the appropriate retrieval mode for the described semantic matching requirement.

Why this answer

Embedding-based vector search maps both the query and the catalog items into a semantic space, so similarity reflects meaning rather than exact wording. Paraphrased questions then retrieve the right products, and ranking by vector distance keeps unrelated items out of the context. Keyword, facet, and scoring-profile approaches remain term- or attribute-driven and cannot perform semantic matching.

Exam trap

The trap here is assuming that adding filters or scoring boosts delivers semantic understanding, when only embeddings capture meaning.

Full explanation →

MCQmedium

Refer to the exhibit. You submit this request to Azure AI Language's conversational language understanding (CLU) for the 'FlightBooking' project. The model correctly identifies the intent as 'BookFlight' and extracts entities: 'Seattle' as FromCity, 'New York' as ToCity, and 'June 15th' as Date. What is the next step for the application?

A.Call a separate booking API with the extracted entities to complete the reservation.

B.Use the CLU response to directly book the flight via the Azure AI Language service.

C.Prompt the user to rephrase the request because the intent is ambiguous.

D.Send another request to CLU to confirm the booking details.

AnswerA

CLU returns only intent and entity predictions; it performs no booking action itself. The application must pass the extracted FromCity, ToCity and Date values to a separate booking API to complete the reservation, which is the required next step.

Why this answer

After CLU extracts the intent and entities, the application must use those entities to call a separate booking API to complete the reservation. CLU itself does not perform bookings; it only provides language understanding. Option B is incorrect because the CLU response cannot directly book a flight.

Option C is incorrect because the intent is already clearly identified as 'BookFlight' and entities are extracted, so no rephrasing is needed. Option D is incorrect because sending another request to CLU would not confirm booking; confirmation is handled by the booking API.

Full explanation →

MCQmedium

You are implementing a RAG (Retrieval-Augmented Generation) solution using Azure AI Search and Azure OpenAI Service. The solution is returning answers that are not relevant to the user query. What is the most likely cause?

A.The max_tokens parameter is set too high.

B.The chunk size is too small.

C.The index includes too many documents.

D.The relevance score threshold is set too low.

AnswerD

A low relevance score threshold lets weakly matching chunks pass into the prompt, so Azure OpenAI grounds its answer on marginal content and returns off-topic responses. Raising the threshold filters these poor matches, directly addressing the irrelevant-answer symptom described in the stem.

Why this answer

A low relevance score threshold in Azure AI Search allows documents with low semantic or vector similarity to be returned as results. When these poorly matched documents are passed to Azure OpenAI Service for answer generation, the model may produce answers that are not relevant to the user query, as the retrieved context is noisy or unrelated.

Exam trap

The trap here is that candidates often confuse the relevance score threshold with other parameters like max_tokens or chunk size, assuming that irrelevant answers stem from generation limits or indexing granularity rather than retrieval quality.

How to eliminate wrong answers

Option A is wrong because the max_tokens parameter controls the length of the generated response, not the relevance of the retrieved content; setting it too high may cause truncation or cost issues but does not directly cause irrelevant answers. Option B is wrong because a chunk size that is too small typically leads to fragmented or incomplete context, which can reduce answer quality but is less likely to cause completely irrelevant answers compared to a low relevance threshold. Option C is wrong because including too many documents in the index does not inherently cause irrelevant answers; the search query and scoring mechanism determine which documents are retrieved, and a large index can still return relevant results if the threshold and ranking are properly configured.

Full explanation →

MCQeasy

You are planning to use Azure AI Content Safety to moderate user-generated content in a social media application. The solution must detect hate speech and self-harm content. Which Content Safety features should you enable?

A.Severity levels for all categories

B.Hate and self-harm content filters

C.Custom categories for hate speech and self-harm

D.Image moderation

AnswerB

Enabling the hate and self-harm content filters directly satisfies the stem's requirement to detect both categories. Azure AI Content Safety classifies text against these severity-based harm categories, so the moderation pipeline can flag or block the specified content types without additional custom models.

Why this answer

Azure AI Content Safety provides pre-built filters for specific harm categories, including hate speech and self-harm. Enabling the 'Hate and self-harm content filters' directly activates the detection models for these categories, meeting the requirement without needing custom categories or additional features.

Exam trap

The trap here is that candidates might think custom categories are needed for specific harm types like self-harm, but Azure AI Content Safety already includes these as built-in categories, so enabling the pre-built filters is the correct approach.

How to eliminate wrong answers

Option A is wrong because severity levels are a configuration setting within each category filter, not a feature to enable; they adjust sensitivity but don't activate detection for specific categories. Option C is wrong because custom categories are for defining new harm types not covered by built-in filters, but hate speech and self-harm are already supported as standard categories, so custom categories are unnecessary and add complexity. Option D is wrong because image moderation is a separate feature for analyzing visual content, but the question focuses on text-based hate speech and self-harm detection; enabling it alone wouldn't address the text requirement.

Full explanation →

MCQmedium

A developer is building a multilingual chatbot using Azure AI Language. The bot must detect the user's language automatically and route the query to the appropriate language-specific model. Which Azure AI Language feature should the developer use?

A.Translator API.

B.Conversational language understanding (CLU) with multilingual project.

C.Language detection API.

D.Custom text classification model.

AnswerC

The Language detection API returns the detected language and ISO code for input text, letting the bot identify the user's language before routing the query to the matching language-specific model. This satisfies the automatic detection requirement without manual selection.

Why this answer

The Language Detection API is the correct choice because it is specifically designed to identify the language of input text automatically, returning a language code and confidence score. This enables the chatbot to route the query to the appropriate language-specific model without requiring any prior training or configuration. The other options either require explicit language specification or are designed for different tasks like translation or intent classification.

Exam trap

The trap here is that candidates often confuse the Translator API's built-in language detection capability with the dedicated Language Detection API, assuming the Translator API is sufficient, but the exam expects you to choose the feature whose primary purpose matches the requirement—pure language detection—rather than a multi-purpose tool.

How to eliminate wrong answers

Option A is wrong because the Translator API is used for translating text from one language to another, not for detecting the source language; it does include language detection as a side feature, but its primary purpose and billing model are centered on translation, making it an indirect and less efficient choice for pure detection. Option B is wrong because Conversational Language Understanding (CLU) with a multilingual project is designed to understand intents and entities across multiple languages, but it requires the user to specify the language or rely on a separate detection step; it does not natively perform automatic language detection on raw input. Option D is wrong because Custom Text Classification is a supervised learning feature that requires labeled training data to classify text into custom categories; it is not designed for language identification and cannot detect languages without extensive training on language-labeled datasets.

Full explanation →

MCQhard

You are developing a bot using Microsoft Bot Framework and Azure AI Language. The bot must handle user intents that change mid-conversation. Which feature should you implement?

A.Prompt dialogs

B.Waterfall dialogs

C.Adaptive dialogs

D.QnA Maker knowledge base

AnswerC

Adaptive dialogs satisfy the mid-conversation intent change by dynamically evaluating language understanding results at each turn, allowing the dialog stack to be restructured or interrupted without restarting the conversation. Their event-driven, declarative model handles context switches that rigid waterfall dialogs cannot, directly meeting the stem's requirement.

Why this answer

Adaptive dialogs are designed for dynamic, event-driven conversations where user intents can change mid-conversation. They use a trigger-based model (e.g., onIntent, onTurn) that allows the bot to react to new intents at any point, unlike linear dialog models. This makes them ideal for handling mid-conversation intent shifts without requiring predefined dialog flows.

Exam trap

The trap here is that candidates often confuse waterfall dialogs (which are sequential and rigid) with adaptive dialogs (which are event-driven and flexible), assuming any dialog can handle mid-conversation changes, but only adaptive dialogs support dynamic interruption and re-routing.

How to eliminate wrong answers

Option A is wrong because prompt dialogs are simple, reusable components for collecting a single piece of input (e.g., text, number) and do not handle intent changes mid-conversation. Option B is wrong because waterfall dialogs follow a fixed, sequential step-by-step flow and cannot dynamically redirect to a different intent once started; they are designed for linear, predictable interactions. Option D is wrong because QnA Maker knowledge base is a question-answering service that matches user queries to predefined Q&A pairs; it does not manage conversational state or handle intent routing.

Full explanation →

MCQmedium

You are developing an Azure AI Language solution to analyze customer support tickets. Each ticket has a subject and a description. You need to automatically classify tickets into categories (e.g., 'billing', 'technical', 'account') and extract the product name mentioned. You have a labeled dataset of 10,000 tickets with category labels and product name annotations. The solution must be cost-effective and easy to retrain as new categories emerge. You want to use a single Azure AI Language resource. Which approach should you use?

A.Use custom text classification for category and key phrase extraction for product name.

B.Use conversational language understanding (CLU) to handle both classification and entity extraction in a single model.

C.Use custom text classification for category and custom named entity recognition for product name extraction.

D.Use custom text classification for category and prebuilt named entity recognition for product name extraction.

AnswerC

Both custom text classification and custom named entity recognition can be trained on the labeled dataset. They can be used within the same Azure AI Language resource, making the solution cost-effective and easy to retrain. This is the best approach.

Why this answer

The most cost-effective and retrainable approach is to use custom text classification for categorizing tickets and custom named entity recognition (NER) for extracting product names. Both are part of Azure AI Language and can be trained on your labeled dataset. Custom NER allows you to define your own entity types (e.g., product name) and is more accurate than prebuilt NER for domain-specific products.

Using a single Azure AI Language resource supports both features.

Exam trap

AI-102 often tests the choice between custom and prebuilt models; candidates may choose prebuilt NER for product names, not realizing that custom NER is needed for domain-specific entities.

How to eliminate wrong answers

Option A is wrong because key phrase extraction is unsupervised and may not accurately extract product names; it's not trainable on your specific annotations. Option B is wrong because Conversational Language Understanding (CLU) is designed for conversational apps (intents and entities) and may not be cost-effective or easy to retrain for this classification task; it's overkill and not optimized for document classification. Option D is wrong because prebuilt NER may not recognize your specific product names, especially if they are not common; custom NER is needed for domain-specific entities.

Full explanation →

MCQmedium

You are building an internal knowledge assistant with Azure OpenAI Service. Responses must be grounded in a curated set of HR policy documents stored in Azure AI Search, and every response must include a citation to the specific source chunk. You already deployed a GPT-4o model and created the search index. You need to add the grounding layer with the least development effort. What should you do?

A.Increase the model temperature and max tokens so the model has more room to recall the HR policies from its pretrained knowledge.

B.Call the Azure AI Search REST API from the client, concatenate the top results into the system message, and post-process the model output to append source links.

C.Fine-tune the GPT-4o deployment on the HR policy documents, then rely on the model to quote the correct policy section.

D.Enable the "On Your Data" feature on the model deployment and point it at the Azure AI Search index.

AnswerD

On Your Data is a built-in Azure OpenAI capability that connects a model deployment directly to an Azure AI Search index, retrieving relevant chunks and returning citations automatically. Because the index already exists, this requires only configuration rather than custom retrieval code, satisfying the least-effort requirement while still grounding every answer in the curated HR documents.

Why this answer

Grounding a deployment in a private index with automatic citations is exactly what the On Your Data capability provides for Azure OpenAI. Since the Azure AI Search index is already built, enabling this feature on the deployment satisfies the grounding and citation requirements with configuration instead of custom code, which matches the least-effort constraint.

Exam trap

The trap here is assuming that fine-tuning or larger token limits can substitute for retrieval when the requirement is verifiable citations from a private document set.

Full explanation →

MCQmedium

You notice a spike in errors (HTTP 429) on a specific day. What is the most likely cause?

A.Network connectivity issues.

B.The number of calls exceeded the rate limit for the service tier.

C.Authentication tokens expired.

D.Invalid API keys were used.

AnswerB

HTTP 429 signals throttling: the subscription or resource tier's requests-per-second or token quota was exceeded, so Azure rejects calls until the window resets. A spike in call volume against a fixed tier limit is the classic trigger, distinguishing it from authentication (401) or server faults (500).

Why this answer

HTTP 429 (Too Many Requests) is a rate-limiting response that occurs when the number of API calls exceeds the allowed threshold for the service tier. In Azure AI services, each pricing tier has a specific requests-per-second (RPS) or requests-per-minute (RPM) limit, and exceeding this limit triggers a 429 error to protect backend resources.

Exam trap

In Azure AI services, HTTP 429 indicates rate limiting rather than service unavailability. Candidates often confuse 429 with 503 (service unavailable) or authentication errors (401/403).

How to eliminate wrong answers

Option A is wrong because network connectivity issues typically result in HTTP 4xx/5xx errors like 503 (Service Unavailable) or 504 (Gateway Timeout), not 429 which is explicitly a rate-limit response. Option C is wrong because expired authentication tokens cause HTTP 401 (Unauthorized) errors, not 429. Option D is wrong because invalid API keys result in HTTP 403 (Forbidden) or 401 errors, not 429.

Full explanation →

MCQhard

You run the Azure CLI command shown in the exhibit. After a few minutes, the deployment fails with a quota error. What is the most likely cause?

A.The SKU name 'Standard' is invalid for Azure OpenAI deployments.

B.The model version '0613' is deprecated and no longer available.

C.The requested capacity of 10 exceeds the available quota for the gpt-4 model in that region.

D.The resource group name 'myResourceGroup' does not exist.

AnswerC

The deployment failed because the requested capacity of 10 exceeds the available quota for the gpt-4 model in that region. Azure OpenAI enforces per-model, per-region capacity quotas, so requesting more than the allotted units triggers a quota error.

Why this answer

The quota error indicates that the requested capacity (10 units) for the gpt-4 model exceeds the available quota in the target region. Azure OpenAI deployments require sufficient model-specific quota, which is region- and model-specific. The error is not related to SKU name validity, model version deprecation, or resource group existence.

Exam trap

The trap here is that candidates might confuse a quota error with a model deprecation or SKU issue, but the error message's explicit mention of 'quota' directly points to capacity limits, not configuration or availability problems.

How to eliminate wrong answers

Option A is wrong because 'Standard' is a valid SKU name for Azure OpenAI deployments; the error message specifically mentions quota, not an invalid SKU. Option B is wrong because model version '0613' is a valid and available version for gpt-4; deprecation would produce a different error (e.g., 'ModelNotFound'), not a quota error. Option D is wrong because if the resource group did not exist, the Azure CLI would fail immediately with a 'ResourceGroupNotFound' error, not after several minutes with a quota error.

Full explanation →

MCQmedium

Refer to the exhibit. You are reviewing a Bicep template for deploying an Azure AI Language resource. After deployment, you need to ensure that the resource uses a private endpoint to block public access. Which additional resource should you include in the template?

A.A service endpoint for Microsoft.CognitiveServices

B.A virtual network peering connection

C.A virtual network gateway

D.A Private Endpoint resource linked to the Cognitive Services account

AnswerD

A Private Endpoint resource creates a private IP address within your virtual network and connects it to the Cognitive Services account via a private link, removing public exposure. This satisfies the requirement to block public access after deployment.

Why this answer

A Private Endpoint resource, when linked to the Cognitive Services account via the `privateLinkServiceId` property, assigns a private IP address from a virtual network to the Azure AI Language resource. This blocks all public access by default when the resource's `publicNetworkAccess` property is set to 'Disabled', ensuring traffic only flows over the Microsoft backbone network through Azure Private Link.

Exam trap

The trap here is that candidates confuse service endpoints (which only filter source traffic but leave the public endpoint active) with private endpoints (which completely remove public accessibility), leading them to incorrectly select Option A.

How to eliminate wrong answers

Option A is wrong because a service endpoint for Microsoft.CognitiveServices does not block public access; it only restricts source traffic to a specific virtual network subnet while still allowing public endpoints to be reachable from the internet. Option B is wrong because virtual network peering connects two virtual networks but does not provide a private IP address or block public access to an Azure AI resource. Option C is wrong because a virtual network gateway is used for site-to-site VPN or ExpressRoute connections, not for creating a private endpoint to an Azure PaaS service.

Full explanation →

MCQhard

An agent built with Azure AI Foundry Agent Service performs a long-running operation by calling a function tool that starts a batch job. The batch job can take up to 30 minutes, and the agent currently times out because the function tool waits for completion. The team wants the agent to continue the conversation after the job finishes. Which pattern should the team implement?

A.Increase the function tool's HTTP client timeout to 30 minutes so the call waits for the job

B.Have the function tool return immediately with a job identifier, then resume the agent when the job completes via a new run or thread message

C.Move the batch job into the code interpreter tool so it runs inside the agent's session

D.Enable parallel tool calls so the agent can start the job and keep responding at the same time

AnswerB

Returning a job identifier lets the tool call finish quickly, so the run is not blocked. When the batch job completes, an external trigger can submit the result back to the agent, either by starting a new run or by adding a message to the thread. This asynchronous pattern matches the requirement that the agent continue the conversation after a long operation finishes.

Why this answer

Long-running work should be decoupled from the agent run. Returning a job identifier immediately keeps the tool call fast, and an external completion event can then deliver the result back to the agent by starting a new run or posting to the thread. Blocking the tool call, moving the job into code interpreter, or enabling parallel calls all fail to provide a reliable continuation path after the job finishes.

Exam trap

The trap here is assuming a longer timeout solves long-running work, when the real need is an asynchronous callback that resumes the agent.

Full explanation →

MCQeasy

You are deploying an Azure AI solution that uses multiple Azure AI services. You need to monitor the solution and receive alerts when the service quota for transactions per second (TPS) is exceeded. What should you use?

A.Azure Service Health

B.Azure Advisor

C.Azure Cost Management

D.Azure Monitor metrics and alerts

AnswerD

Azure Monitor collects metrics from Azure AI services, including transaction counts and throttling events. You can create alert rules based on metrics such as TotalCalls or Throttling to notify when TPS limits are approached or exceeded. This provides proactive monitoring and alerting.

Why this answer

Azure Monitor metrics and alerts allow you to track the transaction rate and configure alerts when thresholds are breached. By using metrics such as TotalCalls or Throttling, you can receive notifications when TPS limits are exceeded, enabling timely action.

Exam trap

The trap here is confusing service-level monitoring with resource-level monitoring; Azure Service Health reports on Azure-wide issues, not your resource's quota usage.

Full explanation →

MCQeasy

A university is developing an app for students to take photos of handwritten notes and convert them to digital text. The app must support multiple languages including English and Spanish. The solution should use a pre-built AI service. Which Azure service should you use?

A.Azure AI Document Intelligence with a custom model

B.Azure AI Vision Read API (OCR)

C.Azure AI Language with custom entity recognition

D.Custom Vision with a custom handwriting recognition model

AnswerB

Azure AI Vision's Read API performs optical character recognition on images, extracting handwritten and printed text. It supports multiple languages including English and Spanish, and is a pre-built service, meeting the university's requirements without custom model training.

Why this answer

Azure AI Vision's Read API (OCR) is a pre-built service specifically designed to extract printed and handwritten text from images, and it supports multiple languages including English and Spanish out of the box. It requires no training and is the correct choice for a general-purpose handwriting-to-text scenario.

Exam trap

AI-102 often tests the confusion between OCR services (Vision Read API) and document-understanding services (Document Intelligence), tricking candidates into picking Document Intelligence for simple handwriting OCR.

How to eliminate wrong answers

Option A is wrong because Azure AI Document Intelligence with a custom model requires labeled training data and is intended for structured document extraction (forms, invoices), not general handwriting OCR. Option B is correct. Option C is wrong because Azure AI Language with custom entity recognition extracts entities from text, not text from images — it does not perform OCR.

Option D is wrong because Custom Vision is an image classification/object detection service, not a handwriting recognition service, and training a custom model for handwriting is unnecessary when the Read API already handles it.

Full explanation →

100

Multi-Selectmedium

Which TWO actions should you take to reduce the latency of an Azure AI Computer Vision OCR call on a large image?

Select 2 answers

A.Use a CPU-bound compute instance.

B.Resize the image to a smaller resolution before calling the API.

C.Increase the API timeout value.

D.Use the Read API asynchronously.

E.Deploy the Cognitive Services container on-premises.

AnswersB, D

Smaller images process faster.

Why this answer

Options B and D are correct. B: Resizing reduces processing time. D: Using the Read API asynchronously allows the client to poll, avoiding timeout.

A: Increasing timeout doesn't reduce latency. C: Using CPU doesn't help. E: Cognitive Services container on-premises might add network latency.

Full explanation →

101

Multi-Selectmedium

Which THREE are valid uses of Azure AI Vision Image Analysis 4.0? (Select three.)

Select 3 answers

A.Extract printed text from an image using OCR

B.Transcribe spoken audio from a video file

C.Detect objects in an image and return bounding boxes

D.Generate a human-readable caption for an image

E.Translate text found in an image to another language

AnswersA, C, D

Image Analysis 4.0 includes a Read OCR feature that extracts printed and handwritten text from images via the Read API, satisfying the OCR use case. It returns text lines and words with bounding polygons, so printed text extraction is a supported capability of this service.

Why this answer

Option A is correct because Image Analysis 4.0 includes a Read OCR feature that extracts printed and handwritten text from images and documents, returning lines and words. Option C is correct because the object detection capability identifies objects in an image and returns their coordinates as bounding boxes along with confidence scores. Option D is correct because Image Analysis 4.0 can generate captions (and dense captions) describing the content of an image in natural language.

Option B is not valid because audio transcription is a speech service capability (e.g., Azure AI Speech), not Image Analysis. Option E is not valid because translating text is performed by Azure AI Translator; Image Analysis only extracts the text via OCR, it does not translate it.

Exam trap

The trap here is that candidates often confuse the capabilities of Azure AI Vision with those of Azure AI Speech or Azure AI Translator, assuming Image Analysis can handle audio transcription or text translation when it strictly processes visual content only.

Full explanation →

102

MCQeasy

You are developing a generative AI application that must comply with responsible AI principles. Which Azure AI service should you use to detect and filter harmful content in both input prompts and output responses?

A.Microsoft Purview

B.Azure AI Content Safety

C.Azure OpenAI Service

D.Azure AI Language

AnswerB

Azure AI Content Safety provides dedicated moderation APIs that classify harmful content across categories such as hate, violence and self-harm, applying them to both prompts and generated responses. This directly satisfies the responsible AI requirement to detect and filter harmful content at input and output, which general language services do not cover.

Why this answer

Azure AI Content Safety is the dedicated Azure service for detecting and filtering harmful content such as hate speech, violence, self-harm, and sexual content in both user prompts and AI-generated responses. It provides configurable severity levels and integrates directly with generative AI workflows to enforce responsible AI policies, making it the correct choice for this requirement.

Exam trap

Microsoft often tests the distinction between a service that provides AI capabilities (Azure OpenAI Service) and a service that enforces safety policies (Azure AI Content Safety), leading candidates to mistakenly choose the model provider instead of the dedicated safety tool.

How to eliminate wrong answers

Option A is wrong because Microsoft Purview is a data governance and compliance service focused on data classification, labeling, and auditing, not on real-time content safety filtering of AI inputs and outputs. Option C is wrong because Azure OpenAI Service provides the generative AI models themselves but does not include built-in content filtering; it relies on separate services like Azure AI Content Safety or its own content filters for safety. Option D is wrong because Azure AI Language offers natural language processing capabilities such as sentiment analysis, key phrase extraction, and language understanding, but it does not specialize in detecting or filtering harmful content in generative AI contexts.

Full explanation →

103

MCQhard

You are building a custom text classification solution in Azure AI Language. You have a dataset with 10 categories and 1000 labeled documents. You need to choose the best project type. What should you use?

A.Conversational Language Understanding (CLU)

B.Key Phrase Extraction

C.Prebuilt Text Classification API

D.Custom text classification (single or multi-label)

AnswerD

Custom text classification supports both single-label and multi-label projects, letting each document map to one or several of the ten categories. This flexibility matches the dataset's structure, whereas other Azure AI Language project types cannot assign categories to documents.

Why this answer

Custom text classification (single or multi-label) is the correct project type because you have a labeled dataset with 10 categories and need to train a model to classify text into those specific categories. Azure AI Language provides a custom text classification feature that allows you to train a model using your own labeled data, supporting both single-label and multi-label classification scenarios. This is the only option that enables you to build a bespoke classifier tailored to your 10-category dataset.

Exam trap

The trap here is that candidates often confuse Conversational Language Understanding (CLU) with custom text classification, but CLU is specifically for conversational flows (intents and entities) and cannot be used for general document-level classification tasks.

How to eliminate wrong answers

Option A is wrong because Conversational Language Understanding (CLU) is designed for intent classification and entity extraction in conversational contexts (e.g., chatbots), not for general text classification with a fixed set of categories. Option B is wrong because Key Phrase Extraction is an unsupervised feature that extracts key terms from text, not a classification model that assigns predefined labels. Option C is wrong because the Prebuilt Text Classification API only supports a fixed set of built-in categories (e.g., sentiment, language detection) and cannot be trained on your custom 10-category dataset.

Full explanation →

104

MCQhard

You are deploying a conversational AI solution using Microsoft Copilot Studio. The solution must comply with organizational data loss prevention (DLP) policies by preventing sensitive data from being sent to the underlying Azure OpenAI model. What should you configure?

A.Configure content filters in Azure OpenAI Studio

B.Define DLP policies in Microsoft 365 compliance center and apply to Copilot Studio

C.Enable Azure AI Content Safety in the bot's generative AI configuration

D.Set the temperature parameter to 0 to reduce variability

AnswerB

DLP policies in M365 can block sensitive data from being sent to AI models.

Why this answer

Microsoft Copilot Studio integrates with Microsoft 365 DLP policies to prevent sensitive data from being sent to the underlying Azure OpenAI model. By defining DLP policies in the Microsoft 365 compliance center and applying them to Copilot Studio, you can enforce data loss prevention rules that block or restrict the transmission of sensitive information (e.g., credit card numbers, social security numbers) to the generative AI backend. This ensures compliance with organizational security requirements without modifying the AI model itself.

Exam trap

The trap here is that candidates confuse Azure AI Content Safety (which handles harmful content moderation) with DLP policies (which handle sensitive data protection), leading them to select Option C instead of the correct DLP-based approach.

How to eliminate wrong answers

Option A is wrong because content filters in Azure OpenAI Studio are designed to filter harmful or offensive content in model outputs, not to prevent sensitive data from being sent to the model as input; they operate on the response side, not the request side. Option C is wrong because Azure AI Content Safety is a service for detecting and filtering harmful content (e.g., hate speech, violence) in both inputs and outputs, but it does not enforce DLP policies or block sensitive data based on organizational compliance rules; it focuses on safety, not data loss prevention. Option D is wrong because setting the temperature parameter to 0 reduces the randomness of the model's responses, making them more deterministic, but it has no effect on preventing sensitive data from being sent to the model; it controls output variability, not input filtering.

Full explanation →

105

MCQeasy

Your team is building an internal knowledge assistant with Azure OpenAI Service. Legal requires that every generated answer include a traceable reference to the source document so reviewers can verify claims. The documents already live in Azure AI Search, and you want the service to return citation metadata alongside the generated text. Which feature should you configure?

A.Azure OpenAI content filtering with annotating models enabled on the deployment.

B.A system message instructing the model to always mention the file name it used to answer.

C.Logging every request and response to Azure Monitor and querying the logs after the fact.

D.The On Your Data feature configured with an Azure AI Search data source, which returns citations with the response.

AnswerD

On Your Data connects the model to an Azure AI Search index and returns the generated answer together with citation entries that identify the retrieved documents and their content. This gives reviewers the traceable references legal requires without building custom retrieval plumbing. Because the documents already reside in Azure AI Search, this configuration fits the existing environment and produces the attribution metadata in the response payload.

Why this answer

On Your Data integrates an Azure AI Search index as the grounding source and returns the model's answer together with citation information identifying the retrieved documents. Because the documents already live in Azure AI Search and the requirement is traceable references in the output, this feature directly provides the needed attribution without custom retrieval code.

Exam trap

The trap here is assuming that a prompt instruction or diagnostic logging yields verifiable citations, when citation metadata must come from the retrieval integration that actually tracks which documents grounded the answer.

Full explanation →

106

Multi-Selecteasy

Which TWO components are required to create a custom text classification model in Azure AI Language?

Select 2 answers

A.A set of labeled documents

B.A QnA Maker knowledge base

C.A project in Azure AI Language

D.A Language Understanding (LUIS) app

E.An Azure Functions app

AnswersA, C

Custom text classification is a supervised task, so the model learns decision boundaries from human-provided examples. Labelled documents supply those intent or class annotations; without them training cannot begin, making this a mandatory component alongside the Azure AI Language project.

Why this answer

Option A (a set of labeled documents) is correct because custom text classification in Azure AI Language is a supervised learning feature that requires training data in the form of documents tagged with the custom categories (labels) you want the model to learn. Option C (a project in Azure AI Language) is correct because you must create a custom text classification project in Azure AI Language (via Language Studio or the REST API) to hold your dataset, labels, training configuration, and deployed model. Option B (a QnA Maker knowledge base) is incorrect because QnA Maker is for building question-and-answer bots, not for training custom classification models.

Option D (a Language Understanding (LUIS) app) is incorrect because LUIS is a separate conversational language understanding service for intents and entities, not custom text classification. Option E (an Azure Functions app) is incorrect because Azure Functions is a serverless compute service and is not a required component for creating or training a custom text classification model.

Exam trap

The trap here is that candidates often confuse the required components for custom text classification with those for other Azure AI Language features (like custom question answering or conversational language understanding), leading them to select QnA Maker or LUIS as plausible options when they are not applicable.

Full explanation →

107

MCQhard

Your Azure AI Language custom entity recognition model incorrectly extracts 'Microsoft' as an organization when it refers to the company, but fails to extract 'Microsoft' as a product when it refers to the software. How should you improve the model?

A.Reduce the amount of training data to avoid confusion

B.Remove the 'Organization' entity type from the model

C.Add more training sentences without labeling the entity type

D.Label 'Microsoft' as both 'Organization' and 'Product' in different training sentences with appropriate context

AnswerD

Custom entity recognition learns from labelled context, so the same surface form can map to different entity types. Labelling 'Microsoft' as Organization in company-context sentences and Product in software-context sentences teaches the model to disambiguate by surrounding words.

Why this answer

Custom entity recognition models in Azure AI Language learn to distinguish entity types based on context. By labeling 'Microsoft' as 'Organization' in sentences where it refers to the company and as 'Product' in sentences where it refers to the software, you provide the model with the contextual clues needed to disambiguate the same token across different uses. This supervised learning approach directly addresses the model's failure to recognize the product entity.

Exam trap

The trap here is that candidates may think reducing data or removing entity types simplifies the problem, but Azure AI Language models require diverse, labeled examples with context to handle polysemy (same word, different meanings).

How to eliminate wrong answers

Option A is wrong because reducing training data would likely worsen model performance by removing valuable examples, not resolve the ambiguity. Option B is wrong because removing the 'Organization' entity type would prevent the model from correctly extracting 'Microsoft' as an organization, which is a valid extraction in many contexts, and does not solve the product extraction issue. Option C is wrong because adding training sentences without labeling the entity type provides no supervised signal for the model to learn the distinction between 'Organization' and 'Product' for the same token.

Full explanation →

108

MCQeasy

A company wants to build a conversational interface that can understand user intents and extract key information from utterances, such as booking a flight to a specific city on a specific date. They need to implement this using Azure AI Language. Which feature should they use?

A.Conversational language understanding (CLU)

B.Custom text classification

C.Named entity recognition (NER)

D.Key phrase extraction

AnswerA

CLU is designed to understand user intents and extract entities from conversational text. It allows you to define intents like 'BookFlight' and entities like destination and date, then train a model to predict them from utterances. This directly matches the requirement of understanding intents and extracting key information from user input.

Why this answer

Conversational language understanding (CLU) is the Azure AI Language feature built for intent recognition and entity extraction in conversational input. It supports custom intents and entities, making it ideal for building chatbots and voice assistants that need to parse user requests like flight bookings. Other features lack intent understanding or custom entity extraction.

Exam trap

The trap here is confusing entity extraction features like NER with intent understanding, which only CLU provides in Azure AI Language.

Full explanation →

109

MCQhard

You are a developer at a global e-commerce company. You are building a multilingual chatbot using Azure AI Language that supports English, French, German, and Spanish. The chatbot must answer frequently asked questions about order status, returns, and shipping. You plan to use Custom Question Answering with a single project containing questions and answers in all four languages. However, during testing, you notice that queries in French and German often return incorrect answers or no answer, while English and Spanish work well. You need to ensure accurate answers across all four languages. What should you do?

A.Create a separate Custom Question Answering project for each language and route user queries to the appropriate project based on language detection.

B.Use Azure Cognitive Search with semantic ranking to index the QnA pairs.

C.Add synonyms in the project for French and German terms to improve matching.

D.Deploy the same project to multiple regions and use traffic manager.

AnswerA

Custom Question Answering projects are language-bound: a single project cannot reliably match questions across multiple languages, so French and German queries miss. Separate per-language projects, selected via language detection, give each language its own trained question-answer pairs and accurate matching.

Why this answer

Custom Question Answering (CQA) projects are language-specific; a single project cannot reliably handle multiple languages due to differences in tokenization, stemming, and stop-word handling. By creating a separate project per language and routing queries based on language detection (e.g., using Azure AI Language's language detection API), you ensure that each project's model is optimized for its respective language, improving answer accuracy for French and German.

Exam trap

The trap here is that candidates assume a single Custom Question Answering project can handle multiple languages by simply adding translated QnA pairs, overlooking that the underlying NLP pipeline is language-specific and cannot correctly process queries in languages other than the project's configured primary language.

How to eliminate wrong answers

Option B is wrong because Azure Cognitive Search with semantic ranking is a search enhancement for indexed documents, not a solution for multilingual QnA matching; it does not address the language-specific tokenization and model training limitations of a single CQA project. Option C is wrong because adding synonyms only improves lexical matching for individual terms but does not resolve the fundamental issue that CQA's underlying model is trained on a single language's linguistic patterns; it cannot correctly interpret grammar, syntax, or phrasing differences across multiple languages. Option D is wrong because deploying the same project to multiple regions and using Traffic Manager only improves latency and availability, not the accuracy of answers for different languages; the underlying model remains unchanged and still fails for French and German.

Full explanation →

110

MCQmedium

You are designing a solution that reads handwritten notes from patient intake forms. The solution must handle various handwriting styles. Which Azure AI capability should you use?

A.Azure AI Document Intelligence Read model

B.Azure AI Custom Vision

C.Azure AI Vision OCR

D.Azure AI Language

AnswerA

The Read model handles handwriting and printed text.

Why this answer

Azure AI Document Intelligence Read model is specifically designed to extract printed and handwritten text from documents, including patient intake forms. It uses advanced OCR capabilities optimized for varied handwriting styles and document layouts, making it the correct choice for this scenario.

Exam trap

The trap here is that candidates often confuse Azure AI Vision OCR (which is for printed text) with the Document Intelligence Read model (which is specialized for handwriting and document structure), leading them to choose the wrong service for handwriting recognition tasks.

How to eliminate wrong answers

Option B is wrong because Azure AI Custom Vision is used for image classification and object detection, not for extracting text from documents or handwriting. Option C is wrong because Azure AI Vision OCR is a general-purpose OCR that works well for printed text but is not optimized for handwriting recognition. Option D is wrong because Azure AI Language is focused on natural language processing tasks like sentiment analysis and entity recognition, not on extracting text from images or documents.

Full explanation →

111

Multi-Selecthard

You are designing a generative AI solution using Azure OpenAI Service. The solution must support multiple languages and provide consistent quality across languages. Which THREE actions should you take?

Select 3 answers

A.Fine-tune the model on a dataset of a single language

B.Use a model that supports multiple languages (e.g., GPT-4)

C.Provide examples in multiple languages in the prompt

D.Set the temperature to 0 for all requests

E.Test the solution with representative prompts in each language

AnswersB, C, E

Multilingual models handle multiple languages natively.

Why this answer

GPT-4 is a multilingual model pre-trained on diverse language corpora, enabling it to generate coherent and contextually appropriate responses across many languages without additional fine-tuning. This ensures consistent quality by leveraging the model's inherent cross-lingual capabilities, which is essential for a generative AI solution that must support multiple languages.

Exam trap

The trap here is that candidates may think fine-tuning on a single language (Option A) is sufficient for multilingual support, or that setting temperature to 0 (Option D) universally improves consistency, when in fact these actions undermine the required cross-lingual quality and flexibility.

Full explanation →

112

MCQeasy

You need to analyze customer call transcripts to identify positive and negative sentiment. Which Azure AI Language feature should you use?

A.Language Detection

B.Named Entity Recognition

C.Key Phrase Extraction

D.Sentiment Analysis

AnswerD

Sentiment Analysis directly returns per-document and per-sentence sentiment labels with confidence scores, satisfying the requirement to identify positive and negative opinions in call transcripts. It is purpose-built for opinion mining within Azure AI Language, unlike key phrase extraction or entity recognition, which surface terms rather than polarity.

Why this answer

Sentiment Analysis is the correct Azure AI Language feature because it is specifically designed to evaluate text and determine whether the sentiment expressed is positive, negative, or neutral. For customer call transcripts, this feature analyzes each sentence or document and returns a sentiment label and confidence scores, directly addressing the requirement to identify positive and negative sentiment.

Exam trap

The trap here is that candidates often confuse Key Phrase Extraction with Sentiment Analysis, assuming that identifying key topics inherently reveals sentiment, but Key Phrase Extraction provides no sentiment polarity or confidence scores.

How to eliminate wrong answers

Option A is wrong because Language Detection identifies the language of the text (e.g., English, Spanish) and does not evaluate sentiment or emotion. Option B is wrong because Named Entity Recognition extracts entities like people, organizations, and locations from text, but does not assess sentiment polarity. Option C is wrong because Key Phrase Extraction identifies important phrases and topics in the text, but it does not classify sentiment as positive or negative.

Full explanation →

113

MCQeasy

A company uses Azure AI Search to index customer support tickets. They need to automatically extract key phrases from each ticket to improve search relevance. Which built-in skill should they add to the skillset?

A.Key Phrase Extraction

B.Entity Recognition

C.Sentiment Analysis

D.OCR

AnswerA

Key Phrase Extraction is a built-in cognitive skill that runs natural language processing over each ticket's text, returning salient terms directly into the index. It satisfies the requirement to extract key phrases automatically without custom code or model training, improving relevance through enriched searchable fields.

Why this answer

The Key Phrase Extraction skill in Azure AI Search is a built-in cognitive skill that uses natural language processing to extract key phrases from text. It is designed exactly for scenarios like extracting important terms from support tickets to improve search relevance and indexing.

Exam trap

AI-102 often tests the difference between similar cognitive skills; the trap is confusing Key Phrase Extraction with Entity Recognition or Sentiment Analysis when the requirement is specifically to extract key phrases.

How to eliminate wrong answers

Option B is wrong because Entity Recognition identifies entities such as people, places, organizations, and dates, not key phrases. Option C is wrong because Sentiment Analysis determines positive, negative, or neutral sentiment, which does not extract key phrases. Option D is wrong because OCR (Optical Character Recognition) extracts text from images, which is irrelevant to extracting key phrases from text tickets.

Full explanation →

114

MCQmedium

You are building an Azure AI Search knowledge mining pipeline that enriches PDF documents with key phrases. The enrichment must be applied after text extraction and before the data is written to the index. You need to ensure the enriched key phrases are available for downstream skills and are mapped to an index field. Which component of the skillset defines the output of the Key Phrase Extraction skill and its mapping to the index?

A.The input parameter of the indexer

B.The context property of the skill

C.The outputFieldMappings of the indexer

D.The outputs array of the skill definition

AnswerD

In an Azure AI Search skillset, each skill has an outputs array that names the output produced by the skill and specifies the target enriched document node. For the Key Phrase Extraction skill, the output is typically named keyPhrases and is mapped to a node like /document/keyPhrases. This definition makes the output available to subsequent skills and to the indexer's outputFieldMappings, which then maps it to an index field. This is the correct place to define the skill's output.

Why this answer

The outputs array within a skill definition is where you declare the name of the output and the enriched document node it targets. This makes the enriched data available to later skills and to the indexer's output field mappings. The indexer's outputFieldMappings then connects that enriched node to a specific index field.

The context property scopes the skill's input, but does not define its output.

Exam trap

The trap here is confusing the skill's output definition with the indexer's output field mappings, which serve different roles in the enrichment pipeline.

Full explanation →

115

MCQeasy

You need to generate a summary of a long article using Azure OpenAI. The article is 10,000 tokens long. What should you do to fit the article within the model's context window?

A.Split the article into smaller sections and summarize each section separately.

B.Increase the temperature parameter.

C.Use a model with a smaller context window.

D.Set max_tokens to a lower value.

AnswerA

Splitting the article into smaller sections keeps each request within the model's context window, since a 10,000-token article exceeds it. Summarising each chunk separately, then optionally combining the partial summaries, avoids truncation and preserves coverage of the full content.

Why this answer

The article exceeds the model's context window (typically 4096 or 8192 tokens for GPT-3.5/4). Splitting the article into smaller sections and summarizing each separately allows you to process the entire content within the token limits, then combine the summaries for a final coherent output. This is a standard chunking strategy for long documents when using Azure OpenAI.

Exam trap

The trap here is that candidates confuse parameters that control output behavior (temperature, max_tokens) with the fundamental input token limit, leading them to incorrectly believe adjusting these parameters can bypass the context window restriction.

How to eliminate wrong answers

Option B is wrong because increasing the temperature parameter affects randomness and creativity of the output, not the input token limit; it does not help fit a long article into the context window. Option C is wrong because using a model with a smaller context window would make the problem worse, as it reduces the maximum input length, not increase it. Option D is wrong because setting max_tokens to a lower value only truncates the output length, not the input; the article still exceeds the context window and will be rejected or truncated at the input stage.

Full explanation →

116

MCQhard

Your company is building a knowledge base for customer support using Azure AI Search. You have a large dataset of customer emails stored in Azure Blob Storage. The solution must extract key phrases, detect sentiment, and identify customer intents (e.g., complaint, inquiry, feedback). You plan to use built-in AI skills for key phrase extraction and sentiment detection. For intent identification, you need a custom solution because the intents are specific to your business. You have trained a custom Language Understanding (LUIS) model and published it. How should you integrate the LUIS model into the Azure AI Search enrichment pipeline to extract intents?

A.Add a Document Intelligence skill to classify intents.

B.Configure the index to use a custom analyzer to parse intents.

C.Use the built-in Entity Recognition skill to extract intents.

D.Create a custom skill in the skillset that calls the LUIS endpoint and returns the top intent.

AnswerD

A custom skill in the skillset invokes the published LUIS endpoint, passing enriched text and mapping the returned top intent into the index. This satisfies the requirement for business-specific intent extraction that built-in skills cannot provide.

Why this answer

Azure AI Search enrichment pipelines support built-in skills (key phrase extraction, sentiment, entity recognition, OCR, etc.) and custom skills. A custom skill is a Web API endpoint that the skillset calls during enrichment, receiving a JSON payload and returning enriched fields. To integrate a published LUIS model, you create a custom skill (typically an Azure Function) that calls the LUIS prediction endpoint with the document text and returns the top intent and score, which the indexer then maps into the search index.

Exam trap

AI-102 often tests whether candidates know that built-in skills cover only generic NLP tasks — anything business-specific (custom intents, custom classification) requires a custom skill that wraps an external model endpoint.

How to eliminate wrong answers

Option A is wrong because Document Intelligence (formerly Form Recognizer) extracts structured data from documents (forms, invoices, receipts) — it does not perform intent classification. Option B is wrong because custom analyzers in Azure AI Search control tokenization, stemming, and stop words at query/index time; they have nothing to do with calling an ML model during enrichment. Option C is wrong because the built-in Entity Recognition skill extracts named entities (people, places, organizations) via Cognitive Services — it does not classify intents, which is a distinct NLU task.

Full explanation →

117

MCQhard

You deploy a GPT-4o model in Azure OpenAI Service for an internal assistant that summarizes long contracts. Legal requires that every generated summary be traceable to the exact source passages and that the assistant never answer from general world knowledge. You need the model to ground each statement in retrieved text and expose the supporting passages to the caller. What should you configure?

A.Enable the abuse monitoring logging option on the Azure OpenAI resource

B.Increase the model's max_tokens value to fit the entire contract

C.Set temperature to 0 and top_p to 1 on the chat completions call

D.Use the Azure OpenAI On Your Own Data feature with citations enabled

AnswerD

On Your Own Data connects the deployment to an Azure AI Search index and instructs the model to answer only from retrieved chunks, returning citations that map statements to source passages. This directly provides the traceability legal requires and constrains responses to the indexed contracts rather than general knowledge. It is the supported grounding pattern for this scenario.

Why this answer

Grounding the assistant in an indexed corpus and returning citations is what makes each summary statement traceable to a specific contract passage. Azure OpenAI On Your Own Data performs retrieval against Azure AI Search, injects the retrieved chunks, and returns citation metadata with the response. Sampling parameters, output length, and monitoring logs change behavior or observability but never bind statements to sources.

Exam trap

The trap here is believing that lowering temperature removes hallucination and therefore substitutes for retrieval-based grounding.

Full explanation →

118

MCQmedium

You are designing an Azure AI solution that uses an Azure OpenAI resource. The solution must allow developers to call the model from a web app without embedding API keys in client-side code. You need to ensure that the web app can authenticate to the Azure OpenAI resource. What should you implement?

A.Configure the web app to use the Azure OpenAI API key and rotate it every 24 hours using Azure Automation.

B.Store the Azure OpenAI key in Azure Key Vault and have the web app retrieve it at runtime.

C.Generate a shared access signature (SAS) token for the Azure OpenAI resource and pass it in the request header.

D.Enable managed identity on the web app and assign the Cognitive Services OpenAI User role to the identity on the Azure OpenAI resource.

AnswerD

Managed identity allows the web app to authenticate to Azure OpenAI without storing credentials. Assigning the Cognitive Services OpenAI User role grants the necessary permissions to call the model. This is the recommended approach for keyless authentication and integrates with Azure RBAC, eliminating secret management in code or configuration.

Why this answer

Using a managed identity with the appropriate Azure RBAC role allows the web app to authenticate to Azure OpenAI without any secrets. The Cognitive Services OpenAI User role provides the necessary permissions to invoke the model. This approach is secure, scalable, and aligns with Azure best practices for keyless authentication.

Exam trap

The trap here is assuming that storing the API key in Key Vault is sufficient for client-side security, when in fact the key can still be exposed if the client retrieves it.

Full explanation →

119

MCQmedium

You are building a generative AI application with Azure OpenAI Service. The application must generate responses that are grounded in a specific set of internal documents stored in an Azure AI Search index. You want to use the built-in 'On Your Data' feature. Which configuration parameter should you set to ensure the model retrieves relevant documents before generating a response?

A.Enable the 'temperature' parameter to a low value to make responses more factual.

B.Set the model deployment to a fine-tuned model trained on the documents.

C.Use the 'max_tokens' parameter to limit the response length to match document snippets.

D.Set the data source type to Azure Cognitive Search and provide the index name.

AnswerD

Configuring the data source to Azure Cognitive Search with the correct index name enables the On Your Data feature to query the index for relevant documents before generating a response. This grounds the model's output in your proprietary data, reducing hallucinations and ensuring answers are based on the specified documents.

Why this answer

The On Your Data feature in Azure OpenAI Service allows the model to retrieve relevant documents from a specified data source, such as Azure Cognitive Search, before generating a response. By setting the data source type to Azure Cognitive Search and providing the index name, the model can ground its answers in the internal documents. This is the intended method for retrieval-augmented generation with Azure OpenAI.

Exam trap

The trap here is confusing fine-tuning with retrieval-augmented generation, assuming that training a model on documents is equivalent to dynamically retrieving them at inference time.

Full explanation →

120

MCQmedium

You are a solution architect at a news agency. The agency publishes thousands of articles daily. You need to build a knowledge mining solution that enables journalists to search for articles by topic, sentiment, key people, and locations mentioned. The articles are stored as HTML files in Azure Blob Storage. The solution must also provide a summary for each article. You plan to use Azure AI Search with cognitive skills and Azure OpenAI. Which combination of skills and features should you include to meet all requirements with the best performance and accuracy?

A.Use Azure AI Document Intelligence to extract content from HTML, then use Azure AI Language to extract entities and sentiment. Index in Azure AI Search with semantic search.

B.Skillset with Entity Recognition skill, Sentiment skill, Key Phrase Extraction skill, and Text Translation skill. Enable semantic search.

C.Skillset with Entity Recognition skill, Sentiment skill, and Key Phrase Extraction skill. Use Azure OpenAI service to generate summaries via a custom skill that calls the GPT model. Enable semantic search.

D.Skillset with Entity Recognition skill, Sentiment skill, and Text Analytics for Health skill to extract medical terms. Use Azure OpenAI for summarization as a custom skill.

AnswerC

Entity Recognition extracts people and locations, Sentiment scores tone, and Key Phrase Extraction surfaces topics; a custom skill invoking Azure OpenAI generates article summaries. Semantic search then reranks results for relevance, meeting every stated requirement across HTML blob content.

Why this answer

The requirements are topic search, sentiment, key people, locations, and article summaries. Entity Recognition extracts people and locations, Sentiment provides sentiment, Key Phrase Extraction supports topic search, and a custom skill calling Azure OpenAI GPT generates summaries — all wired into an Azure AI Search skillset with semantic search for relevance. This combination directly maps to every requirement without extraneous skills.

Exam trap

AI-102 often tests whether candidates can map each stated requirement to a specific skill — the trap is picking an option with a plausible-sounding but irrelevant skill (Text Translation, Text Analytics for Health, Document Intelligence) while missing a required capability like summarization.

How to eliminate wrong answers

Option A is wrong because Azure AI Document Intelligence is designed for structured document extraction (forms, invoices, PDFs) and is overkill/inefficient for HTML articles; Azure AI Language can extract entities and sentiment but the option omits key phrase extraction and summarization, and Document Intelligence adds cost/latency without benefit for HTML. Option B is wrong because Text Translation is irrelevant — the articles are not stated to be in multiple languages — and the option omits summarization entirely, failing a stated requirement. Option D is wrong because Text Analytics for Health extracts medical terms, which is irrelevant to a news agency, and while it includes summarization, the health skill is a mismatch that wastes resources and could misclassify general news content.

Full explanation →

121

Multi-Selecteasy

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

Select 2 answers

A.Azure AI Face

B.Azure AI Document Intelligence

C.Azure AI Video Indexer

D.Azure AI Computer Vision

E.Azure AI Custom Vision

AnswersB, D

Azure AI Document Intelligence performs OCR plus layout analysis, extracting text from images embedded in documents and returning structured content. This satisfies the requirement to extract text from images, particularly where layout and form structure matter.

Why this answer

Azure AI Document Intelligence (formerly Form Recognizer) includes the Read OCR engine that extracts printed and handwritten text from images and documents. Azure AI Computer Vision provides the OCR API (optical character recognition) which can extract text from images, including both printed and handwritten text, and supports multiple languages. Both services are designed specifically for text extraction from visual content.

Exam trap

The trap here is that candidates may confuse Azure AI Video Indexer's ability to extract text from video frames as a primary image text extraction service, but it is designed for video analysis and indexing, not standalone image text extraction.

Full explanation →

122

MCQhard

A legal firm uses Azure AI Language's custom NER to extract party names, dates, and clauses from contracts. The model performs well on English contracts but poorly on French contracts. The firm wants to improve performance without retraining from scratch. What is the most efficient approach?

A.Create a separate custom NER project for French and train from scratch using French contracts.

B.Retrain the English model with a mix of English and French contracts.

C.Use Azure AI Translator to translate French contracts to English, then use the English model.

D.Use the multilingual option in Azure AI Language custom NER to extend the existing project to include French.

AnswerD

Enabling the multilingual option extends the existing project's training to cover French alongside English, reusing the current labelled data and model rather than building and training a separate project from scratch, which is the most efficient route.

Why this answer

Azure AI Language's custom NER supports a multilingual option that allows you to extend an existing project to include additional languages without retraining from scratch. By enabling this option and adding French labeled data, the model learns to recognize entities in French while retaining its English performance, making it the most efficient approach.

Exam trap

A common pitfall in the AI-102 exam is assuming you must train separate models for each language or rely on translation, when Azure AI Language's built-in multilingual support is the correct and efficient path.

How to eliminate wrong answers

Option A is wrong because creating a separate project and training from scratch is inefficient and ignores the multilingual capability that avoids redundant effort. Option B is wrong because retraining with a mix of English and French contracts without enabling the multilingual option would not properly handle language-specific features and could degrade performance. Option C is wrong because translating French contracts to English introduces translation errors and latency, and the model would still fail on native French text in production.

Full explanation →

123

MCQmedium

You are building an internal assistant on Azure OpenAI that must stream responses to a React web app while hiding the model endpoint key from the browser. The app already authenticates users with Microsoft Entra ID. You need the least-privilege approach that keeps the key out of client code. What should you do?

A.Create an Azure API Management instance that fronts Azure OpenAI, and have the React app call API Management with the user's Entra ID token.

B.Give the React app a system-assigned managed identity and call Azure OpenAI directly from the browser.

C.Store the Azure OpenAI key in Azure Key Vault and have the React app retrieve it at runtime using the signed-in user's token.

D.Embed the Azure OpenAI key in the React build and call the completions endpoint directly from the browser.

AnswerA

API Management can validate the Entra ID token, apply rate limits and policies, and hold the Azure OpenAI key in a named value or Key Vault reference so the browser never sees it. Streaming still works because API Management supports server-sent events passthrough. This uses the existing identity provider and enforces least privilege at the gateway.

Why this answer

The key must never reach the browser, so a server-side intermediary is required. Azure API Management can authenticate the Entra ID token, enforce policies, and inject the Azure OpenAI key from a secure store while preserving streaming responses. Managed identities cannot run in a browser, and retrieving a key client-side exposes it regardless of where it was stored.

Exam trap

The trap here is treating managed identity or Key Vault as something a browser-based single-page app can use directly, when both require server-side Azure compute to function.

Full explanation →

124

Multi-Selectmedium

Which THREE components are part of the Azure AI Bot Service?

Select 3 answers

A.Azure AI Search

B.Azure Bot Service (hosting)

C.Azure AI Language (CLU)

D.Bot Framework SDK

E.Bot Framework Composer

AnswersB, D, E

Azure Bot Service provides the hosting runtime that executes bot logic and exposes messaging endpoints, satisfying the requirement for a deployable bot component. It registers channels and manages connectivity between your bot and services such as Teams or web chat, forming the core platform element alongside the Bot Framework SDK and channel registration.

Why this answer

Azure Bot Service (hosting) (B) is correct because the Azure AI Bot Service provides the hosting/registration resource that exposes a messaging endpoint and connects bots to channels such as Microsoft Teams, Web Chat, and Direct Line. Bot Framework SDK (D) is correct because it is the core development library (for C#, JavaScript/TypeScript, Python, and Java) used to build bots that handle activities, turns, dialogs, and channel connectors. Bot Framework Composer (E) is correct because it is the visual authoring tool included in the Bot Framework tooling for designing, testing, and publishing conversational bots.

Azure AI Search (A) is not part of the Bot Service; it is a separate cognitive search service for indexing and querying content. Azure AI Language (CLU) (C) is also not part of the Bot Service; it is a separate Azure AI Language capability for conversational language understanding that a bot may call but that is not a Bot Service component.

Exam trap

The trap here is that candidates often confuse external AI services like Azure AI Search or Azure AI Language (CLU) as being part of the Azure AI Bot Service, when they are actually separate services that can be integrated but are not core components of the bot service itself.

Full explanation →

125

MCQhard

You are deploying a computer vision model using Azure AI Custom Vision with a small dataset of 200 images per class. The model shows high accuracy on training data but low accuracy on test data. Which action should you take to reduce overfitting?

A.Increase the learning rate

B.Reduce the image size to lower resolution

C.Increase the number of training epochs

D.Increase the training dataset size with more varied images

AnswerD

Overfitting arises when the model memorises limited training examples, so adding more varied images increases feature diversity and improves generalisation to unseen test data. This directly addresses the small dataset constraint of 200 images per class named in the stem.

Why this answer

Increase the training dataset size with more varied images. Overfitting occurs when the model learns noise from a small dataset. Adding more varied images helps the model generalize.

Option A (increase learning rate) is wrong because it may cause divergence or unstable training, not directly reduce overfitting. Option B (reduce image size) is wrong because it can lose important features and may even increase overfitting. Option C (increase training epochs) is wrong because it can actually increase overfitting by allowing the model to memorize more noise.

Full explanation →

126

MCQmedium

You are designing an Azure AI Search enrichment pipeline that processes scanned PDF invoices stored in Azure Blob Storage. You need to extract both printed text and handwritten notes from the documents before sending the content to an Azure AI Language entity recognition skill. The solution must minimize development effort and cost. Which skill should you add to the skillset?

A.Microsoft.Skills.Vision.OcrSkill

B.Microsoft.Skills.Text.MergeSkill

C.Microsoft.Skills.Custom.WebApiSkill

D.Microsoft.Skills.Text.KeyPhraseExtractionSkill

AnswerA

The built-in OcrSkill extracts printed and handwritten text from image files and PDFs, producing a text output that downstream skills can consume. It is a first-party Azure AI Search skill, so no custom code is required. Using it minimizes development effort while supporting the handwriting requirement, and it is billed through the Azure AI services attached to the search service.

Why this answer

The built-in OCR skill reads printed and handwritten text from images and PDFs and writes the extracted text into the enrichment tree. Because it is a first-party Azure AI Search skill, it requires no custom code, which satisfies the goals of low effort and predictable cost. Other text skills assume machine-readable input and cannot perform image recognition.

Exam trap

The trap here is assuming that any text-analysis skill can process scanned documents directly, when image content must first be converted to text by an OCR skill.

Full explanation →

127

MCQeasy

Your application needs to determine whether two photos of the same person are of the same individual, even if they are from different angles. Which Azure AI service should you use?

A.Azure AI Video Indexer

B.Azure AI Custom Vision

C.Azure AI Vision OCR

D.Azure AI Face

AnswerD

Azure AI Face provides face verification and identification, comparing facial features across images to determine whether two photos show the same individual. Its detection and matching models tolerate pose and angle variation, satisfying the stem's requirement to compare photos taken from different angles.

Why this answer

Azure AI Face provides face verification APIs that compare two faces and return a confidence score indicating whether they belong to the same person. It uses deep learning models trained to handle variations in pose, lighting, and expression, making it ideal for matching photos of the same individual from different angles.

Exam trap

The trap here is that candidates may confuse the generic 'detect faces in video' capability of Video Indexer with the dedicated face verification API of the Face service, or assume Custom Vision can be trained for face matching without realizing it lacks built-in pose-invariant comparison.

How to eliminate wrong answers

Option A is wrong because Azure AI Video Indexer is designed for extracting insights from video content (e.g., speech, faces, objects) and does not provide a direct face comparison API for still images. Option B is wrong because Azure AI Custom Vision requires training a custom model with labeled images for specific classification or object detection tasks, not for out-of-the-box face verification across pose variations. Option C is wrong because Azure AI Vision OCR (Optical Character Recognition) extracts text from images and has no capability to analyze or compare facial features.

Full explanation →

128

MCQmedium

You are using Azure AI Document Intelligence to process a large batch of PDF forms. The forms have varying layouts and handwriting. You need to extract text and key-value pairs. Which custom model type should you train?

A.Custom template model

B.Prebuilt-layout model

C.Custom neural model

D.Custom composed model

AnswerC

Custom neural models handle unstructured documents with varying layouts and mixed handwriting, generalising from labelled samples across diverse form structures. Custom template models require consistent visual layout, so they fail on the varying layouts described; the neural model extracts text and key-value pairs reliably.

Why this answer

Custom neural model. Azure AI Document Intelligence offers custom template models for forms with fixed layouts and custom neural models for forms with varying layouts and handwriting. Neural models use deep learning to handle variability in structure and handwriting, making them ideal for this scenario.

Option A (Custom template model) assumes a fixed layout and fails with varying layouts. Option B (Prebuilt-layout model) is a prebuilt model that extracts text, tables, and selection marks but not customized key-value pairs for your forms. Option D (Custom composed model) is a combination of multiple models, but the primary choice for varying layouts is the neural model, not composed.

Therefore, C is the best choice.

Full explanation →

129

MCQmedium

You are deploying a generative AI assistant on Azure OpenAI Service that must summarize user-supplied financial reports. Security policy requires that report contents never leave your Azure tenant and that the model must not be fine-tuned. You need to provision the resource so that prompt and completion data are not retained for human review and are not used to train any shared model. What should you configure?

A.Set the Azure OpenAI resource's 'Content logging' to Disabled in Azure Monitor diagnostic settings.

B.Submit an Azure OpenAI limited access modification request to disable abuse monitoring for the subscription.

C.Create the deployment in a separate resource group and assign the Cognitive Services User role only to the application's managed identity.

D.Deploy the model with a content filter policy set to block high-severity categories and enable prompt shields.

AnswerB

Approved limited access modification for abuse monitoring removes the standard human review and retention of prompts and completions for the approved subscription. Because the requirement is that report contents are neither retained for review nor used to train shared models, this is the correct control. Fine-tuning is unaffected, and data stays within the tenant boundary.

Why this answer

The requirement is about data handling by the service itself, not about access control or content safety. The only mechanism that removes default human review and retention of prompts and completions for an Azure OpenAI resource is an approved limited access modification for abuse monitoring on the subscription. Content filters, diagnostic settings, and RBAC all address different concerns and leave the retention behavior unchanged.

Exam trap

The trap here is assuming that disabling diagnostic content logging or adding content filters changes how the service retains prompts, when data-handling behavior is governed by the abuse-monitoring modification process instead.

Full explanation →

130

MCQhard

You are using Azure AI Custom Vision to detect defects in fabric rolls. After training an object detection model, you notice that it often misses small tears. You have a large dataset of labeled images, but the tears are very small relative to the image size. What should you do to improve the model's detection of small tears?

A.Train a new model with images where the tears are larger in the frame

B.Increase the number of training iterations

C.Use the 'General' domain instead of 'Product' domain

D.Increase the model's confidence threshold

AnswerA

To improve detection of small objects, you can crop or zoom in on the regions containing the tears so they occupy a larger portion of the image. This gives the model more pixels to learn from and makes the features more salient. Retraining with such images helps the model detect small tears more reliably, as it focuses on the relevant details.

Why this answer

Small object detection is challenging because the objects occupy few pixels. By training with images where the tears are larger in the frame—through cropping or zooming—you provide the model with more detailed features to learn from. This approach directly addresses the issue and improves the model's ability to detect small tears.

Exam trap

The trap here is thinking that increasing iterations or changing domains will solve small object detection, but the key is to make the objects larger in the training images.

Full explanation →

131

MCQeasy

You are building a chatbot using Azure AI Bot Service and Language Service. The bot must recognize user intent for 'check order status'. How should you configure the Language Service?

A.Create a custom intent classification project

B.Deploy a QnA Maker knowledge base

C.Configure a sentiment analysis endpoint

D.Use the prebuilt entity extraction model

AnswerA

A custom intent classification project in Azure AI Language trains a model on your labelled utterances, returning the top intent for 'check order status'. This satisfies the stem's requirement to recognise user intent, since conversational language understanding maps utterances to intents rather than extracting entities or answering questions.

Why this answer

To recognize user intent for 'check order status', you need a custom intent classification project in Azure Language Service. This project type uses a trained model to map utterances to specific intents, such as 'CheckOrderStatus', which is exactly what the chatbot requires. Prebuilt models or QnA Maker do not provide custom intent recognition.

Exam trap

The trap here is that candidates confuse intent recognition with entity extraction or QnA, assuming any Language Service feature can handle intents, but only custom intent classification (or conversational language understanding) is designed for this purpose.

How to eliminate wrong answers

Option B is wrong because QnA Maker is designed for question-answering over a knowledge base, not for classifying user intent from natural language utterances. Option C is wrong because sentiment analysis determines the emotional tone of text, not the user's intent or goal. Option D is wrong because prebuilt entity extraction identifies named entities like dates or locations but does not classify the overall intent of a user message.

Full explanation →

132

MCQmedium

You are deploying an Azure AI Language service solution for a multilingual customer support chatbot. The solution must support real-time translation between English, Spanish, and French. Which Azure resource should you provision?

A.Azure AI Speech

B.Azure OpenAI Service

C.Azure AI Language

D.Azure AI Translator

AnswerD

Azure AI Translator provides real-time text translation across English, Spanish and French, satisfying the multilingual chatbot requirement. It is the dedicated translation resource, unlike Azure AI Language, which handles entity extraction and sentiment rather than cross-language rendering.

Why this answer

Azure AI Translator is the correct resource because it provides real-time text translation across multiple languages, including English, Spanish, and French, via a dedicated REST API. The scenario specifically requires translation between languages, not speech recognition or generative AI, making Azure AI Translator the precise service for this task.

Exam trap

The trap here is that candidates confuse Azure AI Language with Azure AI Translator, assuming the 'Language' service includes translation, when in fact translation is a separate service under the Azure AI Services umbrella.

How to eliminate wrong answers

Option A is wrong because Azure AI Speech handles speech-to-text, text-to-speech, and speech translation, but it is not optimized for pure text translation between multiple languages without audio input. Option B is wrong because Azure OpenAI Service is designed for generative AI tasks like content creation and conversation, not for direct, real-time text translation between specific languages. Option C is wrong because Azure AI Language provides natural language processing capabilities such as sentiment analysis and entity recognition, but it does not include a dedicated real-time translation API; translation is handled by Azure AI Translator.

Full explanation →

133

MCQeasy

You are using Azure OpenAI Service to generate marketing copy. You notice that the output sometimes contains factual inaccuracies about your company's products. Which action can you take to improve factual accuracy?

A.Lower the temperature to 0.

B.Include relevant product information in the system message.

C.Increase the maxTokens to 4000.

D.Add a stop sequence to limit output.

AnswerB

Embedding product facts in the system message grounds generation in authoritative context, directly addressing the factual-accuracy constraint. Unlike fine-tuning, which adjusts model weights, system-message grounding supplies reference material at inference time, so the model conditions its output on your company's actual product details rather than relying on potentially stale pretrained knowledge.

Why this answer

Including relevant product information in the system message provides the model with authoritative context that grounds its responses in factual data. The system message acts as a persistent instruction set that the model uses to shape its outputs, reducing reliance on its internal training data which may be outdated or incomplete. This technique, known as 'grounding,' directly improves factual accuracy by supplying the model with the specific facts it needs to generate correct marketing copy.

Exam trap

The trap here is that candidates often assume lowering temperature or increasing maxTokens will fix factual accuracy, when in reality these parameters control randomness and output length, not the correctness of the underlying information.

How to eliminate wrong answers

Option A is wrong because lowering the temperature to 0 reduces randomness and creativity but does not inject factual data; it only makes the model more deterministic in its token selection, which can still produce inaccuracies if the model lacks the correct information. Option C is wrong because increasing maxTokens to 4000 only extends the maximum length of the output, which does not address the root cause of factual errors and may even allow the model to generate more incorrect content. Option D is wrong because adding a stop sequence limits where the model stops generating text, which controls output length but does not improve the factual accuracy of the content produced.

Full explanation →

134

Multi-Selectmedium

Which TWO Azure AI services can be used to detect objects in images?

Select 2 answers

A.Video Indexer

B.Face API

C.Custom Vision

D.Azure AI Document Intelligence

E.Computer Vision Object Detection API

AnswersC, E

Custom Vision trains a bespoke object detection model on your own labelled images, returning bounding boxes per class. It satisfies the stem's requirement to detect objects, unlike classification-only or face-based services, because you control the labelled dataset and exported model.

Why this answer

Custom Vision (C) is correct because it is an Azure AI service that lets you train and deploy custom image classification and object detection models, returning bounding boxes for detected objects in images. Computer Vision Object Detection API (E) is correct because the Azure AI Vision (Computer Vision) service provides a prebuilt object detection capability that identifies common objects and their bounding box coordinates in an image. Video Indexer (A) is not the right fit here since it focuses on analyzing and indexing video and audio content rather than being an image object detection service.

Face API (B) only detects and analyzes human faces (and related attributes), not general objects. Azure AI Document Intelligence (D) extracts text, key-value pairs, and structured data from documents, so it does not perform general object detection in images.

Exam trap

The trap here is that candidates often confuse the general-purpose Computer Vision Object Detection API (which is pre-trained on common objects) with Custom Vision (which requires custom training), but both are valid for object detection depending on the scenario, and the question asks for two services that can detect objects, making both C and E correct.

Full explanation →

135

MCQeasy

Your company uses Azure AI Search to power a customer support portal. The search index includes product documentation and known issues. Recently, the portal's search performance has degraded, and users report slow response times. You need to identify the cause of the performance issue. What should you check first?

A.Review the search service metrics for high query latency and CPU usage.

B.Check the size of the index storage in the Azure portal.

C.Ensure the index schema does not have too many fields.

D.Verify that the skillset is not running during peak hours.

AnswerA

Reviewing service metrics exposes query latency and CPU saturation, the primary indicators of degraded Azure AI Search performance. High CPU usage or elevated query latency directly satisfies the stem's requirement to identify the cause first, since these metrics reveal whether throttling, expensive queries, or insufficient replicas are driving the slow response times.

Why this answer

High query latency and CPU usage are direct indicators of performance bottlenecks in Azure AI Search. The search service metrics in the Azure portal provide real-time data on query execution time and resource consumption, which are the first signals to investigate when users report slow response times. Checking these metrics helps identify whether the issue stems from excessive query load, insufficient replicas, or inefficient query execution.

Exam trap

The trap here is that candidates may confuse indexing-related metrics (like skillset execution or index size) with query performance metrics, leading them to check storage size or schema complexity instead of the direct performance indicators of query latency and CPU usage.

How to eliminate wrong answers

Option B is wrong because index storage size alone does not directly cause slow query response times; large indexes can be handled efficiently with proper partitioning and replicas, and storage metrics are more relevant to capacity planning than immediate performance degradation. Option C is wrong because having too many fields in the index schema can increase indexing time but does not typically cause slow query response times; query performance is more affected by the number of searchable fields and the complexity of queries, not the total field count. Option D is wrong because skillsets run during indexing, not querying, and their execution does not impact query response times; query performance is independent of indexing operations unless the service is under-provisioned for concurrent workloads.

Full explanation →

136

Multi-Selecthard

Which THREE factors should you consider when choosing between Azure OpenAI Service and Azure Machine Learning for deploying a generative AI model?

Select 3 answers

A.Integration with Microsoft Purview for data governance.

B.Latency requirements: Azure OpenAI may offer lower latency for standard models.

C.Ability to scale to thousands of concurrent requests.

D.Need for custom model architecture: Azure ML supports custom models, Azure OpenAI uses pre-trained.

E.Operational overhead: Azure OpenAI is a fully managed service.

AnswersB, D, E

Azure OpenAI endpoints are optimized for low latency, whereas Azure ML may require additional optimization.

Why this answer

Azure OpenAI Service provides managed endpoints for pre-trained models like GPT-4, which are optimized for low-latency inference out of the box. In contrast, Azure Machine Learning requires you to deploy your own containerized model, which can introduce additional network and compute overhead, making Azure OpenAI the better choice when sub-100ms response times are critical for standard generative AI tasks.

Exam trap

The trap here is that candidates assume 'scalability' is unique to one service, but both Azure OpenAI and Azure ML can handle high concurrency; the real differentiator is latency and customizability, not raw throughput.

Full explanation →

137

MCQeasy

A company uses Azure Custom Vision to classify images of defective parts. After deploying the model, the accuracy is low. The team only has 10 images per class. What is the most effective way to improve accuracy?

A.Use a different classification algorithm.

B.Add at least 50 more images per class with variations.

C.Reduce the image resolution to speed up training.

D.Increase the number of training iterations (epochs).

AnswerB

Ten images per class is far below what Custom Vision needs to generalise; adding at least 50 varied images per class gives the model enough examples to learn distinguishing features, directly addressing the small-dataset constraint that is causing the low accuracy.

Why this answer

Azure Custom Vision relies on deep learning models that require a sufficient number of diverse training images to generalize well. With only 10 images per class, the model is severely underfit and prone to overfitting; adding at least 50 more images per class with variations in lighting, angle, and background provides the necessary data diversity to improve accuracy significantly.

Exam trap

The trap here is that candidates often assume increasing epochs or changing the algorithm will fix low accuracy, but the real bottleneck is insufficient and non-diverse training data, which is the most common cause of poor Custom Vision model performance.

How to eliminate wrong answers

Option A is wrong because Azure Custom Vision automatically selects and tunes the underlying classification algorithm (a convolutional neural network) based on the dataset; manually changing the algorithm is not supported and would not address the core issue of insufficient training data. Option C is wrong because reducing image resolution can discard important fine-grained features needed to detect defects, and Azure Custom Vision already resizes images to a fixed input size (e.g., 224x224) during training, so further reduction harms accuracy rather than improving it. Option D is wrong because increasing training iterations (epochs) with only 10 images per class will cause the model to overfit to the small dataset, memorizing the training examples rather than learning generalizable patterns, leading to poor accuracy on new images.

Full explanation →

138

MCQhard

You are designing an NLP solution to analyze legal documents. The solution must identify specific clauses and parties involved. Which Azure AI service is most appropriate?

A.Custom Named Entity Recognition in Azure AI Language

B.Pre-built Named Entity Recognition in Azure AI Language

C.Text Analytics for Health

D.Immersive Reader

AnswerA

Custom Named Entity Recognition trains a model on your labelled legal data to extract domain-specific entities such as clauses and party names, which prebuilt models cannot recognise. It satisfies the requirement to identify bespoke fields rather than generic persons or organisations.

Why this answer

Custom Named Entity Extraction (Custom NER) in Azure AI Language is the correct choice because it allows you to train a model to recognize domain-specific entities like legal clauses and party names from your own labeled data. Pre-built NER only recognizes generic entity types (e.g., person, organization, location) and cannot be customized for legal terminology. This makes Custom NER the only option that meets the requirement to identify specific clauses and parties unique to legal documents.

Exam trap

The trap here is that candidates often confuse Pre-built NER with Custom NER, assuming the pre-built model can handle domain-specific entities like legal clauses, but it only recognizes generic categories and cannot be retrained.

How to eliminate wrong answers

Option B is wrong because Pre-built Named Entity Recognition only identifies a fixed set of common entity types (e.g., Person, Organization, Location) and cannot be trained to recognize custom legal clauses or specific party roles. Option C is wrong because Text Analytics for Health is designed specifically for medical and healthcare entities (e.g., diagnoses, medications, symptoms) and has no capability to parse legal document structures or clauses. Option D is wrong because Immersive Reader is a tool for improving reading comprehension (e.g., text-to-speech, translation, focus mode) and does not perform any entity extraction or NLP analysis.

Full explanation →

139

Multi-Selecteasy

Which TWO features of Azure AI Search allow you to improve the relevance of search results for users?

Select 2 answers

A.Synonym maps

B.Semantic search

C.Suggesters

D.Scoring profiles

E.Filterable fields

AnswersB, D

Semantic search applies Microsoft's language understanding models to rerank results, promoting passages that are semantically relevant rather than only keyword-matched. This directly improves relevance ranking for users, satisfying the stem's requirement to enhance search result relevance.

Why this answer

Semantic search (B) is correct because it uses Microsoft's language understanding models to rerank results with semantic captions and answers, boosting relevance beyond keyword matching. Scoring profiles (D) are correct because they let you define custom ranking functions (e.g., boosting by freshness, magnitude, or tags) that directly influence the relevance score of returned documents. Synonym maps (A) expand queries with equivalent terms but only broaden matching, not rank relevance.

Suggesters (C) enable type-ahead autocomplete, which improves query input rather than result relevance. Filterable fields (E) restrict the result set with OData filters but do not affect relevance ranking.

Exam trap

The trap here is that candidates often confuse features that expand recall (synonym maps) or improve user experience (suggesters) with features that directly improve relevance ranking, leading them to select A or C instead of the correct scoring profiles and semantic search.

Full explanation →

140

MCQhard

You are deploying a custom text classification model using Azure AI Language. The model must be retrained monthly with new labeled data. You need to automate the retraining process with minimal manual intervention. Which approach should you use?

A.Create an Azure DevOps pipeline that manually retrains the model every month

B.Retrain the model manually using Language Studio each month

C.Use the Azure AI Language REST API to trigger training and deployment on a schedule using Azure Logic Apps

D.Use Azure Machine Learning to host the custom model and automate retraining

AnswerC

Scheduled REST API calls from Logic Apps trigger training and deployment without human input, satisfying the monthly retraining requirement with minimal manual intervention. The API exposes training and deployment operations directly, so the workflow can orchestrate both steps automatically.

Why this answer

Azure Logic Apps can schedule HTTP requests to the Azure AI Language REST API to trigger training and deployment of a custom text classification model. This approach automates the monthly retraining process without manual intervention, using the API's capabilities for model creation, training, and deployment.

Exam trap

The trap here is that candidates may confuse Azure Machine Learning with Azure AI Language, thinking that Azure Machine Learning is the appropriate service for hosting and retraining custom text classification models, when in fact Azure AI Language provides its own REST API for this purpose and is the correct service for custom text classification.

How to eliminate wrong answers

Option A is wrong because an Azure DevOps pipeline that manually retrains the model every month still requires manual intervention to trigger the pipeline, contradicting the requirement for minimal manual intervention. Option B is wrong because retraining manually using Language Studio each month is entirely manual and does not automate the process. Option D is wrong because Azure Machine Learning is not designed to host custom text classification models built with Azure AI Language; it is a separate platform for building and deploying machine learning models, and using it would introduce unnecessary complexity and integration overhead.

Full explanation →

141

MCQhard

A logistics company uses an Azure AI Search indexer to process bills of lading stored in Azure Blob Storage. The indexer uses a skillset with a ShaperSkill that builds a complex object named 'shipment' containing nested fields for carrier, origin, and destination. After a full index run, queries for the carrier field return no results even though the source documents contain the data. You need to make the carrier value searchable. What should you do?

A.Change the ShaperSkill to output a string instead of an object.

B.Increase the indexer batch size so that all documents are processed in a single run.

C.Enable the indexer's cache and rerun the indexer.

D.Add an outputFieldMapping that maps the nested carrier value to a top-level index field.

AnswerD

Values produced inside a ShaperSkill object exist only in the enrichment tree. To persist them in the index, the indexer needs an outputFieldMapping that targets a field defined in the index. Mapping the nested carrier node to a searchable top-level field makes the value retrievable and queryable, resolving the empty query results.

Why this answer

The ShaperSkill constructs a complex object inside the enrichment tree, but that tree is transient. Persisting any part of it requires an outputFieldMapping from the enrichment node to a field in the index. Mapping the nested carrier node to a searchable top-level field is what actually makes the value appear in query results.

Exam trap

The trap here is believing that shaping an object automatically indexes its members, when shaped nodes must be explicitly mapped to index fields.

Full explanation →

142

MCQeasy

You are designing a solution that uses Azure AI Language to analyze customer feedback. The solution must detect sentiment, extract key phrases, and identify named entities. Which feature should you use?

A.Azure AI Language service

B.Azure AI Speech service

C.Azure AI Computer Vision

D.Translator API

AnswerA

Azure AI Language consolidates sentiment analysis, key phrase extraction and named entity recognition within one resource, so a single call satisfies all three requirements without stitching separate services together. Its prebuilt models return sentiment scores, key phrases and entity categories directly, matching the stem's combined detection, extraction and identification constraints.

Why this answer

The Azure AI Language service provides pre-built capabilities for sentiment analysis, key phrase extraction, and named entity recognition (NER) as part of its text analytics features. These three tasks are directly supported by the service's Analyze API, making it the correct choice for analyzing customer feedback text.

Exam trap

The trap here is that candidates may confuse Azure AI Language with other Azure AI services that have overlapping names (e.g., Translator API for language tasks) or assume Speech or Vision services can perform text analysis, but only the Language service provides the specific trio of sentiment, key phrases, and NER.

How to eliminate wrong answers

Option B is wrong because Azure AI Speech service is designed for speech-to-text, text-to-speech, and speech translation, not for analyzing text sentiment, key phrases, or named entities. Option C is wrong because Azure AI Computer Vision focuses on image and video analysis (e.g., object detection, OCR), not on natural language processing tasks like sentiment or entity extraction. Option D is wrong because Translator API is specifically for machine translation between languages and does not include sentiment analysis, key phrase extraction, or named entity recognition.

Full explanation →

143

MCQmedium

A company uses Azure AI Search to index documents from an Azure SQL Database. They have configured an indexer with a skillset that includes a custom skill hosted in an Azure Function. The custom skill enriches each document with a 'category' field. After running the indexer, they notice that the 'category' field is missing in the index for all documents. The Azure Function logs show that it is receiving requests and returning responses. What is the most likely cause?

A.The indexer's data source connection string is invalid, so documents are not being retrieved.

B.The output field mapping for the custom skill is not correctly configured in the indexer.

C.The custom skill's context is set to /document/pages/*, but the documents do not have a 'pages' node.

D.The custom skill's Azure Function is returning a 200 OK status but with an empty response body.

AnswerB

This is correct because if the custom skill's output is not mapped to an index field, the enriched data will not appear in the index. Even though the skill runs successfully, the indexer needs an outputFieldMapping to send the skill's output to the target field. Without it, the category data is discarded, resulting in missing values in the index.

Why this answer

The most likely cause is a missing or incorrect output field mapping. The custom skill executes and returns a response, but if the indexer is not configured to map that output to an index field, the enriched data never gets stored. This is a common configuration oversight: the skill definition includes an output, but the indexer's outputFieldMappings must explicitly associate it with a target field.

Exam trap

The trap here is focusing on the skill's execution or function code, but the missing field is often due to a missing output field mapping in the indexer.

Full explanation →

144

MCQmedium

You are developing an agent by using the Azure AI Foundry Agent Service. The agent must query a proprietary internal REST API that returns JSON data. The API requires an OAuth 2.0 access token for authentication. You need to configure the agent to call this API. What should you do?

A.Configure an OpenAPI tool for the agent, specifying the API's OpenAPI specification and setting up an OAuth 2.0 connection for authentication.

B.Use the Azure AI Foundry SDK to programmatically inject the OAuth token into each request by using a custom middleware component.

C.Add the API endpoint as a knowledge source in Azure AI Search and use integrated vectorization to index the JSON responses.

D.Create a custom tool by using a function calling definition that includes the API endpoint and authentication details, and register it with the agent.

AnswerA

OpenAPI tools in Azure AI Foundry Agent Service allow you to import an API specification and configure authentication, including OAuth 2.0. This provides a secure, managed way for the agent to call the API, handling token acquisition and refresh automatically. This is the recommended approach for integrating external REST APIs with OAuth.

Why this answer

The agent must call an OAuth-protected REST API. Azure AI Foundry Agent Service supports OpenAPI tools, which allow you to import an API specification and configure OAuth 2.0 authentication. This enables the agent to securely call the API, with the service handling token acquisition and refresh.

Other options either lack secure authentication support or are not designed for dynamic API invocation.

Exam trap

The trap here is assuming that function calling definitions can handle authentication, but they only describe the API schema and do not manage credentials securely.

Full explanation →

145

MCQhard

Refer to the exhibit. You have created an assistant with the above configuration. When you send a message 'What is the weather in Seattle?', the assistant responds without calling the function. What is the most likely cause?

A.The 'instructions' are not being followed

B.The 'strict' parameter is set to true

C.The function is missing the 'description' property

D.The 'tool_resources' code_interpreter has empty file_ids

AnswerC

Function definitions require a description so the model can judge when to invoke them. Without it, the model cannot match the weather query to the function and answers from its own knowledge instead of calling it.

Why this answer

The function definition lacks a 'description' property. In the Assistants API, the 'description' field is critical for the model to understand when and why to invoke a function. Without it, the model may not recognize that the function is relevant to the user's query about weather, causing it to respond directly instead of calling the function.

Exam trap

Azure AI often tests the misconception that the 'instructions' field or 'strict' parameter is the primary driver for function calling, when in reality the 'description' property is the key enabler for the model to understand tool relevance.

How to eliminate wrong answers

Option A is wrong because the 'instructions' field is used to set the assistant's behavior and system prompt, but it does not directly control function calling; the model can still ignore instructions if function definitions are incomplete. Option B is wrong because setting 'strict' to true (if supported) would enforce schema adherence, not prevent function calls; it would actually make the model more likely to follow the defined tools. Option D is wrong because 'tool_resources' with empty 'file_ids' for code_interpreter only affects file-based retrieval or code execution, not the decision to call a function; the function tool is defined separately in the 'tools' array.

Full explanation →

146

Multi-Selectmedium

Which TWO actions are required to enable a custom chatbot built with Azure OpenAI to answer questions based on a company's internal PDF documents?

Select 2 answers

A.Use Azure AI Document Intelligence to extract text from PDFs before indexing

B.Deploy Azure AI Content Safety to filter responses

C.Fine-tune the GPT model on the PDF content

D.Ingest the PDFs into an Azure Cognitive Search index

E.Configure the Azure OpenAI deployment to use 'Add your data' with the search index

AnswersD, E

Indexing enables retrieval of relevant content from PDFs.

Why this answer

Azure Cognitive Search provides the indexing and retrieval capabilities needed to make PDF content searchable. By ingesting PDFs into an Azure Cognitive Search index, the chatbot can perform vector or keyword searches over the extracted text, enabling it to retrieve relevant passages to answer user questions. This is the standard approach for grounding a custom chatbot on proprietary documents without modifying the underlying model.

Exam trap

The trap here is that candidates often confuse fine-tuning (option C) with the RAG pattern, mistakenly believing they must retrain the model on proprietary data, when in fact the 'Add your data' feature with a search index is the correct and simpler approach for question-answering over internal documents.

Full explanation →

147

Multi-Selectmedium

Which TWO are valid ways to manage cost when using Azure OpenAI Service in a production application?

Select 2 answers

A.Fine-tune the model to reduce the number of examples needed in prompts

B.Increase the temperature parameter to 1.0

C.Use a smaller model like GPT-3.5-turbo instead of GPT-4 for simpler tasks

D.Provision more PTUs to get a lower rate per token

E.Set the max_tokens parameter to the minimum needed for the response

AnswersC, E

Smaller models have lower per-token costs.

Why this answer

Using a smaller model like GPT-3.5-turbo for simpler tasks directly reduces the per-token cost compared to GPT-4, which is significantly more expensive. Azure OpenAI Service charges based on model tier and token usage, so selecting the appropriate model for the task complexity is a primary cost management strategy.

Exam trap

The trap here is that candidates may confuse fine-tuning with prompt optimization, or assume that increasing PTUs lowers per-token cost, when in fact PTUs are a fixed-cost commitment that increases total expenditure.

Full explanation →

148

MCQeasy

A company uses Azure Content Moderator to moderate text in a chat application. They want to automatically reject messages that contain profanity or personal data. Which API should they use?

A.Review API

B.Video Moderation API

C.Image Moderation API

D.Text Moderation API

AnswerD

The Text Moderation API screens text content and returns classification for profanity and personally identifiable information, letting the application reject offending messages automatically. This satisfies the requirement to detect both categories in chat text, unlike image moderation or the review and list management APIs.

Why this answer

The Text Moderation API (D) is the correct choice because it is specifically designed to scan text content for profanity, personally identifiable information (PII), and other unwanted text patterns. This API returns a moderation score and a list of detected terms, enabling automated rejection of messages that violate the defined policies.

Exam trap

The trap here is that candidates may confuse the Review API with the moderation APIs, not realizing that the Review API is for manual review workflows rather than automated content detection.

How to eliminate wrong answers

Option A is wrong because the Review API is used for human-in-the-loop review workflows, not for automated detection and rejection of profanity or personal data. Option B is wrong because the Video Moderation API is designed to moderate video content, not text messages. Option C is wrong because the Image Moderation API handles image content, not text-based chat messages.

Full explanation →

149

MCQeasy

Your organization wants to implement a document processing pipeline that extracts text from scanned PDFs and identifies named entities. Which two Azure AI services should you use?

A.Azure AI Language

B.Azure AI Custom Vision

C.Azure AI Document Intelligence

D.Azure AI Translator

E.Azure AI Speech

AnswerA, C

Azure AI Language provides the Named Entity Recognition feature, identifying people, places, organisations and other entities within text. It satisfies the stem's entity-identification requirement, consuming text that Document Intelligence has already extracted from the scanned PDFs.

Why this answer

Azure AI Document Intelligence (formerly Form Recognizer) is used to extract text from scanned PDFs via OCR, while Azure AI Language provides pre-built named entity recognition (NER) to identify entities like people, organizations, and locations. Together, they form a complete pipeline: Document Intelligence handles the image-to-text conversion, and Language processes the extracted text for entity extraction.

Exam trap

The trap here is that candidates often confuse Azure AI Document Intelligence with Azure AI Custom Vision, thinking Custom Vision can perform OCR, but Document Intelligence is the dedicated service for document text extraction and layout analysis.

How to eliminate wrong answers

Option B is wrong because Azure AI Custom Vision is designed for image classification and object detection, not for OCR or text extraction from scanned documents. Option D is wrong because Azure AI Translator is a machine translation service that converts text between languages, not for extracting text from images or identifying named entities. Option E is wrong because Azure AI Speech handles speech-to-text and text-to-speech, not OCR or NER from scanned PDFs.

Full explanation →

150

MCQhard

Refer to the exhibit. You have created a Text Analytics resource and retrieved its keys. You want to use the key1 to call the Sentiment Analysis API from a Python application. Which endpoint URL should you use?

A.https://mytextanalytics.cognitiveservices.azure.com/sentiment/v3.1

B.https://mytextanalytics.api.cognitive.microsoft.com/text/analytics/v3.1/sentiment

C.https://mytextanalytics.cognitiveservices.azure.com/analyze

D.https://mytextanalytics.cognitiveservices.azure.com/text/analytics/v3.1/sentiment

AnswerD

Cognitive Services multi-service and Text Analytics resources expose a regional or custom subdomain endpoint ending in cognitiveservices.azure.com, followed by the service path. The sentiment route under /text/analytics/v3.1/ matches the Sentiment Analysis API for this resource.

Why this answer

The Sentiment Analysis API for Azure Cognitive Services Text Analytics uses the endpoint pattern `https://<resource-name>.cognitiveservices.azure.com/text/analytics/v3.1/sentiment`. This is the standard REST API endpoint for sentiment analysis in version 3.1, which requires the `/text/analytics/v3.1/sentiment` path appended to the custom resource domain.

Exam trap

The trap here is that candidates often confuse the legacy domain (`api.cognitive.microsoft.com`) with the current Azure domain (`cognitiveservices.azure.com`), or they mistakenly use the Analyze API endpoint (`/analyze`) when a dedicated sentiment endpoint is required, leading them to pick options B or C.

How to eliminate wrong answers

Option A is wrong because it omits the required `/text/analytics/` path segment and uses an incorrect path `/sentiment/v3.1`; the version should be in the path after `analytics`, not after `sentiment`. Option B is wrong because it uses the legacy domain `api.cognitive.microsoft.com` instead of the current Azure global domain `cognitiveservices.azure.com`, which is required for all new Cognitive Services resources. Option C is wrong because `/analyze` is the endpoint for the Analyze API (which performs multiple tasks like key phrase extraction, entity recognition, and sentiment analysis in a single call), not the dedicated Sentiment Analysis API endpoint.

Full explanation →

Microsoft Azure AI Engineer Associate AI-102 (AI-102) — Questions 76–150