Courseiva

Microsoft Azure AI Engineer Associate AI-102 (AI-102) — Questions 1–75

761 questions total · 11pages · All types, answers revealed

Take a mock examExam hub

Page 1 of 11

Page 2
1
MCQhard

You have a Conversational Language Understanding (CLU) project in Azure AI Language. Users frequently type utterances such as 'Book a flight from Seattle to Tokyo next Friday for two people.' The solution must extract the origin city, destination city, date, and passenger count as separate structured values so the booking system can act on them. You need to configure the project to capture these values. What should you do?

A.Add more intents and assign each utterance to a single intent.
B.Enable sentiment analysis and opinion mining on the CLU project.
C.Add entities to the project and label them in the training utterances, using prebuilt entities where available.
D.Increase the training data by duplicating existing utterances with slight wording changes.
AnswerC

CLU extracts entities that are defined and labeled in the training utterances. Adding entities for origin, destination, date, and passenger count and labeling their spans teaches the model to return those values as structured fields. Using prebuilt components such as geography or number where available improves accuracy without building everything from scratch.

Why this answer

In Conversational Language Understanding, entity extraction depends on defining entities and labeling their occurrences in training utterances. Adding entities for the four values and labeling them, using prebuilt components such as geography and number where applicable, enables the model to return those values as structured fields for the booking system.

Exam trap

The trap here is assuming intents alone can produce structured values, when only labeled entities yield the extracted fields.

Full explanation →
2
MCQhard

Refer to the exhibit. You send this request to the Conversational Language Understanding API. The response includes the intent 'BookFlight' with entities 'FromCity: Seattle' and 'ToCity: Boston', but the 'Date' entity is missing. What is the most likely cause?

A.The stringIndexType should be 'Utf16CodeUnit'
B.The API version does not support entity extraction
C.The endpoint is pointing to the wrong deployment
D.The model was not trained to recognize date entities
AnswerD

Entity extraction depends entirely on labelled training utterances. If the model was never trained with examples containing date entities, the runtime cannot recognise them, so 'Date' is omitted even though intent and city entities resolve correctly.

Why this answer

The Conversational Language Understanding (CLU) API returns only intents and entities that the deployed model was explicitly trained to recognize. If the training data did not include labeled 'Date' entities, the model will not extract them regardless of the input text. The API itself supports entity extraction, and the endpoint and string index type settings do not affect whether a specific entity type is recognized.

Exam trap

The trap here is that candidates may assume the API automatically extracts common entities like dates (similar to LUIS's prebuilt entities), but CLU requires all entities to be explicitly defined and trained in the model.

How to eliminate wrong answers

Option A is wrong because the stringIndexType parameter (e.g., 'Utf16CodeUnit') controls how offsets are returned, not whether entities are extracted; it has no impact on entity recognition. Option B is wrong because all stable versions of the Conversational Language Understanding API (e.g., 2023-04-01) support entity extraction as a core feature. Option C is wrong because pointing to the wrong deployment would cause a deployment-not-found error or return results from a different model, but the response correctly returned the 'BookFlight' intent and two city entities, indicating the correct deployment was used.

Full explanation →
3
MCQhard

A company uses Azure AI Language Service with Custom Entity Recognition to extract invoice fields. The model correctly extracts invoice numbers but fails to extract dates in the format 'dd/mm/yyyy'. The training data includes dates in 'mm/dd/yyyy' format. What is the most likely issue?

A.The training data does not contain examples with the 'dd/mm/yyyy' format
B.The dates exceed the maximum entity length
C.The language detection is incorrectly identifying the locale
D.The model is overfitting to invoice numbers
AnswerA

Custom Entity Recognition learns patterns from labelled examples, so a model trained only on 'mm/dd/yyyy' dates has no signal for the day-first format. Adding 'dd/mm/yyyy' examples is required for it to extract those dates correctly.

Why this answer

Custom Entity Recognition in Azure AI Language Service learns patterns from labeled training data. Since the training data only contains dates in 'mm/dd/yyyy' format, the model has not seen any examples of 'dd/mm/yyyy' and therefore cannot generalize to that format. The model relies on the exact token sequences and date structures present in the training set, so missing format variations directly cause extraction failures.

Exam trap

The trap here is that candidates may assume the model can infer date formats from context or that language detection handles locale-specific formatting, but Custom Entity Recognition strictly learns from labeled examples and does not apply automatic format normalization.

How to eliminate wrong answers

Option B is wrong because the maximum entity length in Custom Entity Recognition is configurable (default 500 characters) and 'dd/mm/yyyy' dates are well within that limit, so length is not the issue. Option C is wrong because language detection is not used for Custom Entity Recognition; the service operates on the provided text without automatic locale detection, and locale is set manually during project creation. Option D is wrong because overfitting to invoice numbers would cause poor performance on other entities, but the model correctly extracts invoice numbers and only fails on dates, indicating a training data coverage gap rather than overfitting.

Full explanation →
4
Multi-Selecthard

Which TWO actions can you take to mitigate the risk of generating harmful content when using Azure OpenAI Service? (Choose two.)

Select 2 answers
A.Set a system message that instructs the model to avoid harmful outputs.
B.Fine-tune the model on a dataset of safe examples.
C.Deploy the model in multiple regions.
D.Configure Azure AI Content Safety filters.
E.Increase the maxTokens parameter to allow longer responses.
AnswersA, D

A system message sets behavioural guardrails at the prompt level, instructing the model to refuse or avoid harmful content before generation. It is a mitigation available through Azure OpenAI Service configuration, complementing rather than replacing platform-level filtering.

Why this answer

Option A is correct because a system message sets the model's behavioral guardrails at inference time, and explicitly instructing the model to refuse or avoid harmful outputs is a documented prompt-engineering mitigation for Azure OpenAI Service. Option D is correct because Azure AI Content Safety filters (the default and customizable content filters in Azure OpenAI) inspect both prompts and completions for categories such as hate, violence, sexual, and self-harm, and block or annotate harmful content before it reaches users. Option B is not correct because fine-tuning on safe examples can shape tone and style but is not a reliable content-safety control and does not replace the platform's content filtering.

Option C is not correct because multi-region deployment addresses availability and latency, not the generation of harmful content. Option E is not correct because increasing maxTokens only allows longer responses and can actually increase exposure to harmful output rather than mitigate it.

Exam trap

The trap is treating fine-tuning as a safety mechanism — candidates pick 'fine-tune on safe examples' because it sounds thorough, but the exam expects you to know that platform-level Content Safety filters plus system messages are the recognized mitigations.

Full explanation →
5
Multi-Selecteasy

Which TWO capabilities are provided by the Azure AI Language service?

Select 2 answers
A.Text translation.
B.Speech-to-text conversion.
C.Custom text classification.
D.Image captioning.
E.Key phrase extraction.
AnswersC, E

Custom text classification is a prebuilt-and-customisable Azure AI Language capability that trains a model on labelled data to assign your own categories to text, satisfying the capability question. It is distinct from translation and speech, which belong to other Azure AI services.

Why this answer

Option C (Custom text classification) is correct because Azure AI Language includes a custom text classification feature that lets you train models to categorize documents or text into user-defined classes. Option E (Key phrase extraction) is correct because Azure AI Language provides prebuilt key phrase extraction to identify the main talking points in unstructured text. Option A (Text translation) is not part of Azure AI Language; translation is handled by Azure AI Translator.

Option B (Speech-to-text conversion) belongs to Azure AI Speech, not Azure AI Language. Option D (Image captioning) is a computer vision capability in Azure AI Vision, not Azure AI Language.

Exam trap

The trap here is that candidates confuse the Azure AI Language service with other Azure AI services (e.g., Translator, Speech, Vision) that handle specific modalities like translation, audio, or images, leading them to select options that belong to those separate services.

Full explanation →
6
MCQeasy

A retailer wants to analyze thousands of product reviews per day and needs to know which aspects of the products customers mention, such as battery life or screen quality, and whether sentiment toward each aspect is positive or negative. The reviews are already stored as text in an Azure SQL Database. Which Azure AI Language feature should you use?

A.Custom text classification with a project trained on labeled reviews
B.Document-level sentiment analysis with the default opinionMining setting disabled
C.Opinion mining with sentiment analysis on the analyze-text endpoint
D.Key phrase extraction on each review to list the most frequent terms
AnswerC

Opinion mining extends sentiment analysis by returning aspect-level assessments, so the response identifies terms such as battery life or screen quality and pairs each with a target and a sentiment. This directly answers the requirement to know which product aspects customers mention and whether sentiment toward each is positive or negative.

Why this answer

Opinion mining, exposed as an extension of sentiment analysis on the analyze-text endpoint, returns aspect-level targets with individual sentiment labels and confidence scores. That output maps directly to the retailer's need to know which product aspects are mentioned and whether sentiment toward each is positive or negative, without the labeling overhead of a custom classification project.

Exam trap

The trap here is stopping at document-level sentiment, which gives one overall polarity and never attributes that polarity to individual product aspects.

Full explanation →
7
MCQeasy

You need to restrict access to an Azure AI Language resource so that only a specific virtual network can call the endpoint. Which configuration should you use?

A.Rotate the shared access keys
B.Enable a service endpoint or private endpoint for the resource
C.Assign a managed identity to the resource
D.Configure an IP firewall rule with the VNet's public IP range
AnswerB

A service endpoint or private endpoint binds the Azure AI Language resource to a specific virtual network, so only traffic from that network reaches the endpoint. This directly enforces the network restriction the scenario requires.

Why this answer

Azure AI Language resources can be isolated to a specific virtual network by enabling either a service endpoint (via the Microsoft.CognitiveServices service tag) or a private endpoint (using Azure Private Link). This configuration ensures that only traffic originating from the designated VNet can reach the resource's endpoint, effectively blocking all public internet access. Service endpoints provide a direct, optimized route from the VNet to the resource, while private endpoints assign a private IP from the VNet to the resource, making it accessible only within the VNet.

Exam trap

The trap here is that candidates often confuse network-level access controls (service/private endpoints) with authentication mechanisms (keys, managed identities) or IP-based firewalls, mistakenly believing that rotating keys or using managed identities can restrict network access, or that a VNet's public IP range is the same as the VNet's internal address space.

How to eliminate wrong answers

Option A is wrong because rotating shared access keys changes the authentication tokens but does not restrict network-level access; any client with the new keys can still call the endpoint from anywhere on the internet. Option C is wrong because assigning a managed identity enables the resource to authenticate to other Azure services (e.g., Azure Key Vault) without storing credentials, but it does not control which networks can reach the resource's endpoint. Option D is wrong because IP firewall rules with the VNet's public IP range are ineffective for restricting access to a specific virtual network, as VNet traffic typically uses private IPs (RFC 1918) and the public IP of a VNet's NAT gateway or load balancer is not the same as the VNet's internal address space; moreover, IP firewall rules cannot distinguish traffic originating from within the VNet versus other sources using the same public IP range.

Full explanation →
8
MCQmedium

You are building a knowledge mining solution for legal documents stored in Azure Blob Storage. The solution must extract entities, key phrases, and relationships from the documents. Which Azure AI service should you use?

A.Azure AI Document Intelligence
B.Azure AI Translator
C.Azure AI Language
D.Azure AI Search
AnswerC

Azure AI Language provides entity recognition, key phrase extraction and relation extraction natively, covering all three requirements in one service. Its custom NER and text analytics capabilities process the legal documents directly, satisfying the stem's demand to extract entities, key phrases and relationships without additional services.

Why this answer

Azure AI Language provides pre-built capabilities for entity extraction, key phrase extraction, and relationship extraction from text. This service includes features like Named Entity Recognition (NER), key phrase extraction, and document analysis that directly meet the requirements for extracting entities, key phrases, and relationships from legal documents. The other options lack one or more of these specific text analytics capabilities.

Exam trap

The trap here is that candidates often confuse Azure AI Document Intelligence (which handles OCR and form extraction) with Azure AI Language (which handles text analytics like entity and key phrase extraction), leading them to pick A when the question explicitly asks for extracting entities, key phrases, and relationships from text.

How to eliminate wrong answers

Option A is wrong because Azure AI Document Intelligence is designed for extracting structured data (like tables, key-value pairs, and text) from scanned documents and forms, not for performing entity, key phrase, or relationship extraction from unstructured text. Option B is wrong because Azure AI Translator focuses on translating text between languages and does not include entity extraction, key phrase extraction, or relationship analysis. Option D is wrong because Azure AI Search is a search indexing and query service that can index data from various sources but does not natively perform entity extraction, key phrase extraction, or relationship extraction; it relies on an external AI enrichment pipeline (often using Azure AI Language) for those tasks.

Full explanation →
9
MCQeasy

You are testing an Azure OpenAI model with the parameters shown in the exhibit. The model generates very short responses. Which parameter should you modify to allow longer responses?

A.Increase frequency_penalty
B.Increase top_p
C.Increase temperature
D.Increase max_tokens
AnswerD

Raising max_tokens directly lifts the hard ceiling on generated output length, which is the constraint producing truncated, very short responses. Temperature, top_p and frequency_penalty shape token selection, not response length, so they cannot extend output beyond the cap. Microsoft Entra ID is unrelated to inference parameters.

Why this answer

The max_tokens parameter controls the maximum number of tokens (words or subwords) the model can generate in a single response. When responses are very short, increasing max_tokens allows the model to produce longer completions up to the specified limit. The other parameters affect randomness, diversity, or probability distribution, not the length cap.

Exam trap

The trap here is that candidates confuse parameters that control output length (max_tokens) with those that control output diversity or creativity (temperature, top_p, frequency_penalty), leading them to incorrectly adjust the latter when the real issue is a token limit.

How to eliminate wrong answers

Option A is wrong because frequency_penalty reduces the likelihood of repeating the same tokens or phrases, which can actually shorten responses by discouraging repetition, not lengthen them. Option B is wrong because top_p (nucleus sampling) controls the cumulative probability threshold for token selection, affecting diversity but not the maximum output length. Option C is wrong because temperature adjusts the randomness of token selection (higher = more creative, lower = more deterministic) and does not impose or remove a length constraint.

Full explanation →
10
MCQhard

You are building a custom named entity recognition (NER) model using Azure AI Language. After labeling 200 documents, you train the model and achieve 85% precision but only 60% recall. Which action is most likely to improve recall?

A.Lower the confidence threshold
B.Increase the training hours
C.Increase the number of labeled documents, especially those containing the target entities
D.Switch to a different Azure AI Language feature
AnswerC

Low recall means the model misses many true entities, typically because training data under-represents their contexts and variations. Adding more labelled documents containing the target entities exposes the model to those patterns, raising recall while precision stays broadly stable.

Why this answer

Low recall in a custom NER model typically indicates that the model is failing to identify many instances of the target entities. Increasing the number of labeled documents, especially those containing the target entities, provides more positive examples for the model to learn from, directly improving its ability to recognize those entities and thus boosting recall.

Exam trap

The trap here is that candidates often confuse confidence threshold tuning with a data quality fix, thinking lowering the threshold will magically fix recall, when in reality it only trades precision for recall without addressing the root cause of insufficient training examples.

How to eliminate wrong answers

Option A is wrong because lowering the confidence threshold would increase the number of predictions (including false positives), which could improve recall but at the cost of significantly reducing precision, and it does not address the underlying issue of insufficient training examples for the target entities. Option B is wrong because increasing training hours does not improve model performance if the training data is insufficient or imbalanced; the model will simply overfit or plateau without more labeled examples. Option D is wrong because switching to a different Azure AI Language feature (e.g., from custom NER to pre-built entity extraction) would not solve the recall problem for custom entities, as pre-built features are not designed to recognize domain-specific entities.

Full explanation →
11
MCQmedium

You are using Azure AI Language's custom question answering feature. Your knowledge base contains 500 FAQ pairs. Users report that the bot returns irrelevant answers when their questions are phrased differently from the stored FAQs. You want to improve the bot's ability to match paraphrased questions without retraining a full model. Which action should you take?

A.Enable the 'chit-chat' personality in the project settings to make responses more conversational.
B.Increase the confidence threshold in the project settings to filter low-scoring answers.
C.Add alternative question phrasings to each FAQ pair in the knowledge base.
D.Switch the project to use a custom text classification model instead of question answering.
AnswerC

Custom question answering uses the alternative questions as additional training data to learn semantic matches. Adding paraphrased variations directly improves the model's ability to match differently worded user queries to the correct FAQ, without requiring a separate retraining pipeline.

Why this answer

Custom question answering learns from the question-answer pairs and their alternative phrasings. Adding alternative questions provides the model with more surface forms of the same intent, which improves semantic matching for paraphrased user queries without requiring a separate model training process.

Exam trap

The trap here is assuming that tuning thresholds or enabling chit-chat improves semantic matching, when the real lever is adding alternative question phrasings to the knowledge base.

Full explanation →
12
MCQmedium

You are building a knowledge mining solution for a legal firm to extract clauses from contracts. The contracts are stored as PDFs in Azure Blob Storage. You need to design the solution to minimize cost while ensuring high accuracy for clause extraction. Which approach should you use?

A.Use Azure AI Custom Vision to detect clause regions in scanned documents.
B.Use Azure OpenAI GPT-4 to process each PDF and extract clauses using prompts.
C.Use Azure AI Search with a blob indexer to extract clauses during indexing.
D.Use Azure AI Document Intelligence with a custom extraction model trained on contract clauses.
AnswerD

A custom Document Intelligence extraction model learns the specific clause layouts and terminology in the firm's contracts, delivering high accuracy. It is consumption-priced per page, avoiding the higher ongoing cost of building and hosting a bespoke machine learning pipeline, satisfying the minimise-cost constraint.

Why this answer

Azure AI Document Intelligence (formerly Form Recognizer) with a custom extraction model is designed to extract structured data from documents like contracts, and it supports training on your own labeled clauses for high accuracy. It is cost-effective for batch processing of PDFs and integrates with Azure AI Search for knowledge mining. This approach balances accuracy and cost better than the alternatives.

Exam trap

AI-102 often tests the service selection trap, where candidates choose Azure OpenAI or Custom Vision for document extraction, but the correct tool for structured clause extraction with cost efficiency is Document Intelligence custom models.

How to eliminate wrong answers

Option A is wrong because Custom Vision is an image classification/object detection service, not designed for text extraction from documents; it would require converting PDFs to images and would not accurately extract clause text. Option B is wrong because using GPT-4 for each PDF is expensive at scale and may not guarantee consistent, structured extraction; it is also overkill for clause extraction when a specialized model exists. Option C is wrong because Azure AI Search with a blob indexer extracts text content but does not perform clause-level extraction or custom entity recognition; it would index the full text, not isolate clauses.

Full explanation →
13
MCQmedium

Refer to the exhibit. You called the Named Entity Recognition API on a document. Which entity type is "Seattle"?

A.Organization
B.Location
C.Person
D.City
AnswerB

"Seattle" denotes a geographic place, so the Named Entity Recognition model classifies it under the Location entity category, which covers cities, countries and regions. This satisfies the stem's requirement to identify the entity type returned for a place name, rather than Person, Organisation or DateTime.

Why this answer

The Named Entity Recognition (NER) API in Azure AI Language identifies 'Seattle' as a Location entity because it is a recognized geographical place. The API uses a pre-trained model that categorizes entities into types such as Location, Person, Organization, etc., and 'Seattle' falls under the Location type based on its semantic context in the document.

Exam trap

The trap here is that candidates may confuse the specific instance (e.g., 'City') with the official entity type label used by the API, leading them to choose 'City' instead of the correct 'Location' type.

How to eliminate wrong answers

Option A is wrong because 'Seattle' is not an organization; it is a city, and the NER API would classify it as a Location, not an Organization (which typically refers to companies, agencies, or institutions). Option C is wrong because 'Seattle' is not a person; the NER API's Person type is reserved for names of individuals, not places. Option D is wrong because 'City' is not a standard entity type in the NER API's output; the API uses broader categories like Location, and 'City' is a subtype or specific instance within Location, not a top-level entity type.

Full explanation →
14
MCQmedium

You are implementing a Retrieval Augmented Generation pattern with Azure AI Search and Azure OpenAI. Users complain that answers occasionally cite documents the user is not authorized to see, because the index contains all departments' content. You need to restrict retrieval so that each user only receives chunks from documents their identity permits. What should you do?

A.Create a separate index per department and have users manually select which index to search.
B.Add a system message instructing the model to ignore any content the user is not allowed to see.
C.Apply a semantic ranker to the query and rely on relevance scoring to push unauthorized content lower in the results.
D.Add a security filter on a document access field in Azure AI Search, and pass the user's group or object IDs as an OData filter in the query.
AnswerD

Azure AI Search supports filtering on fields such as a collection of allowed group IDs. By storing each document's permitted principals and passing the caller's group or object IDs from the token into an OData filter, the query returns only chunks the identity is authorized to read. This enforces security at retrieval time, which is the correct layer for RAG, before content ever reaches the model or the user.

Why this answer

Security trimming for RAG must occur at retrieval. Storing each document's allowed principals in a filterable field and passing the caller's group or object IDs as an OData filter ensures the search response never contains unauthorized chunks. Relevance ranking, prompt instructions, and manual index selection do not enforce authorization and can leak restricted content before the model or user sees it.

Exam trap

The trap here is treating relevance ranking or prompt instructions as if they provided access control.

Full explanation →
15
MCQmedium

Your team is building a generative AI application on Azure OpenAI that must call an internal order-lookup REST API whenever a user asks about an order status. The model must decide when to call the API, and the application must execute the call and return the result to the model for the final answer. Which capability should you implement?

A.Semantic ranker enabled on an Azure AI Search index that stores order records.
B.Function calling (tools) on the chat completions request, defining the order-lookup API as a tool.
C.A system message instructing the model to answer order questions only from its training data.
D.A JSON mode response format that forces the model to emit a schema-conforming order status object.
AnswerB

Function calling lets you describe available tools in the request; the model then returns a structured tool call when it determines the API is needed. Your application executes the call and sends the result back as a tool message, after which the model composes the final answer. This exactly matches the required decide-execute-return loop.

Why this answer

The requirement is a model-driven decision to invoke a live REST API and then incorporate the returned data. Function calling provides exactly that contract: tools are declared, the model emits a structured call, the host application performs the HTTP request, and the result is fed back for final generation. The other options either constrain formatting, rely on training data, or only rank search results.

Exam trap

The trap here is equating structured output formatting, such as JSON mode, with tool invocation, when only function calling produces a call the application can execute.

Full explanation →
16
MCQmedium

You are building a solution that uses Azure AI Language to analyze transcribed call-center conversations. The transcripts are stored as plain text in Azure Blob Storage. You need to identify the specific products, dates, and monetary amounts mentioned in each conversation while distinguishing them from generic nouns. You also need to return the character offset and length for each detected mention so the UI can highlight them. Which Azure AI Language feature should you use?

A.Key Phrase Extraction
B.Named Entity Recognition (NER)
C.Extractive Summarization
D.Conversation Summarization
AnswerB

NER returns entity categories such as Product, DateTime, and Quantity with the exact text, offset, and length for each mention, which supports highlighting in the UI. It distinguishes specific entity types from generic tokens, matching the requirement to isolate products, dates, and money amounts from ordinary nouns.

Why this answer

Named Entity Recognition in Azure AI Language is designed to detect entity categories such as Product, DateTime, and Quantity and to return each mention with its offset and length. Those offsets allow the application to highlight exact spans in the transcript, and the category labels let developers filter products from dates and amounts.

Exam trap

The trap here is assuming any phrase extraction feature returns entity categories and offsets, when only NER provides typed entities with span metadata.

Full explanation →
17
MCQmedium

You are designing an AI solution that uses Azure AI Document Intelligence to extract data from invoices. The solution must handle a high volume of documents with varying layouts. Which approach should you use?

A.Use a custom template model with fixed field positions
B.Train a custom neural model with labeled invoices
C.Use the prebuilt invoice model
D.Extract text using OCR and then use regex parsing
AnswerB

A custom neural model handles varying invoice layouts because it learns layout and field patterns from labelled examples rather than relying on fixed templates. This satisfies the high-volume, layout-variance constraint, whereas prebuilt models assume consistent formats and would degrade across suppliers.

Why this answer

Custom neural models in Azure AI Document Intelligence are designed to handle high volumes of documents with varying layouts. Unlike fixed template models, neural models learn from labeled examples and generalize across different invoice structures, making them ideal for diverse, high-volume scenarios.

Exam trap

The trap here is that candidates often assume the prebuilt invoice model (Option C) is sufficient for all invoice scenarios, but it only works for standard layouts and fails when invoices have custom fields or non-standard structures.

How to eliminate wrong answers

Option A is wrong because custom template models rely on fixed field positions and are brittle when layouts vary; they fail if the invoice format changes even slightly. Option C is wrong because the prebuilt invoice model is limited to standard invoice layouts and cannot adapt to custom or highly variable formats, leading to poor extraction accuracy. Option D is wrong because OCR with regex parsing is a brittle, rule-based approach that cannot handle the semantic understanding required for diverse invoice layouts and fails when fields are not in predictable positions or formats.

Full explanation →
18
MCQhard

You are designing a solution that must extract personally identifiable information (PII) from medical records stored in Azure Blob Storage. The solution must redact the PII before storing the results. Which combination of Azure services should you use?

A.Use Azure AI Language's PII detection feature and a custom Azure Function to redact.
B.Use Azure AI Search with cognitive skills for PII detection.
C.Use Azure OpenAI to detect and redact PII.
D.Use Text Analytics for Health and then manually redact.
AnswerA

Azure AI Language's PII detection returns entity offsets and categories, letting a custom Azure Function mask or replace each span before writing results back to Blob Storage. This satisfies the stem's redaction-before-storage constraint, since the Function performs the redaction step rather than relying on a service that only detects.

Why this answer

Azure AI Language's PII detection feature is specifically designed to identify and categorize PII entities in text, and combining it with a custom Azure Function allows you to programmatically redact those entities before storing the results in Blob Storage. This provides a serverless, scalable pipeline that meets the requirement of extracting and redacting PII from medical records without manual intervention.

Exam trap

The trap here is that candidates often confuse Text Analytics for Health (which is for medical entity extraction) with Azure AI Language's PII detection (which is for privacy compliance), leading them to choose Option D despite its lack of redaction capabilities.

How to eliminate wrong answers

Option B is wrong because Azure AI Search with cognitive skills is primarily for indexing and enriching searchable content, not for direct PII redaction before storage; it would require additional custom logic to achieve redaction. Option C is wrong because Azure OpenAI is a general-purpose language model that lacks built-in, deterministic PII detection and redaction capabilities, and relying on it for compliance-grade PII handling introduces risks of inconsistent or incomplete redaction. Option D is wrong because Text Analytics for Health is optimized for extracting medical entities (e.g., diagnoses, medications) and does not natively support PII detection or redaction; manual redaction is error-prone and violates the automation requirement.

Full explanation →
19
MCQmedium

You are building an Azure AI Language solution that analyzes customer support emails. You need to detect the language of each email before routing it to the correct translation pipeline. You call the Language Detection API with the following request body: {"kind": "LanguageDetection", "analysisInput": {"documents": [{"id": "1", "text": "Bonjour, je besoin d'aide avec mon compte."}]}}. What will the response contain?

A.The translated version of the text in English, because the default target language is English.
B.A list of all supported languages with confidence scores for each, allowing you to select the highest-scoring language.
C.An error because the text contains an apostrophe, which is not supported in the JSON payload.
D.A single document result with the detected language code 'fr', a confidence score, and the original text with offsets.
AnswerD

The Language Detection API returns the detected language ISO code, a confidence score between 0 and 1, and the analyzed text with character offsets for each document. For the provided French text, the response includes the language code 'fr' and a high confidence score, enabling downstream routing to French-specific processing.

Why this answer

The Language Detection API returns the detected language code, confidence score, and text metadata for each document. It does not return all languages, translated text, or errors for common punctuation. The correct response includes the ISO code 'fr' and a confidence score, which you can use to route the email appropriately.

Exam trap

The trap here is assuming the API returns multiple candidate languages or performs translation, when it only identifies the single most likely language.

Full explanation →
20
MCQeasy

You need to provide a generative AI solution that can answer questions based on a large set of PDF documents stored in Azure Blob Storage. The solution must support natural language queries and return citations from the documents. Which Azure service combination should you use?

A.Azure Machine Learning and Azure Kubernetes Service
B.Azure Cognitive Search and Azure OpenAI Service with 'on your data'
C.Azure AI Bot Service and Azure Functions
D.Azure AI Document Intelligence and Azure AI Translator
AnswerB

Azure Cognitive Search indexes the Blob-stored PDFs and retrieves relevant passages, while Azure OpenAI Service with 'on your data' grounds responses in those results and returns citations. This pairing satisfies both the natural-language query and citation requirements.

Why this answer

Azure Cognitive Search provides the indexing and retrieval capabilities for the PDF content, while Azure OpenAI Service with the 'on your data' feature enables natural language querying and generates answers grounded in the indexed documents, including citations. This combination directly supports the requirement to query a large set of PDFs in Azure Blob Storage and return citations.

Exam trap

The trap here is that candidates often confuse Azure AI Document Intelligence (which extracts text) with the full search-and-generate pipeline, overlooking that a search index (Cognitive Search) and a generative model (Azure OpenAI) are both required to answer natural language queries with citations.

How to eliminate wrong answers

Option A is wrong because Azure Machine Learning and Azure Kubernetes Service are designed for training and deploying custom ML models, not for out-of-the-box document indexing, natural language querying, or citation generation from PDFs. Option C is wrong because Azure AI Bot Service and Azure Functions are used for building conversational bots and serverless compute, but they lack native capabilities for indexing PDF content and returning citations from documents. Option D is wrong because Azure AI Document Intelligence extracts text and structure from documents, and Azure AI Translator handles language translation, but neither service provides the search indexing or generative AI querying needed to answer natural language questions with citations.

Full explanation →
21
Multi-Selectmedium

You are deploying an Azure AI Language custom text classification model. You need to ensure the model meets performance requirements before promoting it to production. Which two actions should you take? (Choose two.)

Select 2 answers
A.Evaluate the model on a held-out test set that was not used during training.
B.Review the confusion matrix to understand which classes are frequently misclassified.
C.Ensure the model achieves at least 95% accuracy on a cross-validation split.
D.Use the training set to compute accuracy and ensure it is above 90%.
E.Compare the model's performance to a baseline model that always predicts the most common class.
AnswersA, B

Evaluating on a held-out test set never seen during training yields unbiased precision, recall and F1 scores, confirming the model generalises rather than memorising training data. This validation gate must pass before promotion to production, satisfying the performance-verification requirement.

Why this answer

Option A is correct because evaluating the custom text classification model on a held-out test set that was never used during training provides an unbiased estimate of generalization performance, which is essential before promoting the model to production. Option B is correct because reviewing the confusion matrix reveals per-class precision/recall patterns and shows exactly which classes are frequently misclassified, enabling targeted data or label improvements. Option C is incorrect because Azure AI Language does not require a fixed 95% accuracy threshold on a cross-validation split; performance targets are scenario-specific and cross-validation is not the standard evaluation workflow for this service.

Option D is incorrect because computing accuracy on the training set measures memorization rather than generalization and will be optimistically biased. Option E is incorrect because comparing to a majority-class baseline is a useful sanity check but is not one of the required actions for validating a custom text classification model before production.

Exam trap

The trap here is that candidates often assume a fixed accuracy threshold (like 95%) is required for production promotion, but Microsoft Azure AI Language custom text classification does not mandate any specific metric value—the focus is on evaluating generalization via a held-out test set and analyzing misclassifications with the confusion matrix.

Full explanation →
22
MCQeasy

A company wants to build a solution that can identify and redact personally identifiable information (PII) from customer support transcripts. The solution must handle multiple languages. Which Azure AI service should be used?

A.Azure AI Content Safety
B.Azure AI Document Intelligence
C.Azure AI Translator
D.Azure AI Language - PII Detection
AnswerD

Azure AI Language's PII detection feature natively identifies and redacts personal data across multiple languages, directly satisfying the stem's multilingual requirement. Its prebuilt entity models recognise names, addresses, and phone numbers in transcripts without custom training, unlike translation or vision services that lack redaction capability.

Why this answer

Azure AI Language's PII Detection feature is specifically designed to identify and redact personally identifiable information in text across multiple languages. It supports over 30 languages and can detect entities such as names, addresses, phone numbers, and credit card numbers, making it the correct choice for this multilingual PII redaction requirement.

Exam trap

The trap here is that candidates may confuse Azure AI Language's PII detection with Azure AI Content Safety, assuming 'safety' includes privacy, but Content Safety addresses content moderation (e.g., toxicity) rather than PII redaction.

How to eliminate wrong answers

Option A is wrong because Azure AI Content Safety focuses on detecting harmful or offensive content (e.g., hate speech, self-harm) rather than PII entities. Option B is wrong because Azure AI Document Intelligence is optimized for extracting structured data from documents (e.g., invoices, forms) and does not provide native PII detection or redaction capabilities. Option C is wrong because Azure AI Translator is a machine translation service that translates text between languages but does not identify or redact PII; it can be used alongside PII detection but is not a standalone solution for this task.

Full explanation →
23
MCQeasy

You are creating an Azure AI Search index that will be populated from an enrichment pipeline. You need to ensure that the original content of each document is searchable. Which index field should you map the document content to?

A.A field with the key attribute set to true
B.A field with the retrievable attribute set to true
C.A field with the searchable attribute set to true
D.A field with the filterable attribute set to true
AnswerC

A searchable field is analyzed and included in the full-text search index. By mapping the document content to a searchable field, you enable users to query the content using keywords. This is the standard way to make document text searchable. The searchable attribute must be set to true for the field to be queried.

Why this answer

To make document content searchable, the index field must have the searchable attribute set to true. This enables full-text search capabilities, allowing users to query the content using keywords and phrases. Other attributes like filterable or retrievable serve different purposes and do not provide search functionality.

Therefore, mapping content to a searchable field is essential.

Exam trap

The trap here is confusing the retrievable attribute with searchability, assuming that if a field is returned in results it must also be searchable, which is not the case.

Full explanation →
24
MCQeasy

A developer needs to generate a descriptive caption for an image using Azure AI Vision. The image contains a dog catching a frisbee in a park. Which feature of Image Analysis should they use?

A.Caption
B.Read
C.Objects
D.Tags
AnswerA

The caption feature in Image Analysis generates a human-readable sentence that describes the image content, such as 'a dog catching a frisbee in a park'. It is specifically designed to produce descriptive captions. This directly meets the developer's requirement for a descriptive caption.

Why this answer

The caption feature of Azure AI Vision Image Analysis generates a natural language description of an image's content. It is the only feature that produces a sentence like 'a dog catching a frisbee in a park'. Tags, objects, and read serve different purposes: tags provide keywords, objects detect and locate items, and read extracts text.

Therefore, caption is the correct choice.

Exam trap

The trap here is confusing tags with captions, as both provide descriptive information, but only captions produce a full sentence.

Full explanation →
25
MCQhard

A developer is using the Azure AI Vision Image Analysis API to extract text from a photo of a street sign. The sign contains text in both English and Japanese arranged in multiple columns. The developer needs the response to include the detected language and the bounding box for each text line. Which feature should they use?

A.The 'read' feature in Image Analysis
B.The 'detectObjects' feature in Image Analysis
C.The 'caption' feature in Image Analysis
D.The 'tags' feature in Image Analysis
AnswerA

The read feature in Image Analysis performs OCR and returns extracted text along with bounding boxes for lines and words. It also detects the language of the text. This directly provides the language and bounding box information required, and it handles mixed-language and multi-column text in a single call.

Why this answer

The read feature in Azure AI Vision Image Analysis is specifically designed for OCR. It returns extracted text with bounding boxes for lines and words, and it detects the language of the text. This matches the need to capture English and Japanese text in multiple columns and provide bounding boxes for each line.

Other features like caption, detectObjects, and tags do not perform OCR.

Exam trap

The trap here is confusing general image analysis features like captioning or tagging with OCR, when only the read feature extracts text and provides bounding boxes and language detection.

Full explanation →
26
MCQeasy

A financial services firm needs to extract structured fields such as invoice date, vendor name, and total amount from thousands of PDF invoices. The documents vary in layout across vendors. The firm wants a pretrained model that requires no custom training and can return field-level confidence scores. Which Azure AI Document Intelligence model should they use?

A.Read model
B.Custom template model
C.Prebuilt invoice model
D.General document model
AnswerC

The prebuilt invoice model is trained to extract invoice-specific fields including invoice date, vendor name, and total amount, and it returns confidence scores per field. Because it is pretrained, no custom labeling or training is required, which matches the firm's requirement to process varied vendor layouts quickly.

Why this answer

The firm needs invoice-specific structured fields with confidence scores and no custom training. The prebuilt invoice model in Azure AI Document Intelligence is designed exactly for that, recognizing common invoice fields across varied layouts. General document, custom template, and Read models either lack invoice-specific fields, require training, or return only raw text.

Exam trap

The trap here is choosing a custom model when a pretrained invoice model already covers the required fields without training.

Full explanation →
27
MCQeasy

You need to build a solution that extracts the main topics discussed in recorded customer service calls. The audio is already transcribed to text, and you must return the most salient phrases without any predefined categories. Which Azure AI Language feature should you use?

A.Named entity recognition
B.Extractive summarization
C.Key phrase extraction
D.Custom text classification
AnswerC

Key phrase extraction returns the main talking points from a document without requiring predefined categories. Because the transcripts have no labels and the goal is to surface salient topics, this feature matches the requirement directly. It is available through the analyze-text endpoint and returns a list of key phrases per document, which can be aggregated across calls to summarize discussion themes.

Why this answer

Key phrase extraction is the Azure AI Language feature designed to surface the main talking points in text without any predefined categories or training. It returns a list of salient phrases per document, which can be aggregated across call transcripts to reveal discussion themes. Custom classification needs labeled classes, entity recognition targets named entities rather than topics, and extractive summarization returns sentences instead of topic phrases.

Exam trap

The trap here is conflating summarization with key phrase extraction, when the requirement for short salient phrases without predefined categories points specifically to key phrase extraction.

Full explanation →
28
MCQhard

You are a data scientist at a healthcare startup. You have deployed a custom object detection model using Azure Custom Vision to detect tumors in MRI scans. The model was trained on 10,000 labeled scans from a single hospital. After deployment, the model performs well on scans from that hospital but poorly on scans from a different hospital with a different MRI machine. The new hospital's scans have slightly different contrast and resolution. The model's precision drops from 0.92 to 0.65, and recall drops from 0.88 to 0.50. You have access to 500 labeled scans from the new hospital. You need to improve the model's performance on the new hospital's data as quickly as possible with minimal effort. What should you do?

A.Collect more labeled scans from the new hospital and train a new model from scratch.
B.Create a new Custom Vision project and train only on the 500 new scans.
C.Apply image preprocessing to normalize the new hospital's scans to match the old hospital's style, then use the existing model.
D.Use the existing model as a starting point and retrain it with the 500 labeled scans from the new hospital.
AnswerD

Domain shift from differing contrast and resolution causes the precision and recall drop. Retraining the existing Custom Vision model with the 500 labelled new-hospital scans performs transfer learning, adapting features to the new distribution quickly and with minimal effort.

Why this answer

Azure Custom Vision supports transfer learning, allowing you to take an existing trained model and retrain it with new labeled data. By using the 500 labeled scans from the new hospital as a training set, you can fine-tune the model to adapt to the different contrast and resolution characteristics without starting from scratch. This approach is the fastest and requires minimal effort, leveraging the previously learned features while incorporating domain-specific adjustments.

Exam trap

The trap here is that candidates may overestimate the need for large datasets or manual preprocessing, failing to recognize that Azure Custom Vision's built-in transfer learning is designed to efficiently adapt models with minimal new data.

How to eliminate wrong answers

Option A is wrong because collecting more labeled scans and training a new model from scratch is time-consuming and resource-intensive, not the quickest or minimal-effort solution. Option B is wrong because creating a new Custom Vision project and training only on 500 scans ignores the valuable knowledge from the original 10,000 scans, leading to a model with insufficient data and likely poor generalization. Option C is wrong because applying image preprocessing to normalize the new hospital's scans to match the old hospital's style is a manual, error-prone process that may not fully address the underlying domain shift and does not leverage the labeled data for supervised adaptation.

Full explanation →
29
MCQhard

You are managing an Azure AI solution that uses Azure OpenAI. You need to ensure that the solution can only be accessed by applications hosted in a specific Azure Virtual Network. The solution must not be accessible from the public internet. What should you configure?

A.Azure OpenAI managed identity with role-based access control (RBAC).
B.Azure OpenAI network security group (NSG) on the resource's subnet.
C.Azure OpenAI firewall with IP restrictions for the virtual network's NAT gateway.
D.Azure OpenAI private endpoint and disable public network access.
AnswerD

A private endpoint creates a private IP address for the Azure OpenAI resource within your virtual network, and disabling public network access ensures that the resource is not reachable from the internet. This combination restricts access to only resources within the specified virtual network or connected networks, meeting the requirement for private, non-public access.

Why this answer

To restrict Azure OpenAI access to only a specific virtual network and prevent public internet access, you should configure a private endpoint for the Azure OpenAI resource and disable public network access. This ensures that the resource is only accessible via private IP addresses within the virtual network, and all public access is blocked, satisfying the strict isolation requirement.

Exam trap

The trap here is confusing authentication and authorization controls like managed identity or RBAC with network isolation, or assuming that IP restrictions based on a NAT gateway provide the same level of private access as a private endpoint with public access disabled.

Full explanation →
30
MCQeasy

You are deploying an Azure AI solution that uses multiple Cognitive Services resources. You need to ensure that the solution can be deployed to multiple regions and that each region has its own endpoint and key. What should you use to manage the deployment?

A.A single Cognitive Services resource with multiple endpoints configured in the application code.
B.Azure Resource Manager (ARM) templates with parameters for region-specific values.
C.Azure Policy definitions that enforce the creation of Cognitive Services resources in specific regions.
D.Azure CLI scripts that create resources in each region sequentially.
AnswerB

ARM templates allow you to define parameters that can be supplied at deployment time, enabling the same template to be used for multiple regions. You can parameterize the region, resource names, and SKUs. This provides a repeatable, consistent deployment method and supports multi-region scenarios with region-specific endpoints and keys.

Why this answer

ARM templates are the recommended infrastructure-as-code tool for deploying Azure resources consistently across multiple regions. By using parameters, you can customize region-specific settings such as location, endpoint names, and SKUs. This approach ensures repeatability and simplifies management of multi-region deployments.

Exam trap

The trap here is confusing Azure Policy (which enforces rules) with ARM templates (which deploy resources), or assuming a single resource can span regions.

Full explanation →
31
MCQmedium

A security firm wants to analyze live video from cameras at a warehouse gate to count the number of people entering and leaving. They require a ready-to-use service that provides a real-time count and does not require training a custom model. Which Azure AI service should they use?

A.Azure AI Custom Vision
B.Azure AI Face
C.Azure AI Vision Spatial Analysis
D.Azure Video Indexer
AnswerC

Azure AI Vision Spatial Analysis is designed for real-time video analytics, including people counting, and does not require custom model training. It ingests RTSP streams and provides counts of people crossing a defined line or zone. This matches the requirement for a ready-to-use service that counts people entering and leaving.

Why this answer

The security firm needs a ready-to-use service for real-time people counting from live video. Azure AI Vision Spatial Analysis is purpose-built for this scenario, offering pre-trained models that analyze RTSP streams and count people crossing lines or zones without custom training. Other services either focus on static images or require significant custom development, making them unsuitable for immediate deployment.

Exam trap

The trap here is assuming that any Azure AI vision service can process live video streams for people counting, when only Spatial Analysis provides that out-of-the-box capability.

Full explanation →
32
MCQhard

Your team is developing an AI-powered document summarization solution using Azure OpenAI. You need to ensure that the solution complies with Microsoft's Responsible AI principles, specifically transparency. Which configuration should you implement?

A.Configure diagnostic logging to capture all model inputs and outputs.
B.Fine-tune the model on a custom dataset to improve accuracy.
C.Enable content filtering with severity levels high and medium.
D.Add a system message that informs users the summary is generated by AI.
AnswerD

A system message disclosing AI generation directly satisfies transparency by informing users they are reading machine-generated output. Unlike content filters or metadata logging, which address harm prevention and auditability, this mechanism makes the AI's role visible at the point of interaction, meeting the stem's explicit transparency requirement.

Why this answer

Transparency under Microsoft's Responsible AI principles requires that users are aware when they are interacting with an AI system. Adding a system message that explicitly states the summary is AI-generated fulfills this disclosure requirement. Diagnostic logging (A) aids in accountability and debugging but does not directly inform the user.

Fine-tuning (B) improves accuracy but does not address transparency. Content filtering (C) mitigates harmful outputs but does not disclose AI involvement.

Exam trap

The trap here is that candidates confuse 'transparency' with 'accountability' or 'safety' and select diagnostic logging or content filtering, not realizing that transparency specifically requires user-facing disclosure of AI involvement.

How to eliminate wrong answers

Option A is wrong because diagnostic logging captures inputs and outputs for auditing and debugging, but it does not communicate to the end user that the content is AI-generated, which is the core requirement of transparency. Option B is wrong because fine-tuning the model on a custom dataset enhances performance and relevance but has no role in informing users about AI authorship. Option C is wrong because enabling content filtering with severity levels high and medium is a safety measure to block harmful content, not a mechanism for disclosing AI involvement to users.

Full explanation →
33
Multi-Selectmedium

A company is building an agent that uses Azure OpenAI to answer questions from a large document library. The agent must use a Retrieval Augmented Generation (RAG) pattern. Which TWO actions should the team take to implement RAG effectively?

Select 2 answers
A.Ensure the model is large enough to memorize the entire document library.
B.Fine-tune the Azure OpenAI model on the document library.
C.Index the documents into a vector database like Azure Cognitive Search.
D.Train a custom language model from scratch.
E.Use a retrieval step to fetch relevant document chunks before generating a response.
AnswersC, E

RAG requires a searchable vector index so the agent can retrieve semantically similar chunks. Indexing documents into Azure Cognitive Search with embeddings satisfies the retrieval constraint, enabling grounded answers rather than relying solely on the model's parametric knowledge.

Why this answer

Option C is correct because RAG requires an external knowledge store that supports semantic similarity search, and indexing the documents into a vector database such as Azure Cognitive Search (or Azure AI Search) creates embeddings that let the agent retrieve the most relevant passages at query time. Option E is correct because the defining step of Retrieval Augmented Generation is retrieving the top-k relevant document chunks and injecting them into the prompt as grounding context before the model generates its answer, which keeps responses accurate and current without retraining. Option A is not appropriate because no model can memorize an entire large document library, and relying on memorization defeats the purpose of retrieval.

Option B is not appropriate because fine-tuning teaches style and task behavior rather than reliably storing and retrieving factual document content, and it is not the RAG mechanism. Option D is not appropriate because training a custom language model from scratch is prohibitively expensive and unnecessary when Azure OpenAI plus retrieval already solves the problem.

Exam trap

The trap here is that candidates often confuse fine-tuning (which adapts model behavior) with RAG (which augments prompts with retrieved data), leading them to select Option B instead of understanding that RAG requires an external retrieval step and vector index.

Full explanation →
34
MCQhard

You are designing a solution that uses Azure AI Vision to extract text from scanned invoices. The invoices vary in layout and include both printed and handwritten fields. The solution must achieve high accuracy with minimal manual labeling. Which approach should you recommend?

A.Use the Read API to extract all text and then use a custom regex to parse fields.
B.Use the Azure AI Document Intelligence prebuilt invoice model.
C.Train a Custom Vision object detection model to locate fields.
D.Label hundreds of invoices and train a custom Azure AI Document Intelligence model.
AnswerB

The prebuilt invoice model in Azure AI Document Intelligence is trained on varied invoice layouts and extracts printed and handwritten fields with high accuracy, requiring no custom labelling. This satisfies the minimal manual labelling constraint for the scanned invoices.

Why this answer

The Azure AI Document Intelligence prebuilt invoice model is specifically designed to extract common fields from invoices with high accuracy, handling both printed and handwritten text without requiring manual labeling. It uses advanced OCR and deep learning models trained on thousands of invoices, making it ideal for varied layouts and minimal setup.

Exam trap

The trap here is that candidates often overestimate the need for custom training or regex-based parsing, not realizing that Azure provides a prebuilt, high-accuracy model specifically for invoices that requires zero manual labeling.

How to eliminate wrong answers

Option A is wrong because the Read API only extracts raw text without understanding document structure or field semantics, requiring complex and brittle regex patterns that fail with varied invoice layouts. Option C is wrong because Custom Vision object detection is designed for image classification and object localization, not for extracting structured text fields from documents. Option D is wrong because labeling hundreds of invoices to train a custom model is unnecessary and inefficient when a prebuilt model already exists for this specific use case, violating the requirement for minimal manual labeling.

Full explanation →
35
Multi-Selecthard

Which THREE factors should you consider when choosing between Azure AI Custom Vision and Azure AI Vision pre-built models for an image classification task? (Choose three.)

Select 3 answers
A.Availability of labeled training data specific to the domain
B.Image format support (JPEG, PNG)
C.Whether the required labels are covered by the pre-built model
D.Need for real-time inference latency
E.Need to iterate and retrain the model over time
AnswersA, C, E

Custom Vision requires labeled data; pre-built models do not.

Why this answer

Azure AI Custom Vision is specifically designed for scenarios where you have domain-specific labeled training data that is not covered by pre-built models. Custom Vision allows you to upload your own labeled images and train a custom model tailored to your unique classification needs, which is essential when off-the-shelf models fail to recognize your target classes.

Exam trap

The trap here is that candidates often confuse image format support or latency as key differentiators, when in fact both services handle these similarly, and the core distinction lies in the availability of custom labeled data and the need for iterative retraining.

Full explanation →
36
MCQhard

You see the exhibit representing an Azure Bot resource configuration. The bot is not responding to user messages. What should you verify first?

A.Check that the endpoint URL is correct and the web app is running
B.Ensure that LUIS app IDs are provided
C.Verify that the Application Insights key is valid
D.Confirm that the msaAppId is a valid GUID
AnswerA

A bot that receives messages but never replies usually has a broken messaging endpoint. Verifying the endpoint URL and confirming the hosting web app is running tests the actual message delivery path before investigating channels, credentials, or configuration.

Why this answer

The most common reason a bot fails to respond to user messages is that the endpoint URL configured in the Azure Bot resource does not match the actual URL of the running web app, or the web app itself is stopped or unresponsive. Without a correct and reachable endpoint, the Bot Framework Service cannot forward messages to the bot's message handler, causing a complete communication breakdown.

Exam trap

The trap here is that candidates often jump to authentication or AI service configuration issues, but the most immediate and common cause of a non-responsive bot is a simple connectivity or endpoint misconfiguration.

How to eliminate wrong answers

Option B is wrong because LUIS app IDs are only required if the bot uses Language Understanding for intent detection; they are not necessary for basic message handling. Option C is wrong because Application Insights is used for telemetry and monitoring, not for core message routing; an invalid key would not prevent the bot from responding. Option D is wrong because the msaAppId (Microsoft App ID) is used for authentication with the Bot Framework Service, but if it were invalid, the bot would typically fail to register or authenticate, not silently ignore messages; the endpoint and web app availability are more fundamental.

Full explanation →
37
MCQmedium

You are a lead AI engineer for a global retail company. The company is building an AI-powered customer support chatbot using Microsoft Foundry. The chatbot must answer product questions, process returns, and escalate to human agents when needed. The solution uses Azure AI Language for intent recognition and Azure AI Bot Service for bot orchestration. During testing, the chatbot fails to understand customer queries about return policies. The intents 'ProductInquiry' and 'ReturnRequest' are defined, but the model often confuses them. You need to improve intent classification accuracy. The development team has already collected 500 sample utterances for each intent. You have a budget to collect additional data. What should you do?

A.Use Azure AI Translator to translate utterances into multiple languages
B.Collect additional utterances that are similar to the confusing ones and retrain
C.Increase the confidence threshold in the bot configuration
D.Create a new custom language project and retrain from scratch
AnswerB

Confusion between ProductInquiry and ReturnRequest stems from insufficient utterance coverage near the decision boundary. Adding utterances similar to the misclassified examples gives the model discriminative training signal on that boundary, improving classification accuracy more effectively than unrelated data.

Why this answer

Collecting additional utterances that are similar to the confusing ones directly addresses the ambiguity between the 'ProductInquiry' and 'ReturnRequest' intents. By providing more representative examples of edge cases where the intents overlap, the Azure AI Language model can better learn the subtle linguistic patterns that distinguish them, thereby improving classification accuracy without requiring a complete retraining from scratch.

Exam trap

The trap here is that candidates often assume increasing the confidence threshold (Option C) will fix misclassifications, but this only adjusts the prediction cutoff and does not improve the underlying model's ability to differentiate intents, which is a data quality issue, not a threshold tuning issue.

How to eliminate wrong answers

Option A is wrong because translating utterances into multiple languages does not resolve confusion between two intents in the same language; it only adds multilingual support, which is irrelevant to the core issue of intent ambiguity. Option C is wrong because increasing the confidence threshold would only reject more low-confidence predictions, not improve the model's ability to distinguish between similar intents; it could actually increase false negatives by requiring higher confidence for correct classifications. Option D is wrong because creating a new custom language project and retraining from scratch is unnecessary and wasteful; the existing project already has 500 utterances per intent, and the problem is specifically about confusing utterances, not a fundamental flaw in the project setup.

Full explanation →
38
MCQhard

You are developing a solution that uses Azure AI Video Indexer to analyze surveillance videos for suspicious activity. The solution must generate alerts when a person is detected in a restricted area. Which feature should you use?

A.Azure AI Video Indexer sentiment analysis
B.Azure AI Video Indexer people detection and tracking
C.Azure AI Face identify API
D.Object detection in Azure AI Vision
AnswerB

People detection and tracking identifies and follows individuals across video frames, enabling detection of a person entering a defined restricted area. This satisfies the requirement to raise alerts on unauthorised presence, unlike face or emotion features.

Why this answer

Azure AI Video Indexer's people detection and tracking feature is specifically designed to detect and track individuals across video frames, making it ideal for generating alerts when a person enters a restricted area. This feature provides bounding boxes, timestamps, and tracking IDs that enable real-time monitoring and alerting based on spatial rules.

Exam trap

The trap here is that candidates confuse generic object detection (which can detect people as objects) with the dedicated people tracking capability that provides persistent IDs and temporal continuity needed for restricted-area alerts.

How to eliminate wrong answers

Option A is wrong because sentiment analysis in Azure AI Video Indexer detects emotional tone (e.g., happiness, sadness) from audio or text, not physical presence or location of people. Option C is wrong because the Azure AI Face Identify API matches detected faces against a known person group (identification), but does not track movement or detect entry into restricted zones. Option D is wrong because object detection in Azure AI Vision identifies generic objects (e.g., car, dog) and does not specialize in people tracking or spatial alerting within video streams.

Full explanation →
39
MCQeasy

You are using Azure OpenAI Service to generate product descriptions. The output is often too verbose. You need to reduce the length of generated text without changing the model. Which parameter should you adjust?

A.Max tokens
B.Frequency penalty
C.Temperature
D.Top-p (nucleus sampling)
AnswerA

Max tokens caps the total tokens the model may generate, directly truncating verbose completions without altering the deployed model. Setting a lower value constrains output length, satisfying the requirement to shorten product descriptions while leaving the model itself unchanged.

Why this answer

Max tokens controls the total length of the generated response by capping the number of tokens (words/subwords) the model can output. Reducing this value directly truncates the output, making descriptions shorter without altering the model or its behavior. Other parameters influence randomness or repetition but do not enforce a strict length limit.

Exam trap

The trap here is that candidates confuse parameters that affect output style (temperature, top-p, frequency penalty) with the one parameter that directly controls output length (max tokens), leading them to choose a parameter that changes how the model writes rather than how much it writes.

How to eliminate wrong answers

Option B (Frequency penalty) is wrong because it reduces the likelihood of repeating the same tokens or phrases, which can change content style but does not enforce a maximum output length. Option C (Temperature) is wrong because it controls the randomness of token selection (higher = more creative, lower = more deterministic), not the number of tokens generated. Option D (Top-p) is wrong because it limits the cumulative probability of token choices (nucleus sampling), affecting diversity but not the total token count.

Full explanation →
40
MCQmedium

Your company has a large repository of scanned invoices in PDF format. You need to extract invoice number, date, total amount, and vendor name from these PDFs. Which Azure AI service should you use?

A.Azure AI Language
B.Azure AI Document Intelligence
C.Azure AI Search
D.Azure AI Vision
AnswerB

Document Intelligence applies prebuilt invoice models that extract invoice number, date, total amount, and vendor name from scanned PDFs, handling OCR and layout. It directly satisfies the requirement to pull those specific fields from image-based invoices.

Why this answer

Azure AI Document Intelligence (formerly Form Recognizer) is purpose-built for extracting structured data (fields like invoice number, date, total amount, vendor name) from scanned documents such as invoices and receipts. It uses prebuilt models trained specifically on invoice layouts and supports both OCR and key-value pair extraction from PDFs.

Exam trap

A common mistake is to choose Azure AI Vision because it performs OCR on scanned PDFs, but it does not provide prebuilt models for extracting structured fields like invoice data. Azure AI Document Intelligence (formerly Form Recognizer) is the correct service as it specializes in extracting key-value pairs from forms and documents.

How to eliminate wrong answers

Option A is wrong because Azure AI Language is designed for text analytics (sentiment, key phrases, entity recognition) and does not natively extract structured fields from scanned PDF documents. Option C is wrong because Azure AI Search is a search indexing and query service, not a document data extraction tool; it requires pre-extracted content to index. Option D is wrong because Azure AI Vision provides general OCR and image analysis capabilities but lacks the specialized prebuilt invoice models and field extraction logic that Document Intelligence offers.

Full explanation →
41
MCQhard

Your team is developing a chatbot using Azure AI Bot Service with Language Understanding (LUIS). You need to improve intent recognition for user queries that contain typos. What should you recommend?

A.Add synonym lists for common misspellings
B.Use Azure Cognitive Search to correct typos
C.Enable Bing Spell Check in the LUIS app settings
D.Train the model with many examples containing typos
AnswerC

Enabling Bing Spell Check in the LUIS app settings corrects typographical errors in incoming utterances before intent classification, directly satisfying the stem's requirement to improve recognition of queries containing typos. LUIS then evaluates the corrected text against its trained intents, raising confidence scores that misspellings would otherwise depress.

Why this answer

Enabling Bing Spell Check in the LUIS app settings allows the service to automatically correct typos in user utterances before they are processed for intent recognition. This built-in feature integrates directly with LUIS at the endpoint, correcting misspelled words to improve the accuracy of intent and entity extraction without requiring manual data augmentation or external services.

Exam trap

The trap here is that candidates often assume training with typos (Option D) is the best way to handle misspellings, but Microsoft explicitly recommends using Bing Spell Check as a separate preprocessing layer rather than bloating the training data with error-prone examples.

How to eliminate wrong answers

Option A is wrong because synonym lists in LUIS are used to map different words that have the same meaning (e.g., 'car' and 'automobile'), not to handle typos or misspellings; they do not correct character-level errors. Option B is wrong because Azure Cognitive Search is a search indexing and query service, not a spell-checking tool; it cannot be used to correct typos in real-time LUIS utterances. Option D is wrong because while training with many examples containing typos might help the model learn some patterns, it is inefficient and does not generalize well to unseen typos; LUIS's built-in Bing Spell Check is the recommended and more robust approach.

Full explanation →
42
MCQeasy

A developer wants to use Azure OpenAI to generate text from a prompt. Which parameter controls the diversity of the generated output?

A.presence_penalty
B.frequency_penalty
C.temperature
D.max_tokens
AnswerC

Temperature scales the sampling distribution's randomness: near zero the model picks the highest-probability token almost deterministically, while higher values flatten probabilities and increase diversity. It is the parameter controlling output variety, not frequency penalty or top_p alone.

Why this answer

Temperature is the parameter that directly controls the randomness or diversity of the generated output by scaling the logits before applying the softmax function. A higher temperature (e.g., 1.0) increases the probability of less likely tokens, producing more creative and varied responses, while a lower temperature (e.g., 0.1) makes the output more deterministic and focused.

Exam trap

The trap here is that candidates often confuse frequency_penalty or presence_penalty with controlling diversity, but those parameters address repetition and topic novelty, not the fundamental randomness of token selection, which is exclusively governed by temperature.

How to eliminate wrong answers

Option A is wrong because presence_penalty penalizes tokens that have already appeared in the text so far, encouraging the model to introduce new topics or avoid repetition, but it does not directly control the overall diversity or randomness of the token selection. Option B is wrong because frequency_penalty reduces the likelihood of tokens based on how frequently they have occurred in the generated text, which helps prevent repetitive phrasing but does not adjust the probability distribution's entropy like temperature does. Option D is wrong because max_tokens simply limits the maximum number of tokens in the generated response and has no effect on the diversity or randomness of the output.

Full explanation →
43
Multi-Selectmedium

Which THREE are valid parameters when calling the Azure OpenAI Service chat completions API?

Select 3 answers
A.messages
B.index_name
C.temperature
D.max_tokens
E.embedding_model
AnswersA, C, D

Required input.

Why this answer

The `messages` parameter is required in the Azure OpenAI Service chat completions API call. It defines the conversation history and user input as an array of message objects, each with a `role` (system, user, assistant) and `content`. Without this parameter, the API cannot determine the context or prompt for generating a response.

Exam trap

Microsoft often tests the distinction between parameters for different Azure OpenAI API endpoints (chat completions vs. embeddings vs. search), so candidates may confuse `index_name` or `embedding_model` as valid chat completions parameters due to their familiarity with other Azure AI services.

Full explanation →
44
MCQhard

A financial services company uses Azure AI Language's custom text classification to categorize loan applications as 'Approved', 'Denied', or 'Review Required'. The model is trained on historical data but is producing poor accuracy on new applications. The data scientist suspects data leakage between training and test sets. What should the data scientist do to validate this?

A.Increase the training dataset size and retrain the model.
B.Use k-fold cross-validation during training.
C.Adjust the classification confidence threshold.
D.Split the data chronologically and ensure no overlapping data between train and test sets.
AnswerD

Chronological splitting prevents future loan applications leaking into training, which inflates test accuracy and masks poor generalisation. Random splits on time-ordered financial data let later patterns contaminate training, so temporal separation is the correct validation for suspected leakage between train and test sets.

Why this answer

Data leakage occurs when information from outside the training set inadvertently influences the model, often due to overlapping or non-independent data splits. By splitting the data chronologically (e.g., training on older applications and testing on newer ones), the data scientist ensures that no future information leaks into the training process, which directly validates whether temporal leakage is causing poor accuracy. This approach is standard for time-series or sequential data like loan applications, where patterns may shift over time.

Exam trap

The trap here is that candidates often confuse data leakage with model performance issues and choose to increase data or adjust thresholds, not realizing that the core problem is the integrity of the train-test split, which must be validated through chronological separation.

How to eliminate wrong answers

Option A is wrong because simply increasing the training dataset size does not address data leakage; if the leakage exists, more data will only reinforce the spurious correlations. Option B is wrong because k-fold cross-validation randomly shuffles data, which can actually mask or even exacerbate leakage by mixing future and past samples across folds, making it unsuitable for detecting temporal leakage. Option C is wrong because adjusting the classification confidence threshold only changes the decision boundary for predictions, not the underlying data split or leakage issue, so it cannot validate whether leakage exists.

Full explanation →
45
Multi-Selecthard

You are designing an Azure AI solution that uses an Azure AI services multi-service account. The solution must call the service from an Azure Kubernetes Service (AKS) cluster without embedding account keys in application code. You need to configure authentication and authorization so that the workload identity used by the pods can be granted access. (Choose two.)

Select 2 answers
A.Store the account key in an Azure Key Vault secret and reference it from the pod.
B.Assign the Cognitive Services User role to the managed identity.
C.Create a service principal with a client secret and mount the secret into the pod.
D.Configure the AKS cluster to use workload identity federation with Microsoft Entra ID.
E.Enable local authentication on the Azure AI services account.
AnswersB, D

Granting the Cognitive Services User role to the managed identity gives the workload the data-plane permissions needed to call the Azure AI services endpoints. This is the recommended way to authorize API calls without keys, because role assignments are evaluated by Azure RBAC and can be scoped to the specific Azure AI services resource. It directly supports the requirement to avoid embedding account keys in application code.

Why this answer

Keyless access from AKS pods requires two things: the cluster must federate the pod identity with Microsoft Entra ID, and that identity must be granted a role on the Azure AI services resource. Workload identity federation provides the token acquisition mechanism, while the Cognitive Services User role provides the authorization. Together they let the application call the service without storing or handling account keys.

Exam trap

The trap here is thinking that storing the key in Key Vault satisfies a no-keys-in-code requirement, when the application still has to retrieve and use that key.

Full explanation →
46
MCQhard

A healthcare company uses Azure AI Language's custom question answering to build a bot that answers patient FAQs. The knowledge base contains documents with sensitive information. The company needs to ensure that the bot only returns answers from documents that the user is authorized to access. What should they implement?

A.Configure the bot to prompt users for credentials before answering and validate against Azure Active Directory.
B.Enable role-based access control (RBAC) on the Azure AI Language resource.
C.Implement document-level access control by using metadata tags and filtering in the query.
D.Use separate knowledge bases for each user group and route queries based on user identity.
AnswerC

Azure AI Language custom question answering supports metadata on documents, which can be used to tag documents with access levels or user groups. At query time, you can filter results based on metadata, ensuring users only get answers from documents they are authorized to see. This is the recommended approach for document-level security.

Why this answer

The correct solution is to use metadata tags on documents and apply filters during query execution. This allows the custom question answering service to return only answers from documents that match the user's authorization level. Other options either address resource management, require complex duplication, or only authenticate without enforcing access control.

Exam trap

The trap here is confusing authentication with authorization; authenticating a user does not automatically restrict which documents they can retrieve answers from.

Full explanation →
47
MCQmedium

A company uses Azure Document Intelligence to process purchase orders. They have trained a custom model with 10 labeled samples and deployed it as 'purchaseOrderModel'. When analyzing a new purchase order, the extracted 'TotalAmount' field is often incorrect. The company wants to improve the model's accuracy for this field. What should they do?

A.Retrain the model with additional labeled samples that include variations of the 'TotalAmount' field.
B.Switch to the prebuilt invoice model, which automatically extracts total amounts from purchase orders.
C.Increase the model's confidence threshold for the 'TotalAmount' field in the project settings.
D.Add a labeled sample where the 'TotalAmount' field is left blank to teach the model to ignore missing values.
AnswerA

This is correct because custom model accuracy improves with more diverse labeled data. Adding samples that cover different formats, locations, and contexts of the TotalAmount field helps the model generalize better. Azure Document Intelligence learns from labeled examples, so increasing the quantity and variety of training data directly addresses the field's extraction accuracy.

Why this answer

Custom model accuracy in Azure Document Intelligence depends heavily on the quality and quantity of labeled training data. When a specific field like TotalAmount is frequently misidentified, the most effective action is to retrain with more labeled examples that capture the field's variations. This helps the model learn the patterns and contexts associated with that field, leading to better generalization on new documents.

Exam trap

The trap here is assuming that confidence thresholds or prebuilt models can be tweaked to improve a custom model's field accuracy, when the real fix is more and better training data.

Full explanation →
48
Multi-Selecthard

You are implementing a generative AI solution using Azure OpenAI Service. The solution must generate responses that are grounded in your organization's proprietary documents and must return citations that link back to the source documents. You need to configure the deployment to meet these requirements. Which two actions should you perform? (Choose two.)

Select 2 answers
A.Create an Azure AI Search index that contains the proprietary documents with retrievable content and citation fields such as title and URL.
B.Enable diagnostic logging on the Azure OpenAI resource to capture prompt and completion text.
C.Fine-tune the model on the proprietary documents so it can reproduce them verbatim.
D.Associate the Azure AI Search index as a data source on the model deployment by using the Azure OpenAI On Your Data configuration.
E.Deploy a second model instance in a different Azure region for high availability.
AnswersA, D

Grounding and citations require a searchable index whose documents expose content plus metadata fields like title and URL. Azure AI Search provides the retrieval layer that the Azure OpenAI On Your Data feature queries. Without an index containing retrievable content and citation fields, the service cannot retrieve relevant chunks or produce source links, so this action is required.

Why this answer

Grounded responses with citations in Azure OpenAI require a retrieval source and a deployment-level data source binding. An Azure AI Search index holds the proprietary documents with content and citation metadata, and associating that index with the model deployment through On Your Data enables service-side retrieval, prompt injection, and citation generation.

Exam trap

The trap here is assuming fine-tuning can supply both grounding and citations, when citations depend on a search index and a data source binding rather than on model training.

Full explanation →
49
Multi-Selectmedium

You are deploying a knowledge mining solution using Azure AI Search and Azure AI Document Intelligence. The solution must extract text from scanned documents, identify named entities, and index the content. You need to configure the skillset. Which TWO built-in skills should you include in the skillset?

Select 2 answers
A.Merge skill
B.LanguageDetection skill
C.OCR skill
D.EntityRecognition skill
E.KeyPhraseExtraction skill
AnswersC, D

The OCR skill performs optical character recognition on scanned document images, producing text that downstream skills can consume. It satisfies the constraint to extract text from scanned documents, since those files contain no embedded text layer for the indexer to read directly.

Why this answer

The OCR skill (option C) is correct because it is the built-in Azure AI Search cognitive skill that invokes Azure AI Document Intelligence's Read/OCR capability to extract text from scanned documents and images, which is exactly what the scenario requires for scanned input. The EntityRecognition skill (option D) is correct because it is the built-in skill that calls the Text Analytics entity recognition model to identify named entities (people, organizations, locations, dates, and so on) in the extracted text, satisfying the named-entity requirement. The Merge skill (option A) only combines text and offsets from multiple input fields and does not perform OCR or entity extraction, so it is not required here.

The LanguageDetection skill (option B) merely detects the language of the text and is not needed to extract text or identify entities. The KeyPhraseExtraction skill (option E) extracts key phrases rather than named entities, so it does not meet the entity-identification requirement.

Full explanation →
50
Multi-Selecthard

You are using Microsoft Purview to create a knowledge map of your organization's data assets. The solution must automatically scan and classify sensitive data in Azure Blob Storage. You need to configure the scanning and classification. Which THREE actions should you perform?

Select 3 answers
A.Run a full scan of the Blob Storage to discover and classify data.
B.Create custom classification rules for sensitive data types.
C.Apply sensitivity labels to the classified data.
D.Create a scan rule set that includes the desired classification rules.
E.Register the Azure Blob Storage account as a data source in Purview.
AnswersA, D, E

A full scan is what actually triggers Purview to crawl the registered Blob Storage, apply the chosen scan rule set, and populate the knowledge map with discovered assets and classifications. Without running the scan, no classification occurs, so this action satisfies the requirement for automatic discovery and classification.

Why this answer

Option E is correct because before Microsoft Purview can scan any asset, the Azure Blob Storage account must first be registered as a data source in the Purview governance portal, which establishes the connection and allows the account to be managed and scanned. Option D is correct because a scan rule set defines which file types and classification rules are applied during a scan; you must create or select a rule set that includes the desired system or custom classification rules so sensitive data types are detected. Option A is correct because after registering the source and configuring the rule set, you must run a scan (a full scan for initial discovery and classification) so Purview can crawl the Blob Storage, apply the rule set, and populate the knowledge map with classified assets.

Option B is not required because Purview provides built-in system classification rules for common sensitive data types, and custom rules are only needed for organization-specific patterns, which the scenario does not require. Option C is not part of the scanning and classification configuration; sensitivity labels are applied through Microsoft Purview Information Protection and are a separate labeling concern from the data map scanning process.

Exam trap

The trap here is that candidates often confuse the required actions for configuring scanning (registering the source, creating a scan rule set, and running a scan) with optional or subsequent steps like creating custom rules or applying sensitivity labels, leading them to select B or C instead of the correct three.

Full explanation →
51
MCQhard

You are implementing a knowledge mining solution with Azure AI Search that ingests data from Azure Blob Storage. The pipeline includes a custom skill that calls an external API for specialized entity extraction. The custom skill sometimes returns HTTP 429 (Too Many Requests). How should you handle this to ensure reliable indexing?

A.Reduce the batch size in the indexer
B.Increase the skill timeout
C.Configure a retry policy on the custom skill
D.Schedule the indexer to run less frequently
AnswerC

A retry policy on the custom skill handles transient HTTP 429 responses by reissuing the request after a delay, honouring Retry-After where supplied. This keeps indexing reliable without discarding enriched documents, directly addressing the throttling constraint described in the stem.

Why this answer

Configuring a retry policy on the custom skill is the correct way to handle HTTP 429 errors, as it allows the skill to retry the request after a delay, respecting the Retry-After header if provided. This ensures reliable indexing without overwhelming the external API.

Exam trap

AI-102 often tests the misconception that increasing timeout or reducing batch size solves rate limiting, when the correct approach is to implement a retry policy that respects the API's rate limits.

How to eliminate wrong answers

Option A is wrong because reducing batch size may lessen load but does not directly handle 429 errors; it's a mitigation, not a solution. Option B is wrong because increasing the skill timeout does not address rate limiting; the skill would still fail if the API returns 429. Option D is wrong because scheduling the indexer less frequently does not handle transient 429 errors during a run; it only reduces frequency, not the error handling.

Full explanation →
52
MCQmedium

A retail company uses Azure Computer Vision to analyze customer traffic in stores. They process images from security cameras using the OCR API to detect product labels. Recently, the OCR accuracy has decreased for images with poor lighting. Which pre-processing step should the company implement to improve OCR accuracy?

A.Convert images to grayscale before sending to OCR API.
B.Increase the image resolution before calling OCR API.
C.Adjust brightness and contrast of images using image processing.
D.Reduce image size to decrease noise.
AnswerC

Adjusting brightness and contrast directly compensates for the poor lighting that is degrading OCR accuracy. Microsoft Entra ID is irrelevant here; the constraint is image quality, not authentication. Normalising luminance and contrast before sending images to the OCR API improves character segmentation and recognition, satisfying the stem's requirement to raise accuracy for poorly lit camera feeds.

Why this answer

Poor lighting directly reduces the contrast between text and background, which is critical for OCR accuracy. Adjusting brightness and contrast improves the signal-to-noise ratio of the text regions, making character edges more distinct for the Azure Computer Vision OCR engine. This pre-processing step compensates for the lighting deficiency without altering the fundamental image content that the API relies on.

Exam trap

The trap here is that candidates confuse image quality improvements (like resolution or noise reduction) with the specific need to correct lighting-induced contrast loss, which is a distinct pre-processing requirement for OCR in poor illumination.

How to eliminate wrong answers

Option A is wrong because converting to grayscale removes color information that can help distinguish text from similarly-toned backgrounds, and it does not address the root cause of low contrast due to poor lighting. Option B is wrong because increasing resolution does not fix the underlying contrast problem; it may even amplify noise and increase API processing time without improving text legibility. Option D is wrong because reducing image size discards pixel detail, which can make small or thin text characters unreadable for OCR, and it does not mitigate the effects of poor lighting.

Full explanation →
53
Multi-Selecthard

An agent uses Azure OpenAI with function calling to perform actions. The agent is not executing functions correctly. Which THREE factors should the team check to diagnose the issue?

Select 3 answers
A.The temperature parameter is set too high.
B.The token limit is too low, truncating the function definitions.
C.The function parameter schemas are incorrect or incomplete.
D.The function descriptions are ambiguous or missing.
E.The model version is outdated.
AnswersB, C, D

Function definitions consume prompt tokens alongside the conversation. If the model's context window or max token setting is too low, definitions get truncated, so the model receives incomplete schemas and cannot emit valid function call arguments, causing execution failures.

Why this answer

Option B is correct because function definitions are serialized into the prompt sent to the model, so if the token limit (max_tokens/context window) is too low, the function schema can be truncated and the model cannot emit valid function calls. Option C is correct because Azure OpenAI function calling relies on a strict JSON Schema for each function's parameters; incorrect or incomplete schemas (wrong types, missing required fields) cause the model to produce arguments that fail validation or the call to be rejected. Option D is correct because the model selects functions based on their natural-language descriptions, so ambiguous or missing descriptions lead to wrong or no function selection.

Option A is not a primary cause: temperature affects randomness, not whether a syntactically valid function call is produced, and function calling can work at high temperature. Option E is not a required check: while newer models may improve function-calling reliability, an outdated model version is not a standard diagnostic factor for functions failing to execute.

Exam trap

Microsoft often tests the misconception that temperature or model version are primary causes for function-calling failures, when in reality the core issues are token limits, schema correctness, and description clarity.

Full explanation →
54
MCQmedium

You are deploying a question answering solution using Azure AI Language. The solution must be able to provide answers from a set of frequently asked questions (FAQs) in PDF format. What should you do?

A.Use Azure AI Search with cognitive skills.
B.Create a custom question answering project and add the PDF as a source.
C.Use Azure OpenAI with a system prompt containing the PDFs.
D.Use the pre-built question answering in Azure AI Language.
AnswerB

Adding the PDF as a source in a custom question answering project lets Azure AI Language extract question-and-answer pairs directly from the document, satisfying the FAQ-from-PDF requirement. Unlike unstructured text ingestion, this source type parses the PDF's structure into a searchable knowledge base without manual pair creation.

Why this answer

Azure AI Language's custom question answering feature allows you to directly upload a PDF as a knowledge source. The service automatically extracts Q&A pairs from the document, enabling the solution to answer questions based on the FAQ content without needing additional search or cognitive skill pipelines.

Exam trap

The trap here is that candidates often confuse the pre-built question answering (which is a generic, non-customizable service) with the custom question answering project, leading them to choose option D, or they overcomplicate the solution by selecting Azure AI Search with cognitive skills when a direct PDF ingestion capability exists.

How to eliminate wrong answers

Option A is wrong because Azure AI Search with cognitive skills is designed for indexing and enriching unstructured data with AI capabilities, but it does not natively extract Q&A pairs from PDFs or provide a direct question-answering interface; it would require building a custom Q&A pipeline on top of the search index. Option C is wrong because Azure OpenAI with a system prompt containing the PDFs would require manual prompt engineering and does not automatically parse or structure the FAQ content into a queryable knowledge base; it also incurs higher latency and cost for each query. Option D is wrong because the pre-built question answering in Azure AI Language is a general-purpose service that does not support custom sources like PDFs; it only works with predefined, built-in knowledge bases and cannot ingest your specific FAQ document.

Full explanation →
55
Drag & Dropmedium

Drag and drop the steps to set up Azure AI Content Safety for content moderation into the correct order.

Drag or tap steps into the slots.

Steps
Order
1Step 1
2Step 2
3Step 3
4Step 4

Why this order

The correct order for setting up Azure AI Content Safety is: first create the Content Safety resource, then obtain its endpoint and key, call the Content Safety API with the content, analyze the response for moderation results, and finally act on the analysis (e.g., block or flag content). This sequence ensures proper authentication and logical flow from setup to action.

Full explanation →
56
MCQmedium

A company uses Microsoft Copilot for Microsoft 365 to automate email responses. They want to ensure that the Copilot responses comply with their data governance policies and do not expose sensitive information. What should they configure?

A.Microsoft Entra ID conditional access policies
B.Microsoft Purview Data Loss Prevention (DLP) policies
C.Microsoft Intune app protection policies
D.Microsoft Sentinel analytics rules
AnswerB

Microsoft Purview DLP policies inspect Copilot-generated content in Exchange Online and SharePoint, blocking responses that contain sensitive information types or sensitivity labels defined in your governance rules. This directly enforces the compliance constraint by preventing exposure at generation time, rather than relying on post-hoc auditing or user discretion.

Why this answer

Microsoft Purview Data Loss Prevention (DLP) policies are designed to detect and prevent the accidental sharing of sensitive information, such as credit card numbers or personally identifiable information (PII), across Microsoft 365 services. By configuring DLP policies, the company can scan Copilot-generated email responses for sensitive data patterns and block or warn before the email is sent, ensuring compliance with data governance policies.

Exam trap

The trap here is that candidates often confuse data governance and content inspection with access control (Entra ID) or endpoint security (Intune), leading them to select a policy that governs who can access data rather than what data is allowed to be shared.

How to eliminate wrong answers

Option A is wrong because Microsoft Entra ID conditional access policies control authentication and access to applications based on user, device, or location conditions, but they do not inspect or govern the content of Copilot-generated email responses for sensitive data. Option C is wrong because Microsoft Intune app protection policies manage how data is handled within mobile apps (e.g., copy/paste restrictions, encryption at rest), but they do not scan or enforce content-level rules on email responses generated by Copilot. Option D is wrong because Microsoft Sentinel analytics rules are used for security information and event management (SIEM) to detect threats and anomalies in logs, not to enforce data governance or prevent sensitive data exposure in email content.

Full explanation →
57
MCQmedium

You are planning a new Azure AI solution that will use several Azure AI services, including Azure AI Vision and Azure AI Language. The solution must allow developers to authenticate by using their Microsoft Entra ID credentials without storing service keys in code. You need to configure the Azure AI services resource to support this authentication method. What should you do?

A.Enable local authentication on the Azure AI services resource and distribute the keys to developers.
B.Store the Azure AI services keys in Azure Key Vault and provide developers access to the vault.
C.Create a managed identity for the Azure AI services resource and assign it the Cognitive Services User role.
D.Configure the Azure AI services resource to disable local authentication and grant developers the Cognitive Services User role on the resource.
AnswerD

Disabling local authentication forces authentication through Microsoft Entra ID. Granting developers the Cognitive Services User role allows them to use their Entra ID credentials to obtain tokens for accessing the service. This achieves keyless authentication and adheres to the requirement.

Why this answer

To allow developers to authenticate with Microsoft Entra ID, you must disable key-based authentication on the Azure AI services resource and assign the appropriate role, such as Cognitive Services User, to the developers. This ensures that access is granted through Entra ID tokens rather than shared keys, aligning with security best practices.

Exam trap

The trap here is assuming that a managed identity assigned to the Azure AI services resource can be used by developers to authenticate, overlooking that managed identities are for resource-to-resource authentication, not user access.

Full explanation →
58
MCQhard

You are developing a generative AI solution using Azure OpenAI Service. The solution must generate product descriptions based on a set of attributes provided by the user. You need to ensure the output adheres to a specific JSON schema. Which feature should you use to enforce the structure of the model's response?

A.Function calling with a defined function that returns JSON.
B.Temperature set to 0 to make output deterministic.
C.Prompt engineering with few-shot examples of JSON.
D.Structured outputs with a JSON schema definition.
AnswerD

Structured outputs allow you to define a JSON schema that the model's response must follow. This ensures the generated content is valid JSON and matches the specified structure, which is ideal for generating product descriptions with consistent fields. It enforces the schema at the API level, reducing the need for post-processing.

Why this answer

Structured outputs in Azure OpenAI Service enable you to specify a JSON schema that the model's response must conform to. This guarantees that the generated output is valid JSON and includes all required fields, which is essential for generating product descriptions with a consistent format. It is a robust way to enforce structure without relying on prompt engineering alone.

Exam trap

The trap here is assuming that function calling or prompt engineering can guarantee a JSON schema, when only structured outputs provide that strict enforcement.

Full explanation →
59
MCQmedium

You are deploying an Azure AI Language resource that will process customer feedback. The resource must be accessible only from a specific Azure virtual network and must allow access from an on-premises network via a site-to-site VPN. You need to configure network security for the resource. What should you do?

A.Use Azure Private Link to create a private endpoint for the Azure AI Language resource and enable public access for on-premises clients.
B.Configure a private endpoint for the Azure AI Language resource and disable public access.
C.Enable public access and configure IP firewall rules to allow only the on-premises public IP address.
D.Configure a service endpoint for Azure AI Language on the virtual network subnet and enable public access for on-premises clients.
AnswerB

A private endpoint creates a private IP address for the Azure AI Language resource within your virtual network, enabling secure access from the VNet and on-premises via VPN. Disabling public access ensures no exposure to the internet. This meets the requirement for restricted access and integrates with your existing network infrastructure.

Why this answer

A private endpoint is the correct solution because it assigns a private IP address from your virtual network to the Azure AI Language resource, allowing secure access from both the VNet and on-premises via VPN. Disabling public access ensures the resource is not reachable from the internet. This configuration satisfies the network isolation requirements.

Exam trap

The trap here is confusing service endpoints with private endpoints; service endpoints do not provide private IP connectivity for on-premises clients and do not disable public access.

Full explanation →
60
MCQmedium

You are designing a knowledge mining solution that ingests documents from SharePoint Online and makes them searchable using Azure AI Search. The solution must extract text from images and perform optical character recognition (OCR) on embedded images within PDFs. Which built-in skill should you include in the skillset?

A.OCR skill
B.Translation skill
C.Key phrase extraction skill
D.Entity recognition skill
AnswerA

The OCR skill extracts text from embedded images and scanned PDF content within the enrichment pipeline, feeding recognised text downstream for indexing. It directly satisfies the requirement to perform optical character recognition on images embedded in PDFs from SharePoint Online.

Why this answer

The OCR skill (Optical Character Recognition) is the correct built-in skill for extracting text from images and performing OCR on embedded images within PDFs in Azure AI Search. It specifically handles image files (e.g., JPEG, PNG) and embedded images in PDFs, outputting text that can be indexed and searched. Other skills like translation, key phrase extraction, or entity recognition do not perform text extraction from images.

Exam trap

The trap here is that candidates may confuse the OCR skill with other text-processing skills like key phrase extraction or entity recognition, mistakenly thinking those can also extract text from images, but only the OCR skill is designed for image-to-text conversion.

How to eliminate wrong answers

Option B is wrong because the Translation skill translates text from one language to another, but it cannot extract text from images or perform OCR on embedded images. Option C is wrong because the Key Phrase Extraction skill identifies key phrases from existing text, but it does not extract text from images or perform OCR. Option D is wrong because the Entity Recognition skill identifies entities (e.g., people, organizations) from text, but it cannot extract text from images or perform OCR on embedded images.

Full explanation →
61
MCQhard

You are responsible for managing costs for multiple Azure AI services in your organization. You notice that provisioned throughput units (PTUs) for Azure OpenAI are not fully utilized. What is the most cost-effective action to optimize spending?

A.Configure auto-scaling to reduce PTUs during low usage
B.Move the resource to a different region with lower pricing
C.Stop the Azure OpenAI service when not in use
D.Reduce the number of provisioned PTUs or switch to pay-as-you-go
AnswerD

Underutilised provisioned throughput units incur cost regardless of consumption, so reducing the PTU count or reverting to pay-as-you-go token billing eliminates spend on idle capacity. This directly addresses the stem's unused-PTU constraint, matching cost to actual demand.

Why this answer

Provisioned throughput units (PTUs) represent a fixed capacity commitment. If PTUs are underutilized, the most cost-effective action is to reduce the number of PTUs or switch to the pay-as-you-go (PAYG) model, which charges only for tokens consumed. This directly aligns with cost optimization by eliminating the fixed cost of unused capacity.

Exam trap

The trap here is that candidates confuse PTUs with standard Azure auto-scaling concepts (like VM scale sets) and assume auto-scaling is available for PTUs, when in fact PTUs are a fixed-capacity model that requires manual adjustment or a switch to PAYG for cost optimization.

How to eliminate wrong answers

Option A is wrong because Azure OpenAI does not support auto-scaling for PTUs; PTUs are a fixed, pre-provisioned capacity model, and scaling must be done manually via quota adjustments. Option B is wrong because moving to a different region does not address underutilization of PTUs; regional pricing differences are typically minimal and do not solve the core issue of paying for unused capacity. Option C is wrong because stopping the Azure OpenAI service (e.g., via the Azure portal) is not a supported operation for the resource itself; you can only pause or delete the resource, which would lose all configurations and data, making it impractical for intermittent use.

Full explanation →
62
MCQmedium

You are building an Azure AI Language solution that ingests support tickets and routes each ticket to the correct department. Each ticket must be assigned exactly one department, and departments are defined only by the examples you label in Language Studio. You need to build the model with the fewest labeling and configuration steps. Which project type and configuration should you use?

A.A custom multi-label classification project with one class per department
B.A custom single-label classification project with one class per department
C.A custom named entity recognition project with one entity type per department
D.A conversational language understanding project with one intent per department
AnswerB

Single-label classification assigns exactly one class per document, matching the requirement that each ticket maps to one department. Each department becomes a class, so labeling a ticket with its department directly trains the model. This is the minimal configuration because no multi-label scoring, entity extraction, or orchestration layer is needed, and the deployed model returns the single predicted department for routing.

Why this answer

The scenario requires exactly one department per ticket and departments defined purely by labeled examples, which is precisely what custom single-label classification provides. Each department maps to one class, so the deployed model returns a single predicted department that can drive routing without additional logic. The other project types either permit multiple labels, extract text spans, or model utterances, none of which enforces a single whole-document category.

Exam trap

The trap here is assuming that more expressive project types such as multi-label classification or entity recognition are always better, when the routing requirement of exactly one department per ticket makes single-label classification the correct and simplest choice.

Full explanation →
63
Multi-Selectmedium

You are designing an agent in Microsoft Copilot Studio that must answer questions by using a large corpus of internal policy documents. The agent must provide accurate citations and must not invent policy details. You need to configure knowledge sources and grounding behaviour. Which two actions should you take? (Choose two.)

Select 2 answers
A.Configure the agent to use generative answers with the knowledge source and review the moderation settings.
B.Disable the knowledge source and rely on the model's pre-trained policy knowledge.
C.Publish the agent immediately and let user feedback correct inaccurate policy statements over time.
D.Add the policy documents as a knowledge source and enable citations in responses.
E.Increase the agent's creativity level to high so it can paraphrase policies more naturally.
AnswersA, D

Generative answers let the agent compose responses grounded in the configured knowledge source, which is how citations and accurate policy summarization are produced. Reviewing moderation settings ensures the generated content meets organizational standards for the policy domain. Together with a knowledge source, this is the supported way to deliver grounded, cited answers.

Why this answer

Grounding the agent in the policy documents and enabling citations ensures answers trace back to authoritative content, while generative answers with reviewed moderation settings produce compliant responses. High creativity and reliance on pre-trained knowledge both increase the chance of fabricated policy details, and deferring accuracy to post-publication feedback does not meet the requirement.

Exam trap

The trap here is assuming that a more creative model produces better policy answers, when creativity increases the likelihood of unsupported details.

Full explanation →
64
MCQhard

You are developing an agentic solution that uses Azure AI Agent Service with a custom function calling tool. The agent needs to call a function that requires authentication to an external API. How should you securely pass the API key to the function?

A.Hardcode the API key in the function code
B.Use Azure Key Vault to store the API key and reference it in the function
C.Store the API key in an environment variable
D.Pass the API key as part of the agent's system prompt
AnswerB

Storing the key in Azure Key Vault keeps the secret out of code and prompt context, satisfying the secure-authentication constraint. The function retrieves it at runtime via its managed identity, so credentials are never exposed in the agent definition or logs.

Why this answer

Azure Key Vault provides a secure, centralized service for storing and managing secrets like API keys. In Azure AI Agent Service, you can configure the function to retrieve the API key at runtime from Key Vault using managed identities, ensuring the key is never exposed in code, configuration, or prompts. This follows the principle of least privilege and aligns with Azure's security best practices for agentic solutions.

Exam trap

The trap here is that candidates often choose environment variables (Option C) because they seem 'secure enough' in local development, but Azure explicitly tests that environment variables are not considered secure for production secrets in cloud-native solutions, especially when audit trails and fine-grained access control are required.

How to eliminate wrong answers

Option A is wrong because hardcoding the API key in the function code violates security best practices, as the key would be exposed in source control, logs, and compiled binaries. Option C is wrong because storing the API key in an environment variable is insecure in cloud environments; environment variables can be leaked through process dumps, logs, or misconfigured container settings, and they lack access control and auditing. Option D is wrong because passing the API key as part of the agent's system prompt would expose the secret in prompt logs, conversation history, and potentially to the language model itself, creating a severe security vulnerability.

Full explanation →
65
MCQmedium

A team is using Azure OpenAI Service to summarize long legal contracts. They observe that summaries sometimes miss clauses located near the end of the document. The contracts are far longer than the model's maximum context window. What should they implement to improve coverage of the entire document?

A.Lower the max_tokens parameter to force the model to read further into the document.
B.Increase the temperature parameter so the model explores more of the contract content.
C.Set the top_p parameter to 1.0 and rely on nucleus sampling to expand context coverage.
D.Split the contract into overlapping chunks, summarize each chunk, then combine the partial summaries into a final summary.
AnswerD

Chunking with overlap followed by a combine step is the map-reduce pattern for documents exceeding the context window. Each chunk is summarized within limits, and overlap preserves clauses that straddle boundaries. Merging partial summaries produces coverage across the whole contract, addressing the missed end-of-document clauses.

Why this answer

Documents longer than the context window must be processed in pieces. Overlapping chunking ensures clauses spanning boundaries survive, and a combine step merges per-chunk summaries into a coherent whole. Sampling parameters and output length settings operate on generation behavior and cannot extend how much source text the model receives.

Exam trap

The trap here is assuming sampling parameters such as temperature or top_p influence how much source content the model can read, when they only affect token selection.

Full explanation →
66
MCQeasy

Your company uses Azure AI Language to analyze customer feedback. You need to extract key phrases from reviews in multiple languages. Which feature should you use?

A.Named entity recognition
B.Language detection
C.Key phrase extraction
D.Sentiment analysis
AnswerC

Key phrase extraction in Azure AI Language identifies the main concepts in text and natively supports multiple languages, satisfying the multilingual reviews constraint. It returns salient terms directly from each document, so no translation pipeline is needed before analysis.

Why this answer

Key phrase extraction is the correct feature because it is specifically designed to identify and extract the most important points or topics from text, regardless of the language. Azure AI Language's key phrase extraction supports multiple languages and returns a list of key phrases that represent the main subjects discussed in the customer feedback, which directly meets the requirement.

Exam trap

The trap here is that candidates often confuse key phrase extraction with named entity recognition, assuming that extracting important names or places is the same as extracting key topics, but NER focuses on specific entity types while key phrase extraction captures broader, contextually important phrases.

How to eliminate wrong answers

Option A is wrong because named entity recognition (NER) identifies and categorizes entities like people, organizations, and locations, not the key topics or phrases that summarize the feedback. Option B is wrong because language detection only identifies the language of the text and does not extract any content or key phrases from the reviews. Option D is wrong because sentiment analysis determines the overall emotional tone (positive, negative, neutral) of the text, not the key phrases or main topics.

Full explanation →
67
MCQhard

A financial services company is building an agent that uses Azure OpenAI to generate investment advice. The agent must be monitored for toxicity and bias. Which combination of services should the team use to implement content safety monitoring?

A.Azure Cognitive Search and Azure AI Language.
B.Azure Bot Service and Azure Logic Apps.
C.Azure Machine Learning and Azure Functions.
D.Azure AI Content Safety and Azure OpenAI content filtering.
AnswerD

Azure AI Content Safety provides configurable harm categories (hate, violence, self-harm, sexual) with severity scoring, while Azure OpenAI content filtering applies policy at the model prompt and completion layer. Together they cover both model-level and application-level toxicity and bias monitoring.

Why this answer

Azure AI Content Safety provides built-in models for detecting harmful content such as hate speech, self-harm, and sexual content, while Azure OpenAI content filtering applies configurable severity-level filters (e.g., low, medium, high) to model inputs and outputs. Together, they enable real-time monitoring of toxicity and bias in generated investment advice, meeting compliance requirements for financial services.

Exam trap

The trap here is that candidates may confuse general AI services (like Azure AI Language or Azure Machine Learning) with the specific, purpose-built content safety and filtering services required for monitoring toxicity and bias in generative AI outputs.

How to eliminate wrong answers

Option A is wrong because Azure Cognitive Search is a retrieval service for indexing and querying data, not a content safety or bias detection tool, and Azure AI Language provides NLP features like sentiment analysis but lacks dedicated toxicity and bias monitoring for generative AI outputs. Option B is wrong because Azure Bot Service is a framework for building conversational agents, and Azure Logic Apps is an integration workflow service; neither includes built-in content safety or bias detection capabilities. Option C is wrong because Azure Machine Learning is a platform for training and deploying custom ML models, and Azure Functions is a serverless compute service; while you could build custom safety logic, they do not provide the pre-built, configurable content filtering and toxicity detection that Azure AI Content Safety and Azure OpenAI content filtering offer out of the box.

Full explanation →
68
Multi-Selectmedium

You are building a generative AI chatbot using Microsoft Copilot Studio. The chatbot must answer questions from a PDF document and a SQL database. Which THREE data sources can you configure? (Choose three.)

Select 3 answers
A.Custom connector to SQL database
B.SharePoint (store the PDF document)
C.Azure AI Search index
D.Dataverse (store SQL data)
E.Azure Blob Storage
AnswersA, B, D

Custom connectors allow integration with SQL databases.

Why this answer

A is correct because Microsoft Copilot Studio supports custom connectors to access external data sources like SQL databases. This allows the chatbot to query the SQL database directly using the connector's API, enabling real-time data retrieval for generative AI responses.

Exam trap

The trap here is that candidates often assume Azure AI Search or Blob Storage are natively configurable in Copilot Studio, but they require additional middleware or custom connectors, unlike SharePoint and Dataverse which are first-party supported sources.

Full explanation →
69
MCQmedium

You see the exhibit from an Azure OpenAI chat completion request. The assistant is not calling the get_weather function when asked about the weather. What is the most likely reason?

A.The required parameter 'location' is missing from the user message
B.Function calling is not supported in the current Azure OpenAI version
C.The function definition is malformed
D.The function is defined in the system message but not in the 'tools' array
AnswerD

Function calling requires each function declared in the request's 'tools' array, which the model inspects to decide when to emit a tool call. Describing get_weather only in the system message leaves no callable tool, so the assistant answers in text instead of invoking it.

Why this answer

D is correct because in Azure OpenAI, functions must be defined in the 'tools' array of the chat completion request to be available for the model to call. Defining a function only in the system message does not register it as a callable tool; the model can see the description but cannot invoke it. Without the function in 'tools', the assistant will ignore the request to call get_weather.

Exam trap

The trap here is that candidates assume listing a function in the system message is sufficient for the model to call it, but Azure OpenAI requires the function definition to be explicitly included in the 'tools' parameter of the API request.

How to eliminate wrong answers

Option A is wrong because the user message does not need to contain the 'location' parameter; the model is expected to extract or prompt for it from the conversation context. Option B is wrong because function calling is fully supported in current Azure OpenAI versions (e.g., gpt-4, gpt-35-turbo with API version 2023-12-01-preview or later). Option C is wrong because a malformed function definition would typically cause an API error or validation failure, not a silent refusal to call the function.

Full explanation →
70
MCQhard

Refer to the exhibit. You are troubleshooting an Azure OpenAI API call that is returning incomplete responses. The response stops mid-sentence. Which parameter should you adjust?

A.Increase max_tokens to 1000.
B.Remove the stop parameter.
C.Increase temperature to 1.0.
D.Increase top_p to 1.0.
AnswerA

max_tokens caps the number of tokens generated in the completion, so a low value truncates output mid-sentence. Raising it to 1000 lets the model finish, directly resolving the incomplete response. Temperature, top_p and presence penalties alter style, not truncation length.

Why this answer

The `max_tokens` parameter controls the maximum number of tokens the model can generate in a single response. When a response stops mid-sentence, it typically means the token limit was reached before the model could complete its output. Increasing `max_tokens` to 1000 provides more room for the model to finish its generation, resolving the truncation issue.

Exam trap

The trap here is that candidates confuse parameters that control output length (`max_tokens`) with those that control output diversity (`temperature`, `top_p`) or early stopping (`stop`), leading them to pick options that change style rather than capacity.

How to eliminate wrong answers

Option B is wrong because removing the `stop` parameter would not fix mid-sentence truncation; the `stop` parameter defines sequences that halt generation early, and removing it could actually make responses longer but does not address a hard token limit. Option C is wrong because increasing `temperature` to 1.0 increases randomness and creativity in the output, but does not affect the maximum length of the response; it could even lead to more verbose or erratic completions. Option D is wrong because increasing `top_p` to 1.0 enables nucleus sampling with all tokens considered, which may alter the diversity of the output but does not extend the token budget; the model will still stop when `max_tokens` is exhausted.

Full explanation →
71
MCQeasy

You need to extract key-value pairs from a large set of invoices. The invoices have a consistent layout but vary in format (PDF, TIFF). Which Document Intelligence model should you use?

A.Custom extraction model
B.Layout model
C.Read model
D.Premade invoice model
AnswerD

The premade invoice model is trained on invoice layout and extracts key-value pairs such as vendor, dates and totals, while accepting PDF and TIFF inputs. Consistent invoice layout matches its domain, avoiding custom model training.

Why this answer

The premade invoice model (D) is specifically designed to extract key-value pairs from invoices, including fields like invoice date, total amount, and vendor details, even when the invoices vary in format (PDF, TIFF). It leverages pre-trained deep learning models optimized for invoice layouts, making it the most efficient choice for this task without requiring custom training.

Exam trap

The trap here is that candidates may confuse the Layout model's ability to extract tables and structure with the specific key-value pair extraction needed for invoices, overlooking that the premade invoice model is purpose-built for this exact use case.

How to eliminate wrong answers

Option A is wrong because a custom extraction model requires labeled training data and is overkill for invoices with a consistent layout, as the premade model already handles this scenario. Option B is wrong because the Layout model extracts text, tables, and structure but does not specifically target key-value pairs like invoice fields, requiring additional post-processing. Option C is wrong because the Read model only performs OCR to extract raw text and does not identify or structure key-value pairs, making it unsuitable for invoice data extraction.

Full explanation →
72
MCQeasy

You are building a solution to detect if a person is wearing a hard hat in construction site images. You have a small dataset of labeled images. Which Azure service should you use?

A.Azure AI Vision Image Analysis
B.Azure Video Indexer
C.Azure AI Document Intelligence
D.Azure Custom Vision
AnswerD

Custom Vision trains a classification or object detection model on your own labelled images, which suits a small domain-specific dataset of hard-hat photos. The prebuilt Computer Vision service cannot reliably detect site-specific safety equipment without custom training.

Why this answer

Azure Custom Vision is the correct choice because it allows you to train a custom image classification model with your own small dataset of labeled construction site images to detect whether a person is wearing a hard hat. Unlike pre-built services, Custom Vision specializes in fine-tuning models for specific visual concepts that are not covered by general-purpose APIs, making it ideal for niche object detection tasks like hard hat detection.

Exam trap

The trap here is that candidates assume Azure AI Vision Image Analysis can handle any visual detection task because of its broad 'Image Analysis' name, but it cannot be customized for niche objects like hard hats, which requires a custom training service like Custom Vision.

How to eliminate wrong answers

Option A is wrong because Azure AI Vision Image Analysis provides pre-trained models for general image analysis (e.g., objects, tags, celebrities) but cannot be retrained on custom datasets like hard hat detection; it lacks the capability to learn new, specific classes from your labeled images. Option B is wrong because Azure Video Indexer is designed for analyzing video content (e.g., extracting insights, speech, faces) and is not suited for static image classification or custom object detection with a small dataset of images. Option C is wrong because Azure AI Document Intelligence is purpose-built for extracting text, tables, and key-value pairs from documents (e.g., invoices, forms) and has no capability for visual object detection or custom image classification.

Full explanation →
73
MCQeasy

You need to extract entities such as dates, locations, and organization names from unstructured text documents. Which Azure AI service should you use?

A.Computer Vision
B.Azure AI Language Service
C.Azure AI Document Intelligence
D.Azure AI Speech Service
AnswerB

Azure AI Language Service includes Named Entity Recognition, which identifies dates, locations and organisations in unstructured text out of the box. No training or custom model is required for these built-in entity categories, directly satisfying the stem's extraction requirement.

Why this answer

Azure AI Language Service provides pre-built capabilities for entity recognition, including extracting dates, locations, and organization names from unstructured text via its Named Entity Recognition (NER) feature. This service is specifically designed for text analytics tasks, making it the correct choice for entity extraction from text documents.

Exam trap

The trap here is that candidates often confuse Azure AI Document Intelligence (which extracts structured data from forms) with Azure AI Language Service (which performs general text analytics like NER), leading them to pick Option C for entity extraction from unstructured text.

How to eliminate wrong answers

Option A is wrong because Computer Vision is designed for analyzing images and video, not unstructured text documents; it extracts visual features like objects, faces, and OCR text, not semantic entities like dates or organizations. Option C is wrong because Azure AI Document Intelligence (formerly Form Recognizer) focuses on extracting structured data from forms and documents using pre-built or custom models, but its primary purpose is layout analysis and key-value pair extraction, not general-purpose entity recognition from unstructured text. Option D is wrong because Azure AI Speech Service handles audio-to-text transcription and speech synthesis, not entity extraction from text; it converts spoken language to text but does not perform semantic analysis like NER.

Full explanation →
74
MCQmedium

You are building a web app that allows users to upload images and receive a list of content tags, such as 'outdoor', 'tree', and 'person'. You want to use a prebuilt Azure AI service and minimize custom development. Which service should you use?

A.Azure AI Vision, using the Image Analysis API with the Tags visual feature
B.Azure AI Document Intelligence, using the prebuilt-read model
C.Azure AI Custom Vision, using a classification model
D.Azure AI Language, using key phrase extraction
AnswerA

The Image Analysis API provides a Tags feature that returns a list of content tags relevant to the image, such as 'outdoor', 'tree', or 'person'. It is prebuilt and requires no custom training, making it ideal for this scenario. You can call the API with the Tags feature enabled to get the desired output.

Why this answer

Azure AI Vision's Image Analysis API includes a Tags feature that returns content tags for an image without any custom training. This prebuilt capability directly satisfies the need for a list of tags such as 'outdoor', 'tree', and 'person'. Other services either require custom training, operate on text, or are document-focused.

Exam trap

The trap here is thinking that Custom Vision is needed for any tagging scenario, but prebuilt Vision tagging covers general concepts without training.

Full explanation →
75
MCQmedium

You are deploying an Azure AI Document Intelligence solution to process invoices. The solution must extract line-item details such as product code, quantity, and unit price. Which prebuilt model should you use?

A.prebuilt-receipt
B.prebuilt-invoice
C.prebuilt-idDocument
D.prebuilt-layout
AnswerB

The prebuilt-invoice model returns structured fields including line items, each with product code, quantity, unit price, and amount, satisfying the stem's line-item extraction requirement. Unlike the general document model, it applies invoice-specific training to locate and label these fields without custom training.

Why this answer

The prebuilt-invoice model is specifically designed to extract line-item details such as product code, quantity, and unit price from invoices. It uses deep learning models trained on thousands of invoice samples to identify and extract structured data, including tables and line items, making it the correct choice for this requirement.

Exam trap

The trap here is that candidates might choose prebuilt-layout thinking it can extract any table data, but it lacks the specialized field mapping and labeling that prebuilt-invoice provides for invoice-specific line items.

How to eliminate wrong answers

Option A is wrong because prebuilt-receipt is optimized for receipt documents, focusing on fields like merchant name, transaction date, and total amount, not the detailed line-item structure (product code, quantity, unit price) found in invoices. Option C is wrong because prebuilt-idDocument is designed to extract information from government-issued identification documents (e.g., driver's licenses, passports), such as ID number, name, and date of birth, and has no capability for invoice line-item extraction. Option D is wrong because prebuilt-layout extracts text, tables, and selection marks from documents without specialized field extraction for invoices; it returns raw table data but lacks the pre-trained model logic to identify and label specific invoice fields like product code or unit price.

Full explanation →

Page 1 of 11

Page 2

All pages

1234567891011

Practice AI-102 by domain

Target a specific domain to shore up weak areas.

Implement computer vision solutionsImplement knowledge mining and information extraction solutionsImplement generative AI solutionsPlan and manage an Azure AI solutionImplement agentic AI solutionsImplement an agentic solutionImplement knowledge mining and document intelligence solutionsImplement natural language processing solutions
See all domains with question counts →