Microsoft Azure AI Engineer Associate AI-102 AI-102 Questions 376–450 | Page 6/13

376

MCQeasy

You need to build a solution that can answer questions based on a set of PDF documents, such as product manuals. The solution should allow users to ask questions in natural language and receive answers with citations. Which Azure AI service should you use?

A.Azure AI Text Analytics

B.Azure Cognitive Search

C.Azure AI Custom Question Answering

D.Azure AI Document Intelligence (formerly Form Recognizer)

AnswerC

Builds a knowledge base from documents and provides answers with citations.

Why this answer

Azure AI Custom Question Answering (part of Azure AI Language) is specifically designed to extract question-answer pairs from semi-structured content like PDFs and provide answers with citations. It uses a deep learning-based extractive QA model that can locate answer spans within documents and return the source text as a citation, directly meeting the requirement for natural language questions and cited answers from product manuals.

Exam trap

The trap here is that candidates often confuse Azure Cognitive Search's document retrieval capability with the extractive QA and citation features of Custom Question Answering, assuming that a search engine alone can provide direct answers with citations without additional AI processing.

How to eliminate wrong answers

Option A is wrong because Azure AI Text Analytics is a pre-built service for sentiment analysis, key phrase extraction, and entity recognition, not for extractive question answering with citations from custom documents. Option B is wrong because Azure Cognitive Search is a search engine that retrieves relevant documents or passages based on keywords or vectors, but it does not natively provide extractive answer spans with citations in a conversational QA format without additional custom components. Option D is wrong because Azure AI Document Intelligence (formerly Form Recognizer) is designed for extracting structured data (e.g., tables, key-value pairs) from forms and documents, not for answering natural language questions with citations.

Full explanation →

377

MCQhard

You are using Azure OpenAI Service to generate marketing copy. You have a requirement to reduce the cost of inference without significantly impacting output quality. Which parameter should you adjust?

A.Decrease the max_tokens.

B.Increase the frequency_penalty.

C.Decrease the temperature.

D.Increase the top_p.

AnswerA

Lower max_tokens limits the output length, reducing tokens consumed and cost.

Why this answer

Decreasing max_tokens directly reduces the number of tokens generated per API call, which lowers the compute cost because Azure OpenAI charges per token (both input and output). Since the requirement is to reduce inference cost without significantly impacting output quality, reducing max_tokens is the most direct and effective parameter. It caps the response length, preventing unnecessarily verbose output while preserving the model's ability to generate high-quality, concise copy.

Exam trap

Microsoft often tests the misconception that temperature or top_p are cost-control parameters, when in fact they only affect output diversity and randomness, not token count or pricing; candidates mistakenly think lowering temperature reduces cost because it 'simplifies' output, but the real cost driver is token length.

How to eliminate wrong answers

Option B is wrong because increasing frequency_penalty reduces the likelihood of repeating the same phrases, which can actually increase token usage (and thus cost) by forcing the model to generate more varied, longer responses to avoid repetition. Option C is wrong because decreasing temperature reduces randomness and makes output more deterministic, but it does not directly affect the number of tokens generated or the cost per call; it changes the sampling behavior, not the length. Option D is wrong because increasing top_p (nucleus sampling) expands the pool of candidate tokens considered, which can lead to longer or more diverse outputs, potentially increasing token count and cost, not reducing it.

Full explanation →

378

Multi-Selecthard

Which THREE considerations are important when designing a custom skill for Azure AI Search that calls an external API for specialized data extraction?

Select 3 answers

A.The API endpoint must be reachable from the search service

B.The skill can only accept one input and produce one output

C.The skill must be written in Python

D.The skill must handle payloads up to 16 MB

E.The skill must complete within 230 seconds

AnswersA, D, E

The search service must be able to call the API over the network.

Why this answer

Correct options are A, D, and E. Custom skills must have an API endpoint reachable from the search service (public or via private endpoint) [A]. They must handle payloads up to 16 MB due to data size limits [D].

They must complete within 230 seconds (the default timeout) [E]. Option B is incorrect because custom skills can accept multiple inputs and produce multiple outputs. Option C is incorrect because custom skills can be written in any language that supports JSON.

Full explanation →

379

Multi-Selecteasy

Which TWO features of Azure AI Content Safety can help you moderate user-generated content in a social media application?

Select 2 answers

A.Self-harm content detection.

B.Hate speech severity detection.

C.PII redaction.

D.Groundedness detection.

E.Prompt injection detection.

AnswersA, B

Detects self-harm content.

Why this answer

Self-harm content detection (A) is a feature of Azure AI Content Safety that specifically identifies text or images related to self-harm, which is a critical category for moderating user-generated content in social media to prevent harm and comply with safety policies. Hate speech severity detection (B) is another core feature that classifies hate speech into severity levels (e.g., low, medium, high), enabling nuanced moderation of offensive content.

Exam trap

The trap here is that candidates may confuse Azure AI Content Safety's features with those of other Azure AI services (like Azure AI Language for PII or Azure OpenAI for prompt injection), leading them to select options that are technically valid in Azure but not part of Content Safety's core moderation capabilities.

Full explanation →

380

MCQmedium

Refer to the exhibit. You execute a search query on an Azure AI Search index and get these results. The query was 'brown fox'. Why is the first result scored higher than the second?

A.The first document has a higher value in a scoring profile field

B.The first document is more similar to the query in vector space

C.The first document was boosted by a semantic ranking function

D.The first document has a higher term frequency and better term proximity for the query terms

AnswerD

Default scoring favors higher term frequency and proximity.

Why this answer

The default scoring algorithm in Azure AI Search uses term frequency (TF) and term proximity. The first document likely has a higher term frequency for 'brown' and 'fox', and the terms are closer together (e.g., appearing as a phrase), resulting in a higher score. Option D correctly identifies this.

Option A is incorrect because no scoring profile is mentioned. Option B is incorrect because vector search is not used. Option C is incorrect because no semantic ranking is configured.

Full explanation →

381

Multi-Selecthard

You are deploying a generative AI model using Azure AI Foundry. The model must be accessible only from within a specific virtual network. Additionally, you need to monitor all API calls for auditing. Which two configurations are required? (Choose two.)

Select 2 answers

A.Assign a managed identity to the model deployment.

B.Configure CORS to allow only the VNet's domain.

C.Enable public network access from selected IP addresses.

D.Enable diagnostic settings to send logs to a Log Analytics workspace.

E.Disable public network access and configure a private endpoint.

AnswersD, E

Logs enable auditing of all API calls.

Why this answer

Enabling diagnostic settings to send logs to a Log Analytics workspace allows you to capture and audit all API calls made to the model deployment. This is essential for monitoring, security auditing, and compliance, as it records detailed telemetry such as request timestamps, caller IPs, and operation names. Option E is correct because disabling public network access and configuring a private endpoint ensures that the model is only accessible from within the specified virtual network, meeting the isolation requirement.

Exam trap

The trap here is that candidates often confuse network-level access controls (like IP whitelisting or CORS) with true VNet isolation via private endpoints, and they overlook that diagnostic settings are the standard Azure mechanism for auditing API calls, not managed identities or CORS.

Full explanation →

382

MCQeasy

You need to implement a solution that searches through a collection of scanned invoices and extracts invoice numbers, dates, and total amounts. The solution must run on a schedule without manual intervention. Which Azure service should you use?

A.Azure Bot Service

B.Azure AI Document Intelligence

C.Azure AI Search with built-in skills

D.Azure AI Foundry model catalog

AnswerB

It extracts structured fields from documents.

Why this answer

Azure AI Document Intelligence (formerly Form Recognizer) is specifically designed to extract structured data like invoice numbers, dates, and total amounts from scanned documents. It can be scheduled to run automatically via APIs or pipelines. Azure Bot Service is for conversational AI, not document extraction.

Azure AI Search with built-in skills can index and search documents but is not optimized for extracting specific fields from invoices. Azure AI Foundry model catalog is for discovering and deploying AI models, not for direct document extraction.

Full explanation →

383

MCQmedium

You are building a solution to extract customer feedback from PDF documents stored in Azure Blob Storage. The solution must extract key phrases and sentiment scores, but you cannot use any pre-built models from Azure AI Language. What should you use?

A.Use the sentiment analysis capability in Azure AI Language

B.Train custom models in Azure AI Language for key phrase extraction and sentiment analysis

C.Use Azure AI Language's pre-built key phrase extraction API

D.Use Azure AI Document Intelligence with the pre-built read model

AnswerB

Custom NER and custom text classification in Azure AI Language are not pre-built; you define the entities and labels, making them a valid solution for key phrase extraction and sentiment analysis.

Why this answer

Azure AI Language supports training custom models for named entity recognition (NER) to extract key phrases and custom text classification to predict sentiment labels. You can first extract text from PDFs using Azure AI Document Intelligence or other OCR services, then pass the text to the custom Language models. This approach meets the requirement without using any pre-built Azure AI Language models.

Options A and C rely on pre-built APIs and are disallowed. Option D only provides OCR and cannot perform semantic analysis.

Exam trap

A common pitfall is to assume that Azure AI Document Intelligence custom models can handle semantic tasks like key phrase extraction or sentiment analysis. Document Intelligence is designed for structured field extraction (e.g., forms), not natural language understanding. When pre-built Language models are prohibited, custom models in Azure AI Language are the correct choice.

How to eliminate wrong answers

Option A is wrong because it uses the pre-built sentiment analysis capability in Azure AI Language, which is explicitly prohibited by the requirement. Option C is wrong because it uses Azure AI Language's pre-built key phrase extraction API, which is also prohibited. Option D is wrong because Azure AI Document Intelligence's pre-built read model only extracts text and layout information from documents, not key phrases or sentiment scores; it does not perform semantic analysis like sentiment detection.

Full explanation →

384

Multi-Selectmedium

Which TWO actions should you take when designing an Azure AI solution that uses Microsoft Foundry to ensure responsible AI practices?

Select 2 answers

A.Implement a human-in-the-loop review for critical decisions

B.Optimize the model for maximum throughput

C.Run an AI fairness assessment on the model

D.Store all training data indefinitely for auditability

E.Remove all explainability metrics to simplify the model

AnswersA, C

Ensures oversight.

Why this answer

Implementing a human-in-the-loop review provides oversight for critical decisions. Option C is correct because running an AI fairness assessment detects bias. Option B is wrong because optimizing for maximum throughput is about performance, not responsible AI.

Option D is wrong because storing all training data indefinitely may violate privacy and is not a recommended practice for responsible AI. Option E is wrong because removing all explainability metrics hinders transparency, which is important for responsible AI.

Full explanation →

385

MCQhard

You are developing an agentic solution that uses multiple AI agents to collaborate on a complex task. To ensure the agents work together effectively, you need to define a clear handoff protocol. Which approach should you use in Azure AI Agent Service to enable agent-to-agent communication?

A.Configure a shared memory store

B.Implement a custom API for agent communication

C.Orchestrate agents sequentially using a script

D.Use Agent Handoff feature

AnswerD

Agent Handoff is designed for seamless context transfer between agents.

Why this answer

Azure AI Agent Service provides a built-in Agent Handoff feature that enables seamless agent-to-agent communication by defining a structured handoff protocol. This feature allows agents to pass tasks and context to each other without custom code, ensuring efficient collaboration in multi-agent systems.

Exam trap

The trap here is that candidates may confuse shared memory or sequential orchestration with a proper handoff protocol, not realizing that Azure AI Agent Service's Agent Handoff feature is specifically designed for dynamic, bidirectional agent-to-agent communication without custom development.

How to eliminate wrong answers

Option A is wrong because a shared memory store is used for persisting state or data across agents, not for defining a handoff protocol for agent-to-agent communication. Option B is wrong because implementing a custom API for agent communication would be redundant and inefficient, as Foundry already provides the Agent Handoff feature for this purpose. Option C is wrong because orchestrating agents sequentially using a script does not enable dynamic agent-to-agent handoffs; it imposes a rigid execution order that lacks the flexibility of a proper handoff protocol.

Full explanation →

386

MCQeasy

You have the above data source definition for Azure AI Search. You want to index only PDF files from the 'documents' container. How should you modify the data source?

A.Set the 'query' field to a blob prefix that corresponds to the folder containing PDF files.

B.Change the container name to 'pdfs' and move all PDFs there.

C.Add a 'fileExtension' property to the container object.

D.Change the connection string to use a different storage account that contains only PDFs.

AnswerA

Filters to that folder.

Why this answer

Azure AI Search data sources support a 'query' property that accepts a blob prefix or virtual directory path. By setting 'query' to the folder path containing PDFs (e.g., 'documents/pdfs/'), the indexer will only process blobs under that prefix, effectively filtering to PDF files without moving or renaming containers.

Exam trap

The trap here is that candidates often assume filtering must be done by file extension or container name, but Azure AI Search's blob indexer supports prefix-based filtering via the 'query' field, which is the simplest and most performant method for selecting blobs from a specific folder or path.

How to eliminate wrong answers

Option B is wrong because changing the container name to 'pdfs' and moving files is an unnecessary manual workaround; Azure AI Search can filter blobs natively using the 'query' field without restructuring storage. Option C is wrong because the container object in Azure Blob Storage does not support a 'fileExtension' property; file extension filtering is done via the indexer's 'indexedFileNameExtensions' configuration or a custom skill, not on the data source definition. Option D is wrong because using a different storage account is an overengineered solution that adds complexity and cost; the existing account can be filtered with the 'query' prefix to achieve the same result.

Full explanation →

387

MCQmedium

You are developing a solution that uses Azure Document Intelligence to extract data from invoices and then uses Azure OpenAI to summarize the extracted data. The solution occasionally produces summaries that omit key fields like the invoice total. What should you do to improve accuracy?

A.Set temperature to 0 to make the output more deterministic

B.Use a larger model like GPT-4 instead of GPT-3.5

C.Increase the max_tokens parameter

D.Define a structured prompt that explicitly requests each field and provide examples

AnswerD

Structured prompt with explicit requests improves adherence to required fields.

Why this answer

The issue is that the summarization prompt lacks explicit instructions for which fields to include. By defining a structured prompt that explicitly requests each key field (e.g., invoice total, date, vendor) and providing examples, you guide the Azure OpenAI model to consistently extract and include those fields in the summary, reducing omission errors. This approach leverages prompt engineering to improve output reliability without changing model parameters or size.

Exam trap

The trap here is that candidates often assume that model size or parameter tuning (temperature, max_tokens) is the primary fix for content omission, when in fact prompt engineering—specifically structured prompts with explicit field requests—is the correct solution for ensuring specific data is included in the output.

How to eliminate wrong answers

Option A is wrong because setting temperature to 0 makes the output more deterministic but does not force the model to include specific fields; it only reduces randomness in token selection, not the likelihood of omitting requested content. Option B is wrong because using a larger model like GPT-4 instead of GPT-3.5 improves general reasoning but does not guarantee that key fields are included unless the prompt explicitly requests them; the omission is a prompt design issue, not a model capability issue. Option C is wrong because increasing max_tokens only allows longer responses but does not influence which content the model chooses to include; the model may still omit fields even with a larger token budget.

Full explanation →

388

Multi-Selecthard

A company uses Azure Content Moderator to moderate user-generated content. They need to ensure that content moderation workflows comply with regional regulations. Which TWO actions should they take?

Select 2 answers

A.Deploy Content Moderator resources in the required geographic regions to meet data residency requirements.

B.Use the free tier to reduce costs while meeting compliance needs.

C.Enable geo-tagging on the Content Moderator API to automatically apply region-specific moderation.

D.Configure the API to automatically reject any content that violates regional laws.

E.Set up human review teams to handle content that requires regional context for moderation decisions.

AnswersA, E

Azure allows choosing a region to store data in compliance with regional regulations.

Why this answer

Deploying Azure Content Moderator resources in the required geographic regions ensures that user-generated content is processed and stored within specific data boundaries, directly addressing data residency regulations. This is a fundamental compliance requirement because Azure resources are region-bound, and data does not leave the selected region unless explicitly configured otherwise.

Exam trap

The trap here is that candidates often assume the API can automatically enforce regional laws (Option D) or that a single global deployment with geo-tagging (Option C) is sufficient, when in fact compliance requires explicit regional resource deployment and human-in-the-loop review for context-sensitive decisions.

Full explanation →

389

MCQmedium

A company uses Azure AI Speech for real-time captioning during live events. They notice a delay of 5 seconds between speech and caption display. Which action should they take to reduce latency?

A.Deploy a custom speech model

B.Use the Speech SDK with intermediate results enabled

C.Switch to batch transcription API

D.Increase the maxAlternatives parameter

AnswerB

Intermediate results reduce perceived latency by displaying partial captions.

Why this answer

Enabling intermediate results in the Speech SDK allows the client to receive partial, real-time recognition hypotheses as the audio is being processed, rather than waiting for the final, fully processed result. This reduces the perceived latency from the full utterance duration (which can be several seconds) to near-instantaneous display of partial captions, directly addressing the 5-second delay.

Exam trap

The trap here is that candidates often confuse latency reduction with accuracy improvements, incorrectly assuming that a custom model or more alternatives will speed up processing, when in fact the solution lies in changing the result delivery mode from final-only to streaming partial results.

How to eliminate wrong answers

Option A is wrong because deploying a custom speech model improves recognition accuracy for domain-specific vocabulary or accents, but does not reduce the fundamental processing latency of the speech-to-text pipeline; it may even add overhead for model loading. Option C is wrong because the batch transcription API is designed for asynchronous, offline processing of pre-recorded audio, not for real-time captioning, and would introduce even greater delays (minutes to hours). Option D is wrong because increasing the maxAlternatives parameter only increases the number of alternative recognition hypotheses returned in the final result, which has no effect on how quickly the first hypothesis is delivered.

Full explanation →

390

MCQeasy

You are responsible for managing costs for Azure AI services used by multiple teams. You notice that costs are higher than expected. Which two actions should you take to reduce costs without impacting performance?

A.Set up budgets and alerts in Cost Management.

B.Migrate all services to a higher tier for better performance.

C.Evaluate and choose the appropriate pricing tier based on usage patterns.

D.Deactivate unused resources to save costs.

AnswerA, C

Alerts help prevent cost overruns.

Why this answer

Setting up budgets and alerts in Cost Management (Option A) allows you to monitor spending and receive notifications, enabling proactive cost control without impacting performance. Evaluating and choosing the appropriate pricing tier based on usage patterns (Option C) ensures you are not over-provisioned, reducing costs while maintaining performance. Both actions reduce costs without impacting performance.

Exam trap

The trap here is that candidates may think deactivating unused resources (Option D) is always a correct cost-saving action, but the question specifically asks for two actions that reduce costs without impacting performance, and the correct pair is A and C, while D is a valid but not selected option in this specific scenario.

How to eliminate wrong answers

Option B is wrong because migrating to a higher pricing tier increases costs and does not reduce them; it may improve performance but contradicts the goal of reducing costs without impacting performance. Option D is wrong because deactivating unused resources can reduce costs, but the question specifies actions that do not impact performance; deactivating unused resources has no performance impact, but it is not listed as correct because the question asks for two specific actions from the provided options, and the correct pair is A and C, not D. Option D is a valid cost-saving measure but is not among the two correct answers in this context.

Full explanation →

391

MCQhard

A company is building a chatbot using Azure Language Service and wants to ensure that the chatbot can understand user intents and extract entities from user utterances. The chatbot must be able to handle multiple intents in a single utterance and must support pre-built entities such as numbers and dates. Which action should the developer take to configure the Language service accordingly?

A.Enable the 'Multiple intents' setting in the Language service project.

B.Upgrade the Language service tier from Standard to Custom.

C.Configure the project to use Orchestration workflow.

D.Set the project language to 'Multilingual' to enable entity recognition.

E.Define list entities for numbers and dates.

AnswerA

This setting allows the model to predict multiple intents for a single utterance.

Why this answer

The Azure Language Service's Conversational Language Understanding (CLU) project includes a 'Multiple intents' setting that, when enabled, allows the model to predict more than one intent per utterance. This is essential for handling compound user inputs where the user expresses multiple goals in a single sentence.

Exam trap

The trap here is that candidates often confuse the 'Multiple intents' setting with Orchestration workflow or assume that pre-built entities require manual list definitions, when in fact the former is a project-level toggle and the latter are automatically available without any custom configuration.

How to eliminate wrong answers

Option B is wrong because the Language service tier (Standard vs. Custom) refers to the pricing and feature set, not the ability to handle multiple intents or pre-built entities; the 'Custom' tier is not a valid upgrade path for CLU. Option C is wrong because Orchestration workflow is used to route utterances to different services (e.g., CLU, QnA Maker, LUIS) based on intent, but it does not itself enable multiple intents within a single CLU project; the multiple intents feature must be enabled at the project level.

Option D is wrong because setting the project language to 'Multilingual' enables the model to process utterances in multiple languages, but it does not directly enable multiple intents or pre-built entity recognition; pre-built entities like numbers and dates are available regardless of the multilingual setting. Option E is wrong because pre-built entities for numbers and dates are built-in and do not require manual definition as list entities; list entities are used for custom, fixed-value entities, not for pre-built types.

Full explanation →

392

Multi-Selecthard

Which TWO Azure AI services can be used to build a conversational AI system that handles multi-turn dialogues with context?

Select 2 answers

A.Azure AI Search

B.Azure AI Language

C.Azure OpenAI Service

D.Azure AI Content Safety

E.Azure AI Bot Service

AnswersB, C

Conversational language understanding supports multi-turn.

Why this answer

Azure AI Language provides conversational language understanding (CLU) capabilities that enable you to build a model capable of understanding multi-turn dialogues by maintaining context across user utterances. Azure OpenAI Service offers advanced language models like GPT-4, which can handle multi-turn conversations with context through prompt engineering and conversation history, making it suitable for building conversational AI systems.

Exam trap

Microsoft often tests the distinction between a platform (Azure AI Bot Service) and the actual AI services that provide the intelligence, leading candidates to mistakenly select Bot Service as a conversational AI service rather than recognizing it as a hosting and orchestration layer.

Full explanation →

393

Multi-Selectmedium

You are designing a generative AI solution using Azure OpenAI Service with your own data indexed in Azure AI Search. Which THREE components are essential for the retrieval-augmented generation (RAG) pattern?

Select 3 answers

A.Data ingestion pipeline to Azure AI Search

B.Azure AI Search index

C.Azure Functions for orchestration

D.Azure API Management for rate limiting

E.Azure OpenAI model

AnswersA, B, E

Data must be ingested into the search index.

Why this answer

A data ingestion pipeline is essential to load and index your data into Azure AI Search, enabling the retrieval step in RAG. Without this pipeline, the search index would have no data to query, breaking the retrieval-augmented generation pattern.

Exam trap

The trap here is that candidates often confuse optional production components (like Azure Functions for orchestration or API Management for rate limiting) with the core, mandatory components of the RAG pattern, which are the data source, search index, and the LLM model.

Full explanation →

394

MCQhard

You deploy the ARM template shown in the exhibit. After deployment, you need to allow access to the Language service from your on-premises application. What should you do?

A.Add an IP rule with your on-premises public IP address.

B.Remove the customSubDomainName property.

C.Set the defaultAction to Allow.

D.Change the SKU to F0 to allow public access.

AnswerA

IP rule allows specific IPs.

Why this answer

The ARM template deploys an Azure Cognitive Services Language service with a network ACL that defaults to denying all traffic (defaultAction: Deny). To allow your on-premises application to access the service, you must add an IP rule that permits traffic from your on-premises public IP address. This is because the network ACL evaluates IP rules before the default action, and adding a rule with your public IP overrides the default deny for that specific source.

Exam trap

The trap here is that candidates often confuse the 'defaultAction' property with a simple on/off switch for public access, not realizing that IP rules are evaluated first and can selectively permit traffic even when defaultAction is Deny.

How to eliminate wrong answers

Option B is wrong because removing the customSubDomainName property would not affect network access; it only controls the endpoint subdomain naming and is unrelated to IP-based access control. Option C is wrong because setting defaultAction to Allow would open the service to all public internet traffic, which is a security risk and not a targeted solution for allowing only your on-premises application. Option D is wrong because changing the SKU to F0 (free tier) does not change network access policies; the F0 SKU still respects the same network ACL rules and does not automatically enable public access.

Full explanation →

395

MCQmedium

A developer is building a chatbot using Azure Bot Service and Language Understanding (LUIS). The bot needs to handle multiple intents, including 'BookFlight', 'CancelFlight', and 'CheckWeather'. During testing, the bot frequently confuses 'BookFlight' and 'CancelFlight' intents. What is the most effective way to improve intent classification accuracy?

A.Reduce the number of intents by merging similar ones.

B.Increase the confidence threshold for intent predictions.

C.Add more entities to the utterances.

D.Add more varied training utterances for 'BookFlight' and 'CancelFlight' intents.

AnswerD

More diverse training data improves the model's ability to distinguish similar intents.

Why this answer

Adding more varied training utterances for the 'BookFlight' and 'CancelFlight' intents directly addresses the root cause of confusion: insufficient or overlapping training data. LUIS relies on diverse utterance patterns to distinguish between semantically similar intents; increasing the quantity and variety of labeled examples improves the model's ability to learn discriminative features, thereby boosting classification accuracy.

Exam trap

The trap here is that candidates often confuse confidence thresholds with model improvement, thinking that raising the threshold will fix misclassifications, when in reality it only masks the problem by rejecting more utterances instead of improving the model's discriminative power.

How to eliminate wrong answers

Option A is wrong because merging similar intents would reduce the bot's functionality and is not a best practice for improving accuracy—it avoids the problem rather than fixing the model's discrimination. Option B is wrong because increasing the confidence threshold only filters out low-confidence predictions but does not improve the underlying model's ability to distinguish between intents; it may cause more utterances to be misclassified or rejected. Option C is wrong because entities are used to extract specific data from utterances, not to differentiate between intents; adding more entities does not help the model learn which intent an utterance belongs to.

Full explanation →

396

MCQhard

You are developing a knowledge mining solution for a legal firm that needs to process thousands of legal contracts stored as PDFs in Azure Blob Storage. The solution must extract clauses, parties, and dates using a custom model. You are using Microsoft Foundry with Azure AI Search and Azure AI Document Intelligence. The custom model must be trained on labeled contract data. After training, you deploy the model and integrate it into the AI Search enrichment pipeline. The pipeline must also perform OCR for scanned contracts. You have configured the following: - A custom classification model in Document Intelligence for document types. - A custom extraction model in Document Intelligence for clauses, parties, and dates. - An Azure AI Search index with fields: clause, party, date. - A skillset with a Document Intelligence skill pointing to the custom extraction model. During testing, the pipeline runs successfully for digital PDFs but fails for scanned PDFs. The error indicates that OCR is not being applied. What should you do to fix the issue?

A.Retrain the custom extraction model with scanned document images.

B.Delete and recreate the index with a different field mapping.

C.Modify the Document Intelligence skill configuration to enable OCR processing.

D.Add an OCR skill to the skillset before the Document Intelligence skill.

AnswerC

Document Intelligence can perform OCR on images; enabling it in the skill allows processing of scanned PDFs.

Why this answer

Scanned PDFs contain images of text, so OCR must be performed to extract text before the custom extraction model can process it. Azure AI Document Intelligence's custom extraction model can perform OCR internally if the skill configuration has OCR enabled. By default, the Document Intelligence skill in the skillset does not automatically enable OCR for scanned documents; you must set the `enableOcr` property to `true` in the skill configuration.

Option A is incorrect because retraining the model with scanned images would not enable OCR; the model still needs text input. Option B is incorrect because recreating the index does not address the missing OCR step. Option D is incorrect because adding a separate OCR skill is not necessary; Document Intelligence can handle OCR internally when configured correctly.

Full explanation →

397

Multi-Selecteasy

Which ONE Azure AI service can be used to perform sentiment analysis on text?

Select 1 answer

A.Azure AI Language

B.Azure AI Search

C.Azure AI Bot Service

D.Azure AI Translator

E.Azure AI Content Safety

AnswersA

Azure AI Language includes sentiment analysis as a key feature, making it a correct choice.

Why this answer

Azure AI Language provides built-in sentiment analysis capabilities as part of its natural language processing (NLP) features, allowing you to assess the sentiment (positive, negative, neutral, or mixed) of text at the document and sentence level. Azure AI Translator does not include sentiment analysis; it only provides text translation. Therefore, only Azure AI Language is correct.

Exam trap

Candidates may mistakenly think Azure AI Translator includes sentiment analysis because it can detect the language of text, but it does not provide sentiment scores. This question tests the specific capabilities of each service.

Full explanation →

398

MCQmedium

You are building a knowledge mining solution for a financial services company that needs to extract key financial terms (e.g., revenue, EBITDA, net income) from annual reports in PDF format. The solution must use a custom skill that runs a Python script to perform the extraction. The Python script is deployed as an Azure Function. You have added the custom skill to the skillset and tested it with a small set of documents. However, when processing the full dataset, the custom skill fails with time-out errors. The Azure Function has a default timeout of 230 seconds. What should you do to resolve the issue without changing the extraction logic?

A.Configure the indexer to process documents in smaller batches.

B.Replace the custom skill with a Document Intelligence custom extraction model.

C.Split the skillset into multiple skillsets and run them sequentially.

D.Change the Azure Function to a Premium plan and increase the function timeout.

AnswerD

Premium plan allows longer timeouts, giving the script more time to execute.

Why this answer

The custom skill fails with time-out errors because the Azure Function's default timeout of 230 seconds is too short for processing large documents. Option D is correct because upgrading to a Premium plan allows you to increase the function timeout (up to 30 minutes with Premium plan, or unlimited with Dedicated plan). Option A is wrong because the indexer's batch size controls how many documents are processed simultaneously, but the timeout is per document (or per skill execution), so smaller batches won't solve the per-document timeout issue.

Option B is wrong because Document Intelligence custom extraction models are pre-built for common forms and are not suitable for custom Python extraction logic. Option C is wrong because splitting the skillset does not change the execution time of the individual skill; the same timeout per invocation would still apply.

Full explanation →

399

Matchingmedium

Match each Azure AI tool to its purpose.

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

Concepts

Matches

Drag-and-drop ML model building

Interactive code development

Command-line management of Azure resources

Programmatic access to Azure services

Run AI services on-premises

Why these pairings

Correct matches: Azure Cognitive Search is an AI search service; Azure Bot Service is for building bots; Azure Cognitive Services offers pre-built AI APIs. Common confusions include swapping Cognitive Search with Machine Learning, and Bot Service with Cognitive Services.

Full explanation →

400

MCQhard

Refer to the exhibit. A developer is training a Language Understanding (LUIS) model. When testing the phrase 'Show me the status of order 98765', the model returns the intent 'OrderStatus' with a low confidence score. What is the most likely reason?

A.The entity 'OrderNumber' should be a list entity.

B.The phrase contains an entity that was not labeled in training.

C.The phrase 'Show me the status' was not seen during training.

D.The model is overfitted to the training data.

AnswerC

The training examples use 'What is the status of' and 'Cancel order'. The model has not seen 'Show me the status' and thus has lower confidence.

Why this answer

The phrase is similar to the training utterance but includes extra words ('Show me the') and a different order number. The model may not generalize well without more varied examples. Also, entities are not labeled in the test phrase.

Full explanation →

401

MCQmedium

You are developing a generative AI solution that uses Azure OpenAI Service. The solution must generate product descriptions in multiple languages. You need to ensure that the model consistently follows specific formatting rules, such as including a bullet list of features. Which strategy should you use?

A.Fine-tune the model with a dataset containing formatted examples.

B.Set a system message with explicit formatting instructions.

C.Increase the max_tokens parameter to allow longer outputs.

D.Adjust the temperature parameter to a lower value.

AnswerB

System messages define behavior and formatting guidelines for the model.

Why this answer

System messages in Azure OpenAI Service allow you to set persistent instructions that guide the model's behavior across the entire conversation. By including explicit formatting rules—such as requiring a bullet list of features—in the system message, you enforce consistent output structure without retraining the model. This approach is efficient, cost-effective, and directly leverages the API's design for controlling response format.

Exam trap

Microsoft often tests the misconception that fine-tuning is the only way to enforce output structure, when in fact system messages provide a lightweight, zero-shot alternative for formatting control.

How to eliminate wrong answers

Option A is wrong because fine-tuning requires a large, curated dataset and significant compute resources; it is overkill for simple formatting rules and introduces risk of overfitting or losing generality, whereas a system message achieves the same goal with zero training overhead. Option C is wrong because increasing max_tokens only extends the maximum length of the response, not the structure or format; it does not enforce bullet lists or any specific formatting rules. Option D is wrong because lowering the temperature parameter reduces randomness and makes outputs more deterministic, but it does not impose explicit formatting constraints like bullet lists; it controls creativity, not structure.

Full explanation →

402

Multi-Selecthard

A company uses Azure Document Intelligence to extract data from tax forms. They need to improve accuracy for a specific field. Which TWO actions should they take?

Select 2 answers

A.Label more examples of the specific field in the training set

B.Increase the batch size in the analysis request

C.Reduce the image resolution to 200 DPI

D.Use the prebuilt-tax.us model

E.Train a custom model using 10 similar forms

AnswersA, E

More labeled examples improve model accuracy for that field.

Why this answer

Labeling more examples of the specific field in the training set directly provides the custom model with additional ground-truth annotations for that field. This increases the model's ability to learn the variations in handwriting, formatting, and layout for that field, which is the most effective way to improve extraction accuracy for a targeted field in Azure Document Intelligence custom models.

Exam trap

The trap here is that candidates often confuse prebuilt models with custom models, assuming that prebuilt models can be retrained or fine-tuned, when in fact they are static and cannot be customized for specific field accuracy improvements.

Full explanation →

403

MCQeasy

An e-commerce company wants to build an agent that helps users track orders, initiate returns, and answer FAQs. The agent should be available on the company's website and mobile app. Which Azure service should the team use to deploy the agent?

A.Azure Logic Apps

B.Azure API Management

C.Azure Bot Service

D.Azure Functions

AnswerC

Azure Bot Service provides bot hosting and channel integration.

Why this answer

Azure Bot Service is the correct choice because it provides a managed environment for building, deploying, and scaling conversational AI agents that can be integrated with multiple channels, including websites and mobile apps. It supports the Bot Framework SDK, which enables the agent to handle order tracking, returns, and FAQs through natural language understanding (NLU) with LUIS or the newer CLU service.

Exam trap

The trap here is that candidates often confuse Azure Bot Service with Azure Logic Apps or Azure Functions, mistakenly thinking that workflow automation or serverless compute alone can serve as a conversational agent, but they lack the essential dialog management, channel integration, and NLU capabilities that Azure Bot Service provides.

How to eliminate wrong answers

Option A is wrong because Azure Logic Apps is a workflow automation service for integrating apps and data, not a conversational agent platform; it lacks built-in support for dialog management, NLU, and multi-channel deployment. Option B is wrong because Azure API Management is used to publish, secure, and monitor APIs, not to host interactive conversational agents; it cannot manage user intents or multi-turn dialogues. Option D is wrong because Azure Functions is a serverless compute service for running event-driven code, but it does not provide the necessary framework for building conversational flows, channel adapters, or state management required for an agent.

Full explanation →

404

MCQhard

You are designing a solution that uses Azure AI Vision to analyze images for moderation. The solution must detect adult content and identify text in images. You need to minimize latency and cost. Which approach should you recommend?

A.Call the Analyze Image API twice: once for adult content and once for OCR

B.Use the Computer Vision 3.2 API with the 'adult' and 'OCR' parameters

C.Call the Analyze Image API with the 'adult' and 'read' visual features

D.Use the Read API for text and the Content Moderator API for adult content

AnswerC

Single call handles both tasks efficiently.

Why this answer

The Analyze Image API in Azure AI Vision supports multiple visual features in a single call, including 'adult' for adult content detection and 'read' for OCR (text extraction). This minimizes latency by avoiding multiple API calls and reduces cost since you are billed per API call, not per feature.

Exam trap

The trap here is that candidates may think separate API calls or older API versions (like Computer Vision 3.2) are required for different tasks, but the Analyze Image API supports multiple visual features in a single call, which is the most efficient approach.

How to eliminate wrong answers

Option A is wrong because calling the Analyze Image API twice doubles both latency and cost, as each call incurs a separate charge and network round-trip. Option B is wrong because the Computer Vision 3.2 API does not support an 'OCR' parameter; OCR is handled via the 'read' visual feature in the Analyze Image API or the dedicated Read API. Option D is wrong because using separate APIs (Read API for text and Content Moderator API for adult content) increases latency and cost due to multiple calls, and the Content Moderator API is a separate service that is not optimized for the same single-call efficiency as the Analyze Image API.

Full explanation →

405

MCQeasy

You need to generate a poem using Azure OpenAI. The poem should be about nature and have a cheerful tone. Which parameter should you adjust to influence the tone?

A.top_p

B.temperature

C.max_tokens

D.system message

AnswerD

System message guides the model's overall behavior and tone.

Why this answer

The system message (D) is the correct parameter to influence the tone of a generated poem because it acts as a high-level instruction that sets the behavior, persona, and style of the model. By including a directive like 'You are a cheerful poet writing about nature,' you directly control the tone without altering randomness or output length.

Exam trap

The trap here is that candidates confuse parameters that control randomness (temperature, top_p) with those that control instruction-following and style (system message), leading them to incorrectly select temperature as the primary tone influencer.

How to eliminate wrong answers

Option A is wrong because top_p controls nucleus sampling—the cumulative probability threshold for token selection—and does not directly set tone; it affects diversity of output, not style. Option B is wrong because temperature adjusts the randomness of token probabilities (higher values increase creativity, lower values make output more deterministic), but it does not specify a cheerful tone; it only influences how likely the model is to choose less probable tokens. Option C is wrong because max_tokens limits the length of the generated response and has no impact on the emotional tone or style of the poem.

Full explanation →

406

MCQhard

You are troubleshooting an agent built with Microsoft Copilot Studio. The agent uses a custom topic to check inventory levels. The topic calls a Power Automate flow that returns JSON with 'inStock' boolean. The agent sometimes says 'Item is in stock' even when the flow returns false. What is the most likely cause?

A.The Power Automate flow has a timeout and returns default true.

B.The topic's condition is using a variable that is not being updated with the flow output.

C.The agent's response is based on a different variable that defaults to true.

D.The agent's topic is not parsing the JSON output correctly.

AnswerB

The variable might be stale or not set correctly.

Why this answer

The most likely cause is that the topic's condition is referencing a variable that does not get updated with the flow's output. In Microsoft Copilot Studio, when a Power Automate flow returns data, the output must be explicitly assigned to a topic variable. If the condition checks a different variable (e.g., a default or uninitialized one), it will not reflect the actual 'inStock' value from the flow, leading to incorrect responses like 'Item is in stock' even when the flow returns false.

Exam trap

The trap here is that candidates may assume the issue is with JSON parsing (Option D) or flow timeout (Option A), but the real problem is a variable assignment mismatch, which is a subtle but critical configuration detail in Copilot Studio topic design.

How to eliminate wrong answers

Option A is wrong because a Power Automate flow timeout would typically cause an error or trigger a timeout branch, not silently return a default 'true' value; flows do not have a built-in mechanism to return default true on timeout. Option C is wrong because while the agent's response could be based on a different variable that defaults to true, this is essentially a restatement of the correct cause but lacks the specific mechanism of the variable not being updated with the flow output; the core issue is the variable assignment, not just a default value. Option D is wrong because Copilot Studio automatically parses JSON output from Power Automate flows into structured variables; incorrect parsing would usually result in an error or null value, not a consistent false positive where the agent says 'in stock' when the flow returns false.

Full explanation →

407

MCQmedium

You are using Azure AI Language to extract information from medical research papers. You need to identify terms like 'dosage', 'side effects', and 'contraindications' specific to the medical domain. Which capability should you use?

A.Prebuilt Named Entity Recognition (NER)

B.Custom Named Entity Recognition (NER)

C.PII detection

D.Entity linking

AnswerB

Custom NER allows you to train a model on your specific domain vocabulary.

Why this answer

Custom Named Entity Recognition allows you to train a model to recognize custom entities like medical terms. Option A is wrong because prebuilt NER only recognizes general entities like person, location, etc. Option C is wrong because PII detection is for personal information.

Option D is wrong because entity linking links to external knowledge bases.

Full explanation →

408

MCQeasy

You are developing a generative AI application that uses Azure OpenAI Service. You want to ensure that the application does not generate offensive content. Which Azure service should you use?

A.Azure AI Bot Service

B.Azure AI Content Safety

C.Azure AI Language

D.Azure AI Search

AnswerB

Azure AI Content Safety is designed to detect and filter offensive or harmful content.

Why this answer

Azure AI Content Safety is the correct service because it is specifically designed to detect and filter offensive, inappropriate, or harmful content in text and images. For a generative AI application using Azure OpenAI, this service can be integrated to review prompts and completions in real time, ensuring that generated outputs comply with content policies and do not contain hate speech, violence, or other offensive material.

Exam trap

The trap here is that candidates often confuse Azure AI Content Safety with Azure AI Language's moderation features, but Azure AI Language does not include a dedicated content safety API for offensive content detection, whereas Content Safety is purpose-built for this task.

How to eliminate wrong answers

Option A is wrong because Azure AI Bot Service is a platform for building conversational agents, not a content moderation or safety service; it lacks native capabilities to detect offensive content. Option C is wrong because Azure AI Language provides natural language processing features like sentiment analysis and entity recognition, but it does not include dedicated content safety filters for offensive or harmful content. Option D is wrong because Azure AI Search is a cognitive search service for indexing and querying data, not a content moderation tool; it cannot filter generated content for offensiveness.

Full explanation →

409

MCQmedium

A company uses Azure OpenAI Service to generate product descriptions. They notice that the descriptions sometimes contain factually incorrect information. Which strategy should they use to reduce hallucinations?

A.Increase the temperature parameter to 1.0.

B.Implement Retrieval-Augmented Generation (RAG) by grounding prompts with a knowledge base.

C.Reduce the max_tokens parameter to limit output length.

D.Add a system message instructing the model to be more careful.

AnswerB

RAG provides factual context from a trusted source, reducing hallucinations.

Why this answer

Retrieval-Augmented Generation (RAG) grounds the model's output in a trusted, external knowledge base, providing factual context that directly reduces hallucinations. By retrieving relevant documents and injecting them into the prompt, the model generates responses based on verified information rather than relying solely on its parametric memory, which is the primary cause of factual inaccuracies in Azure OpenAI Service.

Exam trap

The trap here is that candidates often confuse hyperparameter tuning (temperature, max_tokens) or prompt engineering (system messages) as solutions for factual accuracy, when in fact only grounding with external data (RAG) directly addresses the hallucination problem by providing a verifiable source of truth.

How to eliminate wrong answers

Option A is wrong because increasing the temperature parameter to 1.0 increases randomness and creativity in the output, which actually exacerbates hallucinations by encouraging the model to generate less predictable and potentially more fabricated content. Option C is wrong because reducing max_tokens only truncates the output length; it does not address the root cause of factual inaccuracies and may even cut off critical context or reasoning. Option D is wrong because adding a system message to 'be more careful' is a vague instruction that the model cannot reliably interpret to correct factual errors; it lacks the concrete, grounded data source that RAG provides.

Full explanation →

410

MCQmedium

You are building a multilingual support chatbot using Azure AI Language. The chatbot must understand user queries in English, Spanish, and French, and respond in the same language. The solution should minimize latency and cost. What is the recommended approach?

A.Use Azure AI Translator to translate all queries to English, process with an English-only project, then translate responses back.

B.Use a single Azure AI Language project with multilingual support enabled.

C.Create separate Azure AI Language projects for each language and route queries based on detected language.

D.Use a single English-only project and rely on Azure AI Translator for all non-English queries.

AnswerB

Multilingual support handles multiple languages in a single project, reducing latency and cost.

Why this answer

Azure AI Language supports multilingual projects natively, allowing a single model to process queries in English, Spanish, and French without additional translation steps. This minimizes latency (no round-trip translation) and cost (no Translator API consumption), while maintaining response accuracy in the user's original language.

Exam trap

The trap here is that candidates often assume translation is necessary for multilingual support, overlooking Azure AI Language's built-in multilingual capability, which is more efficient and cost-effective.

How to eliminate wrong answers

Option A is wrong because it introduces unnecessary latency and cost by translating every query to English and back, and it risks losing nuance or context during translation. Option C is wrong because maintaining separate projects for each language increases management overhead, duplicates training effort, and does not leverage Azure AI Language's built-in multilingual capability, which is more efficient. Option D is wrong because it forces all non-English queries through translation, adding latency and cost, and fails to use the native multilingual support that Azure AI Language provides for direct processing.

Full explanation →

411

MCQmedium

You are building a knowledge mining solution using Azure AI Search and Azure AI Language. The solution must extract key phrases, entities, and sentiment from customer feedback documents. After processing, the enriched content should be stored in the search index for full-text search. You need to configure the enrichment pipeline. Which two Azure AI services should you integrate?

A.Azure AI Language and Azure AI Search

B.Azure AI Language and a custom skill in Azure Functions

C.Azure AI Translator and Azure AI Search

D.Azure AI Document Intelligence and Azure AI Search

AnswerA

Language provides the required skills; Search indexes the enriched content.

Why this answer

The solution requires key phrase extraction, entity recognition, and sentiment analysis, which are all capabilities of Azure AI Language. Azure AI Search is the other required service because it provides the search index and the enrichment pipeline (skillset) that invokes the AI Language skills to process documents and store the enriched content for full-text search.

Exam trap

The trap here is that candidates may confuse Azure AI Document Intelligence or Azure AI Translator as the source for text analytics, when in fact Azure AI Language is the specific service that provides key phrase, entity, and sentiment extraction as built-in cognitive skills in the search enrichment pipeline.

How to eliminate wrong answers

Option B is wrong because while Azure AI Language is needed, a custom skill in Azure Functions is unnecessary for standard key phrase, entity, and sentiment extraction—these are built-in cognitive skills in Azure AI Search. Option C is wrong because Azure AI Translator handles language translation, not key phrase extraction, entity recognition, or sentiment analysis, which are the required enrichments. Option D is wrong because Azure AI Document Intelligence (formerly Form Recognizer) extracts text and layout from documents but does not perform key phrase, entity, or sentiment analysis; those are AI Language capabilities.

Full explanation →

412

MCQmedium

A company uses Azure AI Search to index product catalogs. The search must support multilingual queries and return results in the user's language. What should you configure?

A.Assign appropriate language analyzers to fields

B.Create synonym maps for each language

C.Configure scoring profiles based on language

D.Enable semantic search

AnswerA

Language analyzers handle language-specific tokenization and stemming.

Why this answer

Azure AI Search allows you to assign language-specific analyzers (e.g., Microsoft English, French, Arabic) to individual fields in the index. When a query is submitted, the search engine uses the analyzer associated with the field to tokenize and normalize the text according to the linguistic rules of that language, enabling accurate multilingual search and returning results in the user's language.

Exam trap

The trap here is that candidates confuse semantic search (which improves relevance via deep learning) with language-specific analysis, not realizing that semantic search still requires language analyzers for proper tokenization and stemming in multilingual scenarios.

How to eliminate wrong answers

Option B is wrong because synonym maps expand query terms to include equivalent terms (e.g., 'car' and 'automobile') but do not handle language-specific tokenization, stemming, or normalization required for multilingual search. Option C is wrong because scoring profiles boost results based on metadata like freshness or field weight, not on language detection or linguistic processing. Option D is wrong because semantic search improves relevance by understanding intent and context (using L2 re-ranking) but does not inherently support multilingual analysis; it still relies on language analyzers for tokenization.

Full explanation →

413

MCQmedium

Your organization has a knowledge base of technical manuals in PDF format. You need to enable users to ask natural language questions and get answers from the manuals. Which solution should you build?

A.Azure AI Search with integrated vectorization and semantic search

B.Azure OpenAI Service with GPT-4o and Azure AI Search as a data source

C.Azure AI Language custom question answering with the documents as sources

D.Azure AI Document Intelligence to extract text and then use Azure AI Search

AnswerC

Provides direct answers from documents.

Why this answer

Azure AI Language custom question answering is specifically designed to ingest documents (like PDFs) and provide a natural language Q&A interface over them. It uses a built-in extractive reader to find answer spans directly from the source text, making it the most straightforward solution for answering questions from a static knowledge base of technical manuals without requiring additional search or vectorization infrastructure.

Exam trap

The trap here is that candidates often confuse retrieval (search) with extraction (question answering), assuming that any solution involving Azure AI Search or GPT-4o is automatically the best for Q&A, when in fact custom question answering is the purpose-built service for direct answer extraction from documents.

How to eliminate wrong answers

Option A is wrong because Azure AI Search with integrated vectorization and semantic search is a retrieval system that returns relevant document chunks or passages, not a direct question-answering service that extracts precise answer spans from the text. Option B is wrong because Azure OpenAI Service with GPT-4o and Azure AI Search as a data source uses a RAG (Retrieval-Augmented Generation) pattern that requires custom orchestration and prompt engineering to generate answers, whereas the question asks for a built-in solution that directly answers from the manuals. Option D is wrong because Azure AI Document Intelligence only extracts text from PDFs (OCR/layout analysis) and does not provide any natural language question-answering capability; pairing it with Azure AI Search still only gives retrieval, not answer extraction.

Full explanation →

414

MCQmedium

A company uses Azure AI Vision to analyze product images in an e-commerce application. The solution uses the Analyze Image API with OCR. Recently, the OCR accuracy has decreased for images with handwritten text on product labels. What should the team do to improve accuracy?

A.Enable OCR in the Analyze Image API configuration.

B.Use Azure AI Document Intelligence prebuilt receipt model.

C.Switch to the Read API for OCR.

D.Retrain the OCR model with additional handwritten samples.

AnswerC

The Read API is better for handwritten text.

Why this answer

The Read API (part of Azure AI Vision) is specifically optimized for extracting text from images, including handwritten text, and provides higher accuracy for OCR scenarios compared to the general-purpose Analyze Image API. The Analyze Image API's OCR feature is designed for printed text and does not handle handwritten content as effectively. By switching to the Read API, the team leverages a dedicated OCR engine that supports both printed and handwritten text, improving accuracy for product labels.

Exam trap

The trap here is that candidates assume all OCR features in Azure AI Vision are equivalent, but the Read API is a separate, more advanced service specifically designed for handwritten and complex text extraction, while the Analyze Image API's OCR is a legacy feature for printed text only.

How to eliminate wrong answers

Option A is wrong because enabling OCR in the Analyze Image API configuration does not change the underlying OCR engine; it still uses the same printed-text-focused OCR that has poor accuracy for handwritten text. Option B is wrong because the Azure AI Document Intelligence prebuilt receipt model is designed for extracting structured data from printed receipts, not for general OCR of handwritten text on product labels. Option D is wrong because the Azure AI Vision OCR models are pre-trained and cannot be retrained with custom samples; retraining is not a supported capability for these APIs.

Full explanation →

415

MCQmedium

A company is building a custom question-answering solution using Azure AI Language. They need to ensure that the model can provide answers from a set of internal documents, but only to authenticated users from the company's Azure Active Directory tenant. The solution should minimize latency and cost. Which deployment option should the team choose?

A.Deploy a custom question-answering model using a dedicated Azure AI Language resource with a private endpoint and managed identity.

B.Use the public endpoint of Azure AI Language with Azure AD token-based authentication.

C.Deploy a serverless endpoint with Azure Functions and Azure Cognitive Search.

D.Use the prebuilt question-answering model from Azure AI Language with a custom answer list.

AnswerA

This provides secure access via private endpoint and managed identity, and dedicated resources ensure low latency and predictable cost.

Why this answer

Deploying a custom question-answering model with a dedicated Azure AI Language resource, a private endpoint, and managed identity ensures that only authenticated users from the company's Azure AD tenant can access the solution via private network connectivity, minimizing latency by avoiding public internet routing and reducing cost by using a dedicated (not serverless) resource that can be right-sized.

Exam trap

The trap here is that candidates often assume Azure AD token-based authentication alone is sufficient for security, overlooking that a public endpoint still exposes the service to internet-based attacks and latency, while private endpoints are required for true network isolation.

How to eliminate wrong answers

Option B is wrong because using the public endpoint with Azure AD token-based authentication still exposes the endpoint to the public internet, increasing latency and security risk, and does not meet the requirement for private access. Option C is wrong because deploying a serverless endpoint with Azure Functions and Azure Cognitive Search introduces additional components and cold-start latency, increasing cost and complexity without providing the native private endpoint and managed identity integration of Azure AI Language. Option D is wrong because the prebuilt question-answering model cannot be customized to answer from a specific set of internal documents; it only provides general answers from a predefined knowledge base, and custom answer lists are limited to static Q&A pairs, not document-based retrieval.

Full explanation →

416

Multi-Selectmedium

Which TWO actions should you take to ensure compliance with data privacy regulations when using Azure AI Language to process customer support transcripts that contain personally identifiable information (PII)?

Select 2 answers

A.Disable diagnostic logging for the resource.

B.Use customer-managed keys (CMK) for encryption.

C.Apply Azure AI Content Safety filters to anonymize personal data.

D.Enable PII detection and redaction in the Azure AI Language service.

E.Configure data residency by selecting the appropriate Azure region.

AnswersD, E

PII detection identifies and redacts sensitive information.

Why this answer

Azure AI Language's PII detection and redaction feature is specifically designed to identify and mask personally identifiable information (PII) in text, such as names, addresses, and social security numbers, directly within the service. This built-in capability ensures that sensitive data is removed or obfuscated before storage or further processing, directly supporting compliance with data privacy regulations like GDPR or HIPAA.

Exam trap

The trap here is that candidates often confuse data encryption (Option B) with data anonymization or redaction, assuming that encrypting the data at rest or in transit is sufficient for privacy compliance, whereas regulations like GDPR require active masking or removal of PII from the content itself, not just cryptographic protection.

Full explanation →

417

MCQeasy

Refer to the exhibit. You have this Azure AI Search indexer configuration. The indexer is failing after processing 6 documents that contain errors. What should you do to ensure the indexer continues processing even if some documents fail?

A.Decrease the batch size to 5

B.Increase batch size to 20

C.Increase maxFailedItems to a higher value, such as 100

D.Remove the schedule to run the indexer on demand

AnswerC

Increasing maxFailedItems allows more failures before stopping.

Why this answer

The indexer is failing after processing 6 documents because the `maxFailedItems` threshold has been reached. By increasing `maxFailedItems` to a higher value like 100, the indexer will continue processing even if more documents fail, as long as the total number of failed items stays below the new threshold.

Exam trap

The trap here is that candidates confuse batch size with failure tolerance, thinking that reducing batch size will prevent the indexer from stopping, when in fact the `maxFailedItems` parameter is the direct control for how many document failures are tolerated before the indexer halts.

How to eliminate wrong answers

Option A is wrong because decreasing the batch size to 5 would reduce the number of documents processed per batch, but it does not affect the `maxFailedItems` threshold that causes the indexer to stop after 6 failures. Option B is wrong because increasing the batch size to 20 would process more documents per batch, but it does not change the `maxFailedItems` limit; the indexer would still stop after 10 failures (default). Option D is wrong because removing the schedule to run the indexer on demand does not alter the failure handling behavior; the indexer would still stop after exceeding `maxFailedItems` regardless of how it is triggered.

Full explanation →

418

MCQmedium

Your organization uses Azure OpenAI Service to generate code snippets. You want to log all user prompts and model responses for auditing purposes. What should you configure?

A.Use Azure API Management to log requests and responses.

B.Store logs in Azure Key Vault.

C.Enable diagnostic settings in Azure OpenAI Service to send logs to a Log Analytics workspace.

D.Use Azure AI Search to index the prompts and responses.

AnswerC

Diagnostic settings capture API call logs including prompts and completions for auditing.

Why this answer

Azure OpenAI Service's diagnostic settings allow you to stream platform logs—including user prompts and model responses—to a Log Analytics workspace. This is the native, built-in mechanism for capturing and analyzing request/response data for auditing, without needing additional services or custom middleware.

Exam trap

The trap here is that candidates often confuse Azure API Management's logging capabilities with Azure OpenAI's native diagnostic logging, assuming you need a separate gateway to capture request/response data, when in fact the service's own diagnostic settings provide this out-of-the-box.

How to eliminate wrong answers

Option A is wrong because Azure API Management is a gateway for managing APIs, not a logging sink for Azure OpenAI Service; it would require routing all traffic through APIM and custom policies to capture payloads, which is unnecessary and adds latency. Option B is wrong because Azure Key Vault is designed for securely storing secrets, keys, and certificates, not for storing large volumes of log data like prompts and responses. Option D is wrong because Azure AI Search is a search indexing service for building search experiences over your own data, not a logging or auditing solution; it cannot capture or store operational logs from Azure OpenAI Service.

Full explanation →

419

MCQmedium

A company uses Azure Computer Vision to moderate user-generated content. The solution must detect adult content and flag it. Which API should you call?

A.Read API

B.Analyze API with visualFeatures set to 'Adult'

C.Detect API

D.Describe API

AnswerB

Detects adult/racy content.

Why this answer

The Analyze API with the visualFeatures parameter set to 'Adult' is the correct choice because Azure Computer Vision's Analyze Image operation includes an 'Adult' category that specifically detects adult, racy, and gory content in images. This API returns a confidence score (0 to 1) for each category, allowing the solution to flag content based on a threshold. The other APIs do not provide adult content moderation capabilities.

Exam trap

The trap here is that candidates may confuse the Analyze API's 'Adult' feature with the 'Description' or 'Tags' features, assuming that general image analysis can detect adult content, but only the explicit 'Adult' visualFeature parameter provides the specialized moderation scores.

How to eliminate wrong answers

Option A is wrong because the Read API is designed for optical character recognition (OCR) to extract printed and handwritten text from images, not for detecting adult content. Option C is wrong because the Detect API is used for object detection (identifying and locating objects within an image), not for content moderation. Option D is wrong because the Describe API generates human-readable captions describing the content of an image, but it does not include adult content classification or scoring.

Full explanation →

420

MCQmedium

You are building a knowledge mining solution for a legal firm that needs to extract key clauses from thousands of scanned contract PDFs. The solution must identify parties, effective dates, and termination conditions. Which Azure AI service should you use as the primary component?

A.Azure AI Document Intelligence

B.Azure AI Vision

C.Azure AI Language

D.Azure AI Search

AnswerA

Azure AI Document Intelligence (formerly Form Recognizer) is designed to extract data from documents, including scanned PDFs.

Why this answer

Azure AI Document Intelligence (formerly Form Recognizer) is the appropriate service for extracting structured data like parties, effective dates, and termination conditions from scanned contract PDFs. It offers prebuilt models for contracts and custom extraction capabilities. Azure AI Language is for text analytics and NLP but lacks the OCR and layout understanding needed for scanned documents.

Azure AI Vision provides OCR but not structured extraction. Azure AI Search is for indexing and searching, not extraction. Therefore, Option A is correct.

Full explanation →

421

MCQhard

Refer to the exhibit. You have this skillset definition for an Azure AI Search enrichment pipeline. You notice that the entity recognition skill is not executing on any document. What is the most likely cause?

A.The entity recognition skill requires a language code that is missing

B.The split skill is not producing pages because the content is too short

C.The entity recognition skill is not registered in the skillset

D.The input source path in the entity recognition skill should be relative to the context, not absolute

AnswerD

The absolute path '/document/pages/*' conflicts with the context; relative path should be used.

Why this answer

The entity recognition skill has its context set to '/document/pages/*', which means it iterates over each page in the split output. However, its input source is '/document/pages/*' (absolute path), which points to the entire page array element rather than the text content within each page. The input source should be relative to the context, e.g., 'text' if the page had a property, or simply the context itself if the page is a string.

Using an absolute path that matches the context can cause the skill to fail because it expects a scalar input, but the context iteration provides the array element. Therefore, the correct fix is to make the input source relative, such as using 'text' or '/document/pages/*/text' appropriately. Option D correctly identifies this issue.

Exam trap

In Azure AI Search skillsets, input source paths must be relative to the skill's context when the context iterates over an array, unless the absolute path points to a property of the array element. Using an absolute path that equals the context path can cause the skill to not execute.

Full explanation →

422

Multi-Selecthard

Which TWO configurations are required to enable incremental enrichment in Azure AI Search?

Select 2 answers

A.Configure the indexer to run in 'once' mode.

B.Enable blob metadata extraction in the indexer.

C.Add a custom skill that outputs a hash of the document content.

D.Define a projection in the skillset to store enriched data.

E.Set the 'cacheKey' property in the skillset to a unique document identifier.

AnswersD, E

Projection stores intermediate state.

Why this answer

Defining a projection in the skillset is required to store enriched data in a knowledge store, which is a prerequisite for incremental enrichment. Incremental enrichment relies on caching enriched content so that only changed documents are reprocessed, and projections define how that enriched data is stored in Azure Storage for later reuse.

Exam trap

The trap here is that candidates often confuse enabling blob metadata extraction (Option B) with the caching mechanism required for incremental enrichment, but metadata extraction alone does not create or manage the enrichment cache.

Full explanation →

423

MCQmedium

Your team is building a custom ChatGPT-like copilot using Azure AI Foundry that answers questions based on internal HR policies stored in SharePoint. The solution must retrieve only the most relevant documents to minimize token usage. Which Azure AI Search feature should you configure?

A.Synonyms

B.Scoring profiles

C.Semantic ranking

D.Filters

AnswerC

Semantic ranking uses deep learning to re-rank results for better relevance.

Why this answer

Semantic ranking is the correct feature because it uses deep neural networks to re-rank search results based on semantic relevance to the query, ensuring only the most contextually appropriate documents are returned. This directly minimizes token usage by reducing the number of irrelevant documents passed to the copilot, which is critical for cost and performance in a ChatGPT-like system.

Exam trap

The trap here is that candidates often confuse semantic ranking with scoring profiles or filters, assuming any ranking or filtering mechanism can achieve semantic relevance, but only semantic ranking uses deep learning to understand query intent beyond keyword matching.

How to eliminate wrong answers

Option A is wrong because synonyms expand queries to include alternate terms (e.g., 'vacation' vs. 'leave'), which can increase the number of retrieved documents and token usage, not minimize it. Option B is wrong because scoring profiles boost results based on static field values or freshness (e.g., 'last modified date'), but they do not re-rank for semantic relevance to the query, so they may still return many irrelevant documents. Option D is wrong because filters narrow results by exact field values (e.g., 'department = HR'), but they cannot assess semantic similarity between the query and document content, so they may miss relevant documents or include irrelevant ones that match the filter criteria.

Full explanation →

424

MCQeasy

You need to provide a team of developers with access to manage Azure AI resources in a specific resource group. The developers should be able to create, read, update, and delete AI resources, but not manage access control (IAM). Which built-in role should you assign?

A.Cognitive Services Contributor

B.Owner

C.Reader

D.Contributor

AnswerD

Contributor allows create/read/update/delete but not IAM.

Why this answer

The Contributor role grants full access to manage all resources within a resource group, including creating, reading, updating, and deleting Azure AI resources, but explicitly denies the ability to manage access control (IAM). This matches the requirement exactly, as developers need full resource management without IAM permissions.

Exam trap

The trap here is that candidates often confuse the Contributor role with the Cognitive Services Contributor role, mistakenly thinking the latter provides broader resource group-level management, when in fact it is limited to Cognitive Services resources only.

How to eliminate wrong answers

Option A is wrong because Cognitive Services Contributor is a built-in role that provides full access to Azure Cognitive Services resources, but it is scoped to the Cognitive Services resource level and does not grant the ability to manage other AI resource types (e.g., Azure Machine Learning, Azure Bot Service) within the resource group. Option B is wrong because the Owner role grants full access to all resources, including the ability to manage access control (IAM), which violates the requirement that developers should not manage IAM. Option C is wrong because the Reader role only allows read-only access to resources; it does not permit create, update, or delete operations.

Full explanation →

425

MCQmedium

Your NLP solution uses custom text classification in Azure AI Language. You need to improve the model's accuracy. Which action should you take?

A.Add more labeled training data with balanced classes.

B.Increase the number of training epochs.

C.Decrease the batch size.

D.Use a different pretrained model.

AnswerA

More labeled data improves accuracy.

Why this answer

Custom text classification in Azure AI Language relies on supervised learning, where model accuracy is directly proportional to the quantity and quality of labeled training data. Adding more labeled examples with balanced class distributions reduces bias, improves generalization, and helps the model learn distinguishing features more effectively, which is the most impactful action for improving accuracy.

Exam trap

The trap here is that candidates often confuse hyperparameter tuning (epochs, batch size) with data quality improvements, but Azure AI Language's custom text classification abstracts away most training hyperparameters, making data augmentation the only viable lever for accuracy gains.

How to eliminate wrong answers

Option B is wrong because increasing the number of training epochs can lead to overfitting, where the model memorizes the training data rather than learning generalizable patterns, and Azure AI Language's custom text classification does not expose epoch tuning as a user-configurable parameter. Option C is wrong because decreasing the batch size affects training stability and convergence speed but does not inherently improve model accuracy; it can even introduce noise and slower convergence without addressing data quality or class balance. Option D is wrong because Azure AI Language's custom text classification uses a fixed pretrained model (e.g., BERT-based) that is not user-selectable; the service automatically fine-tunes the underlying model, so choosing a different pretrained model is not an available action.

Full explanation →

426

Matchingmedium

Match each Azure AI scenario to the appropriate service.

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

Concepts

Matches

Computer Vision

Speech Translation

Form Recognizer

Text Analytics

QnA Maker

Why these pairings

Correct matches: Chatbot -> Bot Service, Image moderation -> Computer Vision, Sentiment -> Text Analytics. Common confusions include associating Bot Service with image tasks or mistaking Cognitive Search for text analytics.

Full explanation →

427

MCQmedium

You are designing a knowledge mining solution for a publishing company that needs to extract metadata from thousands of book manuscripts in various formats (PDF, Word, EPUB). The solution must identify authors, publication dates, and chapter titles. You are using Microsoft Foundry with Azure AI Search and Azure AI Document Intelligence. The manuscripts are stored in Azure Blob Storage. You need to ensure that the solution can handle all file formats. You have configured a skillset with a Document Intelligence skill for the PDFs and Word documents. However, the EPUB files are not being processed. What should you do to include EPUB files in the enrichment pipeline?

A.Use Azure AI Document Intelligence to extract text from EPUB files directly.

B.Develop a custom skill that converts EPUB files to plain text and add it to the skillset.

C.Modify the Document Intelligence skill to accept EPUB files.

D.Register a new data source type for EPUB in Azure AI Search.

AnswerB

A custom skill can convert unsupported formats into text that the pipeline can process.

Why this answer

Azure AI Search indexer does not natively support EPUB files. The Document Intelligence skill can only process formats it supports (PDF, Word, etc.). Therefore, a custom skill is needed to convert EPUB files to plain text or a supported format before they can be processed by the enrichment pipeline.

Option A is wrong because Azure AI Document Intelligence does not support EPUB directly. Option C is wrong because the Document Intelligence skill cannot be modified to accept new formats. Option D is wrong because registering a new data source type does not enable processing of unsupported file formats.

Full explanation →

428

MCQeasy

You are extracting text from scanned documents that are in French. Which capability of Azure AI Document Intelligence should you use?

A.Custom model

B.Read API

C.Layout model

D.Prebuilt invoice model

AnswerB

Read API supports OCR in over 100 languages.

Why this answer

The Read API is the correct choice because it is specifically designed for extracting printed and handwritten text from scanned documents, including support for multiple languages like French. It performs optical character recognition (OCR) to digitize text without requiring any additional training or customization, making it ideal for general text extraction from scanned documents.

Exam trap

The trap here is that candidates often confuse the Read API with the Layout model, assuming that structural analysis is required for text extraction, but the Read API is the dedicated OCR solution for plain text extraction from scanned documents.

How to eliminate wrong answers

Option A is wrong because a custom model requires labeled training data and is used for extracting specific fields from structured documents, not for general text extraction from arbitrary scanned documents. Option C is wrong because the Layout model extracts text along with structural information like tables and selection marks, which is more than what is needed for simple text extraction from scanned documents. Option D is wrong because the prebuilt invoice model is specialized for extracting fields from invoices (e.g., totals, dates) and is not designed for general text extraction from arbitrary scanned documents in French.

Full explanation →

429

MCQhard

A security company uses Azure Video Analyzer on IoT Edge to detect intrusions. The edge device has limited compute and network. They need to reduce latency. What should they configure?

A.Increase the video resolution sent to the cloud.

B.Set a high minimum confidence threshold for detection.

C.Enable cloud-based processing for all frames.

D.Use multiple AI models simultaneously.

AnswerB

Filters out low-confidence results, saving compute.

Why this answer

Setting a high minimum confidence threshold for detection reduces the number of false positives and the volume of events that need to be processed and transmitted. This directly lowers the computational load on the edge device and reduces network bandwidth usage, thereby decreasing latency for actionable intrusion alerts.

Exam trap

The trap here is that candidates often assume increasing cloud processing (Option C) or using more models (Option D) improves accuracy, but they overlook the critical constraint of limited compute and network on the edge device, which makes local filtering via confidence thresholds the correct latency-reducing strategy.

How to eliminate wrong answers

Option A is wrong because increasing video resolution sent to the cloud increases the data size per frame, which consumes more network bandwidth and processing time on the edge device, worsening latency rather than reducing it. Option C is wrong because enabling cloud-based processing for all frames would require continuous high-bandwidth uploads from the edge device, defeating the purpose of edge processing and increasing end-to-end latency due to network round trips. Option D is wrong because using multiple AI models simultaneously increases the computational load on the resource-constrained edge device, leading to higher processing latency and potential queueing delays.

Full explanation →

430

Multi-Selecthard

You are planning an Azure AI solution that processes sensitive customer data. Compliance requires encryption at rest using a customer-managed key (CMK) stored in Azure Key Vault. The AI resource must also be accessible only from specific virtual networks. Which two configurations are necessary? (Choose two.)

Select 2 answers

A.Assign a system-assigned managed identity to the Azure AI service.

B.Deploy Azure Firewall to inspect outbound traffic.

C.Enable customer-managed key encryption for the Azure AI service.

D.Enable public network access from selected IP addresses.

E.Configure private endpoints for the Azure AI service.

AnswersC, E

CMK is needed for encryption at rest with customer control.

Why this answer

This is a multi-select question requiring two correct answers. Option C is correct because enabling customer-managed key encryption for the Azure AI service meets the compliance requirement for encryption at rest with CMK stored in Azure Key Vault. Option E is correct because configuring private endpoints ensures the AI resource is accessible only from specific virtual networks, providing network isolation.

Option A is incorrect because a system-assigned managed identity is not required for CMK or network access; it is used for authentication. Option B is incorrect because Azure Firewall inspects outbound traffic, which is not mandated by the requirements. Option D is incorrect because public network access must be disabled when using private endpoints; allowing selected IP addresses does not achieve the required virtual network restriction.

Full explanation →

431

MCQmedium

You are developing an agentic solution that uses Azure AI Search as a knowledge base. The agent needs to retrieve the most relevant documents based on a user query. You notice that the agent sometimes returns irrelevant results. Which configuration should you adjust?

A.Change the language analyzer to a different language

B.Increase the 'top' parameter to retrieve more candidate documents

C.Disable semantic ranking

D.Decrease the minimum search score threshold

AnswerB

More candidates increase chance of relevant results.

Why this answer

Increasing the 'top' parameter retrieves more candidate documents from Azure AI Search, which gives the agent a larger pool of potentially relevant results to rank and filter. This can help when the initial set of top documents misses relevant content, as the agent can then apply its own reasoning or additional reranking to select the best matches. The 'top' parameter controls the number of search results returned, not the relevance scoring itself.

Exam trap

The trap here is that candidates often confuse the 'top' parameter with relevance scoring or assume that adjusting scoring thresholds (like minimum score) will fix retrieval gaps, when in fact the issue is simply that the initial candidate set is too small.

How to eliminate wrong answers

Option A is wrong because changing the language analyzer affects how text is tokenized and stemmed for linguistic processing, but it does not address the core issue of retrieving irrelevant results due to an insufficient number of candidate documents. Option C is wrong because disabling semantic ranking would remove the reranking capability that improves relevance by understanding query intent, which would likely worsen, not improve, result quality. Option D is wrong because decreasing the minimum search score threshold would include more low-relevance documents, potentially increasing irrelevant results rather than reducing them.

Full explanation →

432

MCQhard

Refer to the exhibit. You are calling the Azure AI Language API for entity linking. What is the primary purpose of this request?

A.To identify entities in the text and link them to a knowledge base.

B.To extract named entities from the text without linking.

C.To extract key phrases from the text.

D.To analyze the sentiment of the text.

AnswerA

Entity linking maps entities to known entities in a knowledge base.

Why this answer

The request is configured for entity linking, which is a specific capability of the Azure AI Language API that identifies named entities in the text and resolves them to a unique identifier in a knowledge base (such as Wikipedia or a custom knowledge graph). This goes beyond simple named entity recognition (NER) by providing a canonical link, enabling disambiguation of entities with the same name (e.g., 'Washington' as a state vs. a person). The response includes both the entity name and a URL to the knowledge base entry, confirming the primary purpose is linking to a knowledge base.

Exam trap

The trap here is that candidates confuse Named Entity Recognition (NER) with Entity Linking, assuming both simply 'find entities,' but the key differentiator is that entity linking explicitly resolves entities to a knowledge base with a unique identifier and URL, which is the core purpose of this request.

How to eliminate wrong answers

Option B is wrong because extracting named entities without linking is the function of Named Entity Recognition (NER), not entity linking; the request explicitly uses the 'entityLinking' task, not 'entities'. Option C is wrong because key phrase extraction is a separate API capability (KeyPhraseExtraction) that identifies important terms without entity resolution or linking. Option D is wrong because sentiment analysis is performed by the SentimentAnalysis task, which returns sentiment scores and opinions, not entity links or knowledge base references.

Full explanation →

433

Multi-Selectmedium

You are building a solution that uses Azure AI Language to analyze customer support transcripts. You need to detect personally identifiable information (PII) and also redact the detected PII from the text. Which TWO features should you use? (Select TWO.)

Select 2 answers

A.PII detection

B.Translation

C.Sentiment analysis

D.PII redaction

E.Key phrase extraction

AnswersA, D

PII detection identifies PII entities in text.

Why this answer

PII detection (Option A) is correct because it is the Azure AI Language feature specifically designed to identify categories of personally identifiable information such as names, addresses, phone numbers, and social security numbers within text. PII redaction (Option D) is correct because it is the companion feature that replaces the detected PII entities with placeholder tokens (e.g., '********') or masks them, enabling safe sharing of transcripts. Together, they fulfill the requirement to both detect and redact PII from customer support transcripts.

Exam trap

The trap here is that candidates may confuse 'PII detection' with 'PII redaction' as a single feature, or mistakenly think that 'Key phrase extraction' or 'Sentiment analysis' can identify personal data, when in fact only the dedicated PII detection and redaction features handle that task.

Full explanation →

434

MCQhard

Your organization uses Azure OpenAI Service with a data source configured as 'Azure OpenAI on your data'. You notice that the responses include outdated information even though the underlying data source has been updated. What is the most likely cause?

A.The model is using a cached version of the prompt

B.The index in Azure Cognitive Search has not been refreshed

C.The data source is configured to sync only daily

D.The Azure CDN is caching the responses

AnswerB

The search index must be re-indexed to reflect data changes.

Why this answer

When using Azure OpenAI Service with 'Azure OpenAI on your data', the responses are generated by querying an Azure Cognitive Search index that contains your data. If the underlying data source has been updated but the responses still include outdated information, the most likely cause is that the index in Azure Cognitive Search has not been refreshed to reflect those updates. The model itself does not store or cache the data; it relies on the index at query time, so an outdated index directly leads to outdated responses.

Exam trap

The trap here is that candidates may confuse the model's lack of awareness of data updates with caching mechanisms (like CDN or prompt caching), when the real issue is the decoupled indexing pipeline in Azure Cognitive Search that requires explicit refresh.

How to eliminate wrong answers

Option A is wrong because the model does not cache the prompt; each request is processed independently, and caching would not cause outdated information from the data source—it would only affect repeated identical prompts. Option C is wrong because while a sync schedule could cause delays, the question states the data source 'has been updated' and the responses are outdated, implying the index is not refreshed regardless of schedule; the default sync behavior is not the core issue. Option D is wrong because Azure CDN is used for static content delivery, not for caching Azure OpenAI responses, which are dynamic and not routed through CDN.

Full explanation →

435

MCQmedium

Your organization is migrating on-premises machine learning models to Azure. The models are used for real-time inference. You need to choose a service that provides managed endpoints with autoscaling and supports custom containers. Which service should you use?

A.Azure Machine Learning managed online endpoints

B.Azure Functions

C.Azure Kubernetes Service (AKS) with manual scaling

D.Azure AI Services custom vision

AnswerA

Supports custom containers, autoscaling, and managed infrastructure.

Why this answer

Azure Machine Learning managed online endpoints are the correct choice because they provide fully managed, autoscaling endpoints specifically designed for real-time inference. They support custom container images, allowing you to deploy any model packaged as a Docker container, and handle traffic splitting, health checks, and scaling automatically without managing underlying infrastructure.

Exam trap

The trap here is that candidates often confuse Azure Kubernetes Service (AKS) as the only option for custom containers, overlooking that Azure Machine Learning managed endpoints natively support custom containers with autoscaling, eliminating the operational burden of managing a Kubernetes cluster.

How to eliminate wrong answers

Option B (Azure Functions) is wrong because Azure Functions is a serverless compute service for event-driven, short-lived tasks, not optimized for real-time ML inference with custom containers; it lacks native autoscaling for ML workloads and does not provide managed endpoints with traffic splitting or model versioning. Option C (Azure Kubernetes Service with manual scaling) is wrong because while AKS can host custom containers, the requirement specifies 'managed endpoints with autoscaling'—manual scaling contradicts autoscaling, and AKS requires significant cluster management overhead, unlike the fully managed endpoint service. Option D (Azure AI Services custom vision) is wrong because Custom Vision is a pre-built AI service for image classification and object detection, not a general-purpose platform for deploying custom ML models with custom containers; it does not support arbitrary custom containers or managed endpoints for real-time inference.

Full explanation →

436

MCQmedium

A company deploys a custom question answering project in Azure AI Language. Users report that the bot sometimes returns irrelevant answers. The knowledge base contains hundreds of QnA pairs. You need to improve answer relevance without retraining the model. What should you do?

A.Increase the confidence score threshold in the project settings.

B.Reduce the number of QnA pairs to decrease ambiguity.

C.Add alternate phrases to existing QnA pairs for common user queries.

D.Enable active learning to let the bot suggest new questions based on user queries.

AnswerC

This improves matching without retraining.

Why this answer

Adding alternate phrases to existing QnA pairs directly improves the bot's ability to match user queries to the correct answer without retraining. In Azure AI Language's custom question answering, the model uses a ranker that compares the user's input against the questions and alternate phrases in the knowledge base. By providing more varied phrasings for common queries, you increase the likelihood of a high-confidence match, thereby reducing irrelevant answers.

Exam trap

The trap here is that candidates often confuse 'active learning' (a feature for suggesting new questions) with a direct method to improve current answer relevance, when in fact it is a long-term knowledge base enhancement tool that does not immediately affect matching accuracy.

How to eliminate wrong answers

Option A is wrong because increasing the confidence score threshold would filter out more low-confidence matches, potentially causing the bot to return no answer or default responses for valid queries, rather than improving relevance of the answers it does return. Option B is wrong because reducing the number of QnA pairs would shrink the knowledge base and could remove valid answers, increasing the chance of irrelevant or no matches for legitimate user questions. Option D is wrong because enabling active learning is a feature that suggests new questions to add to the knowledge base based on user queries, but it does not directly improve answer relevance for existing pairs; it requires manual review and addition of those suggestions, and does not affect the current matching behavior.

Full explanation →

437

MCQhard

Your knowledge mining solution uses Azure AI Search with a custom skill that calls an Azure Function to perform complex data validation. The custom skill returns an error for some documents, but the indexer continues without raising an error. What is the most likely cause?

A.The indexer is configured with 'allowSkillsetToExecuteIfError' set to false.

B.The indexer is configured with 'allowSkillsetToExecuteIfError' set to true.

C.The Azure Function returns HTTP 200 with an error message in the body.

D.The custom skill's output field mapping is incorrect.

AnswerB

When true, skill errors are treated as warnings and the indexer continues processing other documents.

Why this answer

By default, indexers continue on error. To stop on skill errors, set 'allowSkillsetToExecuteIfError' to false. The indexer treats skill errors as warnings by default.

Full explanation →

438

MCQeasy

You plan to use Azure AI Search to index a large number of text documents stored in Azure Blob Storage. The documents are in English. You want to automatically extract key phrases from the content during indexing. What should you add to the skillset?

A.Key Phrase Extraction skill

B.Sentiment skill

C.Language Detection skill

D.Entity Recognition skill

AnswerA

This skill extracts key phrases from text.

Why this answer

The Key Phrase Extraction skill is the correct choice because it is specifically designed to identify and extract important phrases from text content, which aligns with the requirement to automatically extract key phrases during indexing. This skill is part of Azure AI Search's cognitive skillset and operates on English text to produce a list of key phrases per document.

Exam trap

The trap here is that candidates may confuse Entity Recognition (which extracts specific named entities) with Key Phrase Extraction (which extracts general important phrases), leading them to select the wrong skill for the requirement.

How to eliminate wrong answers

Option B is wrong because the Sentiment skill is used to determine the emotional tone (positive, negative, neutral) of text, not to extract key phrases. Option C is wrong because the Language Detection skill identifies the language of the text (e.g., English, Spanish), but does not extract key phrases from the content. Option D is wrong because the Entity Recognition skill extracts named entities such as people, organizations, and locations, not general key phrases.

Full explanation →

439

MCQeasy

You are building a generative AI solution using Azure OpenAI Service. You want to deploy the model to a specific Azure region to minimize latency for users in Europe. Which deployment parameter must you set?

A.The model version (e.g., 0613).

B.The temperature parameter.

C.The Azure region of the Azure OpenAI resource.

D.The deployment name.

AnswerC

Region determines physical location.

Why this answer

The Azure region of the Azure OpenAI resource directly determines the physical data center location where the model is deployed. By selecting a region in or near Europe (e.g., France Central, Sweden Central, UK South), you minimize network round-trip latency for European users. The region is set at the resource creation level and cannot be changed after deployment.

Exam trap

The trap here is that candidates confuse deployment parameters (region, capacity, model version) with inference parameters (temperature, max tokens), leading them to select temperature or model version as the answer for minimizing latency.

How to eliminate wrong answers

Option A is wrong because the model version (e.g., 0613) specifies which iteration of the GPT model to use, not the geographic deployment location; it affects model behavior and capabilities, not latency. Option B is wrong because the temperature parameter is a runtime inference setting that controls randomness in output generation, not a deployment parameter that influences where the model runs. Option D is wrong because the deployment name is a user-defined label for the model endpoint (e.g., 'gpt-35-turbo-deployment') and has no impact on the Azure region or network latency.

Full explanation →

440

MCQmedium

You are building a solution to extract key information from scanned invoices. The invoices are in PDF format and contain both printed and handwritten fields. Which Azure AI service should you use?

A.Language Service

B.Speech Service

C.Computer Vision

D.Azure AI Document Intelligence (formerly Form Recognizer)

AnswerD

Azure AI Document Intelligence (formerly Form Recognizer) is the correct choice because it is built for extracting information from forms and documents, including printed and handwritten text, and provides prebuilt models for invoices.

Why this answer

Azure AI Document Intelligence (formerly Form Recognizer) is the correct service because it is specifically designed to extract text, key-value pairs, and tables from documents like scanned invoices, and it supports both printed and handwritten text. Option A (Language Service) is used for text analytics and NLP, not structured document extraction. Option B (Speech Service) handles audio transcription and speech, not documents.

Option C (Computer Vision) can extract text via OCR but is less specialized than Document Intelligence for invoice-specific extraction.

Full explanation →

441

MCQhard

An image was submitted to Azure Content Moderator's Image Moderation API. The application uses a threshold of 0.5 for Category1 (adult) to trigger a review. Based on the exhibit, what should the application do with this image?

A.Auto-approve the image because ReviewRecommended is false.

B.Escalate the image to a human reviewer because Category1 score exceeds the threshold.

C.Auto-reject the image because no offensive terms were found.

D.Log the inconsistency and continue processing without action.

AnswerB

The high score warrants a review despite the ReviewRecommended flag being false.

Why this answer

The application's threshold for Category1 (adult) is 0.5, and the exhibit shows a Category1 score of 0.6, which exceeds this threshold. Even though ReviewRecommended is false, the application's custom threshold overrides the default recommendation, requiring escalation to a human reviewer for manual judgment. This aligns with Azure Content Moderator's design where you can set your own thresholds to trigger reviews regardless of the API's built-in recommendation.

Exam trap

The trap here is that candidates assume ReviewRecommended is the sole determinant for action, ignoring that custom thresholds defined in application logic can override the API's default recommendation, leading to incorrect auto-approval.

How to eliminate wrong answers

Option A is wrong because ReviewRecommended is false only indicates the API's default recommendation, but the application's custom threshold of 0.5 is exceeded, so auto-approval would bypass the required human review. Option C is wrong because auto-rejection based on offensive terms is irrelevant; the issue is adult content scoring, not text-based offensive terms, and the Image Moderation API does not use term-based filtering for images. Option D is wrong because logging and continuing without action ignores the explicit threshold violation, which demands a defined action (escalation) rather than passive logging.

Full explanation →

442

Multi-Selecteasy

You are building a generative AI application using Azure OpenAI Service. You need to ensure that the application handles user data securely. Which TWO practices should you implement?

Select 2 answers

A.Disable data logging in the Azure OpenAI Service resource

B.Store prompts and completions in Azure Blob Storage

C.Use HTTPS for all API calls

D.Use system messages to instruct the model not to store data

E.Configure content filtering to block sensitive data

AnswersA, C

Disabling logging prevents storage of user data.

Why this answer

Disabling data logging in Azure OpenAI Service ensures that prompts and completions are not stored by the service for monitoring or abuse detection. This is a key security practice for handling sensitive user data, as it prevents inadvertent retention of confidential information. By default, Azure OpenAI may log data for operational purposes, so explicitly disabling logging is necessary to meet data privacy requirements.

Exam trap

The trap here is that candidates confuse content filtering (which blocks harmful outputs) with data security controls (which prevent data retention), leading them to select Option E instead of recognizing that disabling logging is the direct method to stop data storage.

Full explanation →

443

MCQhard

You are reviewing the configuration of an agent deployed via Azure AI Agent Service, as shown. The agent fails to authenticate when calling the reset_password function, which requires a token from Microsoft Entra ID. What is the most likely issue?

A.The conversation_starters are preventing the agent from calling functions

B.The agent's authentication is set to api_key instead of managed identity or OAuth

C.The reset_password function is missing required parameters

D.The model version is outdated and does not support function calling

AnswerB

API key cannot authenticate to Microsoft Entra ID.

Why this answer

The agent's authentication is set to api_key instead of managed identity or OAuth. Azure AI Agent Service supports multiple authentication methods, but when calling a function that requires a token from Microsoft Entra ID (e.g., reset_password), the agent must use either a managed identity (for Azure resources) or OAuth 2.0 client credentials flow to obtain a valid token. Using an api_key does not provide the necessary Entra ID token, causing authentication failures for the function call.

Exam trap

The trap here is that candidates may assume authentication failures are always due to missing permissions or incorrect function parameters, rather than recognizing that the authentication method itself (api_key vs. managed identity/OAuth) is the root cause when Entra ID tokens are required.

How to eliminate wrong answers

Option A is wrong because conversation_starters are merely initial prompts or suggestions for the user and do not affect the agent's ability to call functions or authenticate. Option C is wrong because the question states the agent fails to authenticate, not that the function call fails due to missing parameters; missing parameters would result in a different error (e.g., validation error), not an authentication failure. Option D is wrong because model version does not impact authentication mechanisms; function calling is supported across modern models, and the issue is specifically about token acquisition from Entra ID, not model capability.

Full explanation →

444

MCQeasy

Your organization has a large repository of technical manuals in PDF format. You need to build a chatbot that can answer questions about the content of these manuals. Which combination of Azure services should you use?

A.Azure AI Search and Azure OpenAI

B.Azure AI Speech and Azure OpenAI

C.Azure AI Language and Azure AI Document Intelligence

D.Azure AI Document Intelligence and Azure Bot Service

AnswerA

Search indexes the manuals; Azure OpenAI provides conversational Q&A (RAG pattern).

Why this answer

Azure AI Search provides the indexing and retrieval capabilities needed to search through the PDF content, while Azure OpenAI (specifically GPT models) can generate natural language answers based on the retrieved passages. This combination enables a RAG (Retrieval-Augmented Generation) pattern where the search engine finds relevant text chunks from the manuals and the language model formulates a coherent answer.

Exam trap

The trap here is that candidates often confuse Azure AI Language (which handles text analytics) with the search and generative AI capabilities needed for a question-answering system, or they incorrectly assume that Azure Bot Service alone can handle document-based Q&A without a search backend.

How to eliminate wrong answers

Option B is wrong because Azure AI Speech is used for speech-to-text and text-to-speech, not for searching or understanding document content; it does not index PDFs or retrieve relevant passages. Option C is wrong because Azure AI Language provides pre-built NLP capabilities like entity recognition or sentiment analysis, but it is not designed for full-text search over a large repository of PDFs; Azure AI Document Intelligence is for extracting text from documents, not for answering questions. Option D is wrong because Azure AI Document Intelligence extracts text from PDFs but does not index or search that text, and Azure Bot Service is a framework for building chatbots but lacks the search and generative AI components needed to answer questions from a document repository.

Full explanation →

445

MCQhard

Refer to the exhibit. A developer is debugging an Azure web app that uses Azure AD authentication. The app frequently fails to authenticate users. What is the most likely cause of the error?

A.The JWT token has expired.

B.The web app is not registered in Azure AD.

C.The web app is using an incorrect client secret.

D.The web app's outbound traffic is blocked by a firewall or VNet restriction.

AnswerD

The error shows failure to obtain configuration from the Azure AD endpoint, which requires outbound HTTPS access.

Why this answer

The error indicates that the web app cannot reach the Azure AD metadata endpoint, likely due to a network restriction or firewall blocking outbound HTTPS traffic.

Full explanation →

446

MCQhard

A company is using Azure Cognitive Service for Language to analyze customer support transcripts. They want to identify custom categories (e.g., 'billing', 'technical support') using a custom text classification model. After training and deploying the model, they receive many false positives for the 'billing' category. What is the best first step to improve model accuracy?

A.Add more training data to all categories to improve overall model performance.

B.Use a different Azure AI service, such as key phrase extraction, to identify billing-related content.

C.Review the training data for the 'billing' category and correct any mislabeled examples.

D.Increase the confidence threshold for the 'billing' category to reduce false positives.

AnswerC

Correcting mislabeled examples improves the model's ability to distinguish categories.

Why this answer

False positives for a specific category like 'billing' most often stem from mislabeled or ambiguous training examples in that category. By reviewing and correcting the training data for 'billing', you directly address the root cause of the model's confusion, which is the most effective first step in custom text classification model improvement.

Exam trap

The trap here is that candidates often jump to a threshold adjustment (Option D) as a quick fix, but Azure's custom text classification models require data quality improvements first, as confidence thresholds only affect prediction output, not model accuracy.

How to eliminate wrong answers

Option A is wrong because adding more training data to all categories indiscriminately does not target the specific false-positive issue with 'billing' and could even introduce more noise or imbalance. Option B is wrong because key phrase extraction is an unrelated Azure AI service that extracts terms, not a classification model; it cannot replace or fix a custom text classification model's accuracy. Option D is wrong because increasing the confidence threshold only filters out low-confidence predictions but does not correct the underlying misclassification pattern; it may reduce false positives at the cost of increasing false negatives, without improving model understanding.

Full explanation →

447

MCQhard

A manufacturing company uses Azure AI Custom Vision to detect defects on a production line. The model was trained with 500 images per class and achieves 95% accuracy. After deployment, the model's accuracy drops to 80% due to changes in lighting conditions. What is the most effective first step to improve the model's robustness?

A.Reduce the probability threshold to increase recall.

B.Capture additional images under the new lighting and retrain the model.

C.Use Azure AutoML to automatically find the best algorithm.

D.Add more images from the original lighting conditions to the training set.

AnswerB

Adding representative data from the new conditions is the best practice.

Why this answer

The drop in accuracy is caused by a domain shift—specifically, new lighting conditions that were not represented in the original training set. The most effective first step is to capture additional images under the new lighting and retrain the model, as Custom Vision relies on diverse, representative training data to generalize to real-world variations. This directly addresses the root cause by expanding the training distribution to include the new lighting scenario, which is a fundamental principle of supervised learning in computer vision.

Exam trap

The trap here is that candidates may confuse a performance tuning action (like adjusting the probability threshold) with a data quality fix, or assume AutoML can magically fix any accuracy drop, when in fact the root cause is a classic domain shift that requires representative retraining data.

How to eliminate wrong answers

Option A is wrong because reducing the probability threshold increases recall but also increases false positives, which does not improve robustness to lighting changes—it only trades precision for recall without addressing the underlying distribution shift. Option C is wrong because Azure AutoML is designed for automated model selection and hyperparameter tuning, but the problem here is a data distribution mismatch, not a need for a different algorithm; AutoML cannot compensate for missing lighting variations in the training data. Option D is wrong because adding more images from the original lighting conditions does not help the model learn to handle the new lighting; it only reinforces the existing bias toward the old lighting, leaving the domain shift unaddressed.

Full explanation →

448

MCQeasy

Your company uses Azure OpenAI Service to generate marketing content. You need to ensure that the generated content does not contain offensive language. Which feature should you enable?

A.Azure AI Content Safety filters.

B.Audit logging for all API calls.

C.Data encryption at rest.

D.Rate limiting on the endpoint.

AnswerA

Content Safety filters can block offensive language.

Why this answer

Azure AI Content Safety filters are specifically designed to detect and block offensive, inappropriate, or harmful language in text and images. By enabling these filters on your Azure OpenAI Service deployment, you can configure severity thresholds for categories like hate, self-harm, sexual, and violence content, ensuring generated marketing content meets safety policies.

Exam trap

The trap here is that candidates confuse operational features like logging or rate limiting with content moderation, assuming any security-related setting can filter offensive language, when only Azure AI Content Safety provides the specific content filtering capability.

How to eliminate wrong answers

Option B is wrong because audit logging records API calls for monitoring and compliance but does not actively filter or block offensive content in responses. Option C is wrong because data encryption at rest protects stored data from unauthorized access but has no role in analyzing or moderating generated text for offensive language. Option D is wrong because rate limiting controls the number of requests per time period to prevent abuse or overload, not to inspect or filter the content of responses.

Full explanation →

449

MCQmedium

Refer to the exhibit. A developer is testing the Text Analytics sentiment analysis API and receives a 401 error. What is the most likely cause?

A.The API version is not supported.

B.The endpoint URL is incorrect.

C.The resource group is misspelled.

D.The subscription key is invalid or expired.

AnswerD

The error message clearly states invalid subscription key, and the curl command uses 'wrongkey'.

Why this answer

The curl command uses 'wrongkey' instead of one of the valid keys shown in the output.

Full explanation →

450

MCQhard

A retail company uses Azure AI Vision to analyze store shelf images for product availability. The solution uses an object detection model trained on custom products. Recently, the model's performance dropped significantly due to new packaging designs. You need to improve the model's accuracy with minimal manual effort. What should you do?

A.Use Azure AI Vision model customization with active learning

B.Adjust the confidence score threshold to reduce false negatives

C.Use the Image Analysis 4.0 dense captioning feature

D.Collect a new set of images and retrain the model from scratch

AnswerA

Active learning selects the most informative images for labeling, minimizing manual effort.

Why this answer

Azure AI Vision model customization with active learning is the correct choice because it allows the model to automatically identify images where it is uncertain (low confidence predictions) and prioritize those for labeling and retraining. This minimizes manual effort while directly addressing the performance drop caused by new packaging designs, as the model iteratively improves on the specific data distribution shift without requiring a full retraining from scratch.

Exam trap

The trap here is that candidates often assume retraining from scratch (Option D) is the only way to fix model drift, underestimating the power of active learning to efficiently handle distribution shifts with minimal manual effort.

How to eliminate wrong answers

Option B is wrong because adjusting the confidence score threshold only changes the trade-off between precision and recall; it does not improve the model's underlying ability to recognize new packaging, and lowering the threshold would increase false positives without fixing the root cause. Option C is wrong because Image Analysis 4.0 dense captioning generates descriptive captions for regions of an image, but it is not designed for object detection or model retraining to adapt to new visual features like packaging changes. Option D is wrong because collecting a new set of images and retraining from scratch requires significant manual effort (data collection, labeling, and training) and is not minimal compared to the iterative, semi-automated approach of active learning.

Full explanation →

Microsoft Azure AI Engineer Associate AI-102 (AI-102) — Questions 376–450