AI-102 Exam Questions and Answers

You are building an agentic solution using Microsoft Semantic Kernel. The agent must autonomously decide when to call an external API to fetch real-time data. You want to minimize token usage and avoid unnecessary API calls. Which planner configuration should you use?

Use a SequentialPlanner with a stepwise strategy and explicit function parameter constraints

Stepwise strategy ensures API is called only when needed.

Use a ManualInvoke kernel with a sequential planner

Use a ParallelPlanner with a function-calling model

Use an AutoInvoke kernel with a greedy action planner

Why: Option D is correct because a SequentialPlanner combined with a stepwise strategy ensures the agent only invokes the API when the context explicitly requires it, reducing token waste. Option A is incorrect because AutoInvoke with greedy planning may over-call APIs. Option B is incorrect because ManualInvoke requires user intervention, defeating autonomy. Option C is incorrect because ParallelPlanner may invoke multiple APIs concurrently, increasing token usage.

You are implementing an agentic solution using Azure AI Agent Service. The agent needs to maintain conversation context across multiple turns. You configure the agent with a custom prompt that includes a 'system message' and 'few-shot examples'. However, after a few turns, the agent starts repeating the same responses. What is the most likely cause?

The 'max_tokens' parameter is set too low for the response

The 'max_context_length' is set too low, causing earlier turns to be truncated

Truncation loses context, leading to repetition.

The agent is using a 'context recycling' feature that resets after each turn

The temperature is set too high, causing the model to become deterministic

Why: Option B is correct because a low maximum context length truncates earlier conversation turns, causing the agent to lose context and repeat responses. Option A is incorrect because the token limit per request affects the length of individual responses, not repetition. Option C is incorrect because the Azure OpenAI API does not recycle context unless explicitly configured. Option D is incorrect because changing temperature affects randomness, not repetition due to context loss.

Your organization is deploying an agentic solution using Microsoft Copilot Studio. The agent must be able to escalate to a human agent when it cannot resolve a user's request. You need to ensure that the escalation includes the full conversation history. What should you configure?

Add an 'End conversation' action and configure a fallback

Add a 'Create a ticket' action in the topic

Add a 'Start a new topic' action with context variables

Add a 'Transfer conversation' action and set it to include the full transcript

Transfer conversation sends history to human agent.

Why: Option C is correct because 'Transfer conversation' automatically passes the full transcript to the human agent. Option A is incorrect because 'Create a ticket' does not transfer the conversation. Option B is incorrect because 'End conversation' terminates without escalation. Option D is incorrect because 'Start a new topic' does not include history.

You are designing an agentic solution using Azure AI Agent Service with a custom skill that calls an external REST API. The API has rate limits: 100 requests per minute per client. You need to ensure the agent respects this limit without degrading user experience. Which approach should you take?

Use a token bucket rate limiter with a shared counter stored in Azure Cache for Redis

Token bucket allows burst and respects limits globally.

Set a fixed delay of 600ms between each API call

Configure the skill to handle HTTP 429 responses with retry-after logic

Implement exponential backoff in the skill code

Why: Option D is correct because a token bucket with a shared counter across all agent instances ensures global rate limiting without dropping requests unexpectedly. Option A is incorrect because 'retry-after' headers require the agent to handle 429 responses, which can degrade UX. Option B is incorrect because exponential backoff can cause delays. Option C is incorrect because a fixed delay per request does not adapt to varying traffic.

You are building an agentic solution using Microsoft Semantic Kernel. The agent needs to orchestrate multiple plugins. One plugin returns a large dataset that exceeds the model's context window. What is the best way to handle this?

Split the data into multiple smaller API calls and combine results

Truncate the data to fit the context window

Configure the plugin to return only a subset of the data and mark the rest as sensitive

Use a summarization plugin to condense the data before passing it to the model

Summarization preserves key info and reduces size.

Why: Option B is correct because summarizing the data reduces its size while preserving key information. Option A is incorrect because truncation may lose critical data. Option C is incorrect because splitting into multiple calls increases latency and tokens. Option D is incorrect because marking as sensitive does not solve the size issue.

You are using Microsoft Copilot Studio to create an agent that helps users reset their passwords. The agent should first verify the user's identity using multi-factor authentication (MFA) before proceeding. Which feature should you configure?

Add a variable to store the user's identity status

Configure Authentication settings to require Microsoft Entra ID authentication with MFA policy

Authentication settings enforce sign-in with MFA.

Add a Power Automate flow that calls Microsoft Entra ID MFA

Use a 'Sign in' topic trigger from the customer channel

Why: Option C is correct because 'Authentication' in Copilot Studio allows you to require Microsoft Entra ID sign-in with MFA. Option A is incorrect because 'Power Automate flow' can call MFA but is not the primary configuration. Option B is incorrect because 'Variable' does not enforce authentication. Option D is incorrect because 'Topic trigger' initiates the topic but does not enforce authentication.

Want more Implement an agentic solution practice?

All Implement computer vision solutions questions

Domain 2: Implement computer vision solutions

A retail company uses Azure Computer Vision to analyze customer traffic in stores. They deploy a custom object detection model to count customers and detect occupancy. After deployment, the model consistently underestimates the number of customers during peak hours. The company has retrained the model with more data but the issue persists. What is the most likely cause?

The model is not being batch-processed for inference.

The training data does not adequately represent peak-hour scenarios.

Data drift or lack of representative samples for peak hours leads to underestimation during those times.

The model is overfitting to the training data.

The Computer Vision API version is outdated.

Why: The model consistently underestimates customer counts during peak hours, which indicates a distribution shift between the training data and the inference environment. Even after retraining with more data, the issue persists because the additional data likely still lacks sufficient representation of peak-hour scenarios (e.g., high density, occlusion, rapid movement). In Azure Custom Vision, object detection models learn from labeled examples; if the training set does not include diverse peak-hour images with varied lighting, crowd densities, and angles, the model will fail to generalize to those conditions.

A hospital uses Azure Custom Vision to classify X-ray images as normal or abnormal. The model achieves 98% accuracy on the test set. However, during deployment, the model misclassifies many abnormal cases as normal, causing missed diagnoses. The hospital has a class imbalance where abnormal cases are only 5% of the data. What should the data scientist do first to address this?

Increase the number of training epochs.

Add more normal X-ray images to the dataset.

Switch to a different object detection algorithm.

Use oversampling or class-weight techniques to balance the training.

Balancing the dataset or adjusting loss weights improves minority class recall.

Why: Option D is correct because the primary issue is class imbalance, where abnormal cases constitute only 5% of the data. Oversampling (e.g., SMOTE) or class-weight techniques adjust the training process to give more importance to the minority class, directly addressing the model's bias toward the majority class and reducing false negatives. This is a standard preprocessing step in Custom Vision and other ML frameworks before tuning hyperparameters or changing algorithms.

A company uses Azure Face API to verify employee identities for building access. They need to ensure that only live faces are used, not photos or videos. Which feature should they enable?

Set a high confidence threshold for face matching.

Face identification with a large person group.

Enable liveness detection using session-based verification.

Liveness detection checks for spoofing attacks.

Face detection with attributes such as age and emotion.

Why: Option C is correct because Azure Face API's liveness detection with session-based verification is specifically designed to prevent spoofing attacks using photos, videos, or masks. It analyzes subtle cues such as micro-movements, texture, and depth to confirm the presence of a live person, ensuring that only live faces are accepted for identity verification.

A developer is building an application to extract text from scanned invoices using Azure Computer Vision's Read API. The invoices contain a mix of printed and handwritten text. The developer needs to ensure the highest accuracy for both types. Which parameter should they set in the API call?

Set the 'language' parameter to 'en' for English handwriting.

No special parameter; the Read API automatically handles both.

Read API OCR works on both printed and handwritten text without additional parameters.

Specify the 'model-version' as '2022-04-30'

Use the 'mode' parameter set to 'Handwriting'

Why: The Read API in Azure Computer Vision is designed to extract text from images and documents, and it automatically handles both printed and handwritten text without requiring any special parameter. Setting the 'language' parameter to 'en' is optional and only improves accuracy for language-specific text, but it does not enable or disable handwriting recognition. Therefore, no additional parameter is needed to achieve the highest accuracy for both types.

A company uses Azure Custom Vision to build a classifier for defect detection on a manufacturing line. They have labeled images of products with and without defects. Which TWO actions should they take to improve model performance?

Train for more iterations without validation.

Use images with balanced numbers of defect and non-defect samples.

Balanced datasets prevent bias toward majority class.

Set the learning rate manually using the Custom Vision API.

Increase the number of images per tag, including variations in lighting and angle.

More diverse data improves robustness.

Reduce the number of images per tag to avoid overfitting.

Why: Option B is correct because balanced datasets prevent the model from becoming biased toward the majority class (e.g., non-defect images), which is critical for defect detection where defects are rare. Azure Custom Vision uses a weighted loss function during training, and class imbalance can cause the model to predict the majority class for most inputs, reducing recall for defects. Balanced samples ensure the model learns discriminative features for both classes equally.

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

Collect more labeled scans from the new hospital and train a new model from scratch.

Create a new Custom Vision project and train only on the 500 new scans.

Apply image preprocessing to normalize the new hospital's scans to match the old hospital's style, then use the existing model.

Use the existing model as a starting point and retrain it with the 500 labeled scans from the new hospital.

Transfer learning with new data quickly adapts the model to the new domain with minimal effort.

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

Want more Implement computer vision solutions practice?

All Implement knowledge mining and information extraction solutions questions

Domain 3: Implement knowledge mining and information extraction solutions

You are building a knowledge mining solution to extract insights from a large set of PDF contracts. The solution must identify parties, dates, and monetary amounts. Which Azure AI service should you use as the primary extraction engine?

Azure AI Language (custom NER)

Azure AI Search with integrated vectorization

Azure OpenAI Service with GPT-4o

Azure AI Document Intelligence

Designed for extracting fields from forms and documents.

Why: Option C is correct because Document Intelligence (formerly Form Recognizer) is specialized in extracting structured fields from documents. Option A is wrong because Azure AI Search is for indexing and searching, not extraction. Option B is wrong because Azure OpenAI can extract entities but is not the most cost-effective for this specific scenario. Option D is wrong because AI Language is for text analytics but not optimized for document layout analysis.

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

The indexer encountered errors and marked the documents as failed

Indexer errors prevent documents from being indexed.

The index does not have a semantic configuration

The search service has insufficient replicas

The search service is throttled due to high query volume

Why: Option B is correct because if the indexer fails, documents are not indexed. Option A is wrong because throttling would affect all documents, not a subset. Option C is wrong because the search service would report failures for other reasons. Option D is wrong because missing semantic configuration affects ranking, not indexing.

You are designing a knowledge mining solution that must extract entities from scanned handwritten forms. The forms contain signatures and checkboxes. Which combination of Azure AI services should you recommend?

Azure AI Document Intelligence with a custom neural model and Azure AI Language for entity linking

Custom neural models support handwriting; Language can enrich entities.

Azure AI Document Intelligence with a premade model and Azure AI Computer Vision

Azure AI Computer Vision (OCR) and Azure AI Search with integrated vectorization

Azure Cognitive Search and Azure AI Document Intelligence with a premade model

Why: Option A is correct because Document Intelligence can extract handwriting and layout, and AI Language can post-process entities. Option B is wrong because Computer Vision OCR is for printed text only. Option C is wrong because Cognitive Search is not an extraction service. Option D is wrong because AI Document Intelligence already includes OCR; adding Computer Vision is redundant.

Your knowledge mining solution uses Azure AI Search. Users complain that search results are not relevant. You have enabled semantic search but results still lack context. What should you do to improve relevance?

Ensure the index includes a semantic configuration with title and content fields

Semantic configuration is required for semantic ranking to work.

Increase the number of partitions to handle more data

Configure a scoring profile with boosting based on metadata

Increase the number of replicas to improve query performance

Why: Option D is correct because semantic ranking uses captions and answers; without them, it is less effective. Option A is wrong because simple scoring profiles do not use AI. Option B is wrong because more replicas improve throughput, not relevance. Option C is wrong because increasing partition count improves indexing speed.

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

Custom extraction model

Layout model

Read model

Premade invoice model

Built specifically for invoices.

Why: Option B is correct because the premade invoice model is designed for common invoice layouts. Option A is wrong because the layout model extracts text and tables, not key-value pairs. Option C is wrong because custom extraction requires training data. Option D is wrong because the read model only extracts text.

Your knowledge mining solution ingests documents from multiple tenants. Each tenant's data must be isolated and searchable only by that tenant. You have a single Azure AI Search service. How should you implement multi-tenancy?

Use separate skillsets for each tenant

Create a separate search service for each tenant

Use a single index with a tenant ID field and filter queries by that field

Index-level security with filters is the recommended approach.

Use separate data sources within the same index

Why: Option C is correct because index-level security with filters is the recommended pattern for multi-tenancy. Option A is wrong because multiple services are costly and unnecessary. Option B is wrong because data sources are not security boundaries. Option D is wrong because skillsets are stateless and do not provide isolation.

Want more Implement knowledge mining and information extraction solutions practice?

All Implement image and video processing solutions questions

Domain 4: Implement image and video processing solutions

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

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

Increase the confidence threshold in the API call.

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

Custom models need retraining with new labeled data.

Recreate the Computer Vision resource in a different region.

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

A healthcare provider uses Azure Video Analyzer for Media to extract insights from surgical videos. They need to ensure that no patient health information (PHI) is stored in the transcriptions. What is the best approach?

Use Azure AI Content Safety to post-process transcriptions.

Disable indexing for videos containing PHI.

Enable content moderation in Video Analyzer for Media settings.

Content moderation can detect and redact PHI.

Use the 'delete' API to remove all transcripts after processing.

Why: Option C is correct because enabling content moderation in Azure Video Analyzer for Media automatically redacts or flags sensitive content, including PHI, from transcriptions during the indexing process. This built-in feature ensures that PHI is not stored in the output without requiring additional post-processing or manual deletion.

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

Use a different classification algorithm.

Add at least 50 more images per class with variations.

More data improves model accuracy.

Reduce the image resolution to speed up training.

Increase the number of training iterations (epochs).

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

A news organization uses Azure Video Indexer to generate transcripts of live broadcasts. They notice that the speaker names are not appearing in the transcript. What is the most likely cause?

The video resolution is too low for OCR.

The speaker identification model has not been trained with voice samples.

Speaker names require custom voice identification.

The video format is not supported.

The language is not set correctly.

Why: Speaker names are missing because Azure Video Indexer's speaker identification feature requires pre-trained voice samples to match speakers to their identities. Without a custom voice model trained on known speakers' audio, the service can only label speakers as 'Speaker #1', 'Speaker #2', etc., but cannot assign actual names. This is a supervised learning process where the model must be trained with labeled voice samples before it can recognize and name speakers.

A company uses Azure Face API to detect faces in a crowd. They need to comply with GDPR and delete face data after 30 days. What should they implement?

Enable data encryption at rest.

Set a retention policy on the Face API resource.

Use the Face API Delete operation to remove stored face IDs.

Explicitly deleting face data meets GDPR requirements.

Recreate the Face API resource every 30 days.

Why: Option C is correct because the Azure Face API stores face data (e.g., face IDs) in a PersonGroup or FaceList, and GDPR compliance requires the ability to delete personal data on demand. The Face API provides a dedicated Delete operation (e.g., FaceList - Delete Face or PersonGroup Person - Delete) to remove stored face IDs and associated face data, enabling the 30-day deletion requirement without recreating the resource.

A developer is building a mobile app that uses Azure Computer Vision to analyze images. The app needs to handle many requests with low latency. Which pricing tier should they choose?

Video Analyzer S1 tier.

Free F0 tier.

Custom Vision S0 tier.

Computer Vision S1 tier.

S1 offers up to 30 calls per second.

Why: The Computer Vision S1 tier is designed for production workloads requiring high throughput and low latency, making it suitable for a mobile app that handles many image analysis requests. Unlike the Free F0 tier, which has strict rate limits (e.g., 20 transactions per minute), the S1 tier offers higher transactions per second (e.g., up to 10 TPS) and guaranteed performance for real-time scenarios.

Want more Implement image and video processing solutions practice?

All Implement natural language processing solutions questions

Domain 5: Implement natural language processing solutions

A company is building a chatbot using Azure Cognitive Service for Language. They need to ensure that user utterances are correctly mapped to the appropriate intent in a custom question answering project. What should they configure?

Add synonyms and phrase list to the LUIS application.

Add alternative question phrases to the QnA pairs in the custom question answering project.

Define entities in the custom question answering project to capture key information.

Add synonyms and phrase list to the custom question answering project.

Synonyms and phrase lists help match varied user utterances to the correct QnA pair.

Why: Option D is correct because adding synonyms and phrase lists to a custom question answering project directly improves the mapping of user utterances to intents by normalizing variations in phrasing. This configuration allows the project to recognize equivalent terms (e.g., 'cost' and 'price') as the same intent, ensuring accurate intent mapping without requiring exact matches.

A development team is using Azure Cognitive Service for Language to extract key phrases from customer reviews. They notice that some reviews are not being processed, and the API returns a 400 error code. What is the most likely cause?

One of the reviews exceeds the maximum character limit for a single document.

The service limits each document to 5,120 characters.

The reviews contain characters that are not valid UTF-8.

The request contains more than 5 documents.

The reviews are written in a language not supported by the service.

Why: The Azure Cognitive Service for Language key phrase extraction API enforces a maximum document size of 5,120 characters per document. When a single review exceeds this limit, the API returns a 400 Bad Request error because the request payload violates the service's input constraints. This is the most common cause of 400 errors in batch text analysis operations.

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?

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

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

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

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

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

Why: Option C is correct because 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.

A company wants to use Azure AI Translator to translate customer emails from English to French. They need to ensure that the translation preserves the tone and formality of the original text. What should they configure in the request?

Set the 'category' parameter to 'general' to use a standard translation model.

Set the 'scope' parameter to 'document' to ensure context-aware translation.

Set the 'formality' parameter to the desired level (e.g., 'formal' or 'informal').

The formality parameter controls the tone of the translation.

Set the 'language' parameter to 'fr' and the 'from' parameter to 'en'.

Why: Option C is correct because Azure AI Translator provides a 'formality' parameter that allows you to specify the desired level of formality (e.g., 'formal' or 'informal') in the translated text. This parameter directly controls the tone and register of the output, ensuring that the translation preserves the original email's tone and formality, which is critical for customer communications.

A development team is using the Azure Cognitive Service for Language to perform sentiment analysis on social media posts. They notice that the returned sentiment scores are often neutral for posts that are clearly positive or negative. What is the most likely reason?

The service does not support sentiment analysis for social media language.

The posts are too short, causing the sentiment detection to default to neutral.

Short texts provide insufficient context for accurate sentiment detection.

The service is not configured to detect mixed sentiment.

The posts are in a language that is not supported by the sentiment analysis API.

Why: The Azure Cognitive Service for Language sentiment analysis API has a minimum text length requirement for reliable scoring. When input text is very short (e.g., a few words or a single sentence), the model lacks sufficient context to confidently assign a positive or negative score, so it defaults to a neutral score (often around 0.5). This is a documented behavior of the API, not a limitation of social media language support.

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

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

Overlapping content confuses the model.

Use a stratified split of training and testing data.

Stratified split maintains class distribution across sets.

Oversample the minority classes to balance the dataset.

Remove all stop words from the training data.

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

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

Want more Implement natural language processing solutions practice?

All Implement generative AI solutions questions

Domain 6: Implement generative AI solutions

A company wants to generate personalized product descriptions for its e-commerce site using Azure OpenAI. They need to ensure the model's output adheres to brand guidelines and does not generate prohibited content. Which approach should they use?

Use a system message with brand guidelines and apply content filtering.

System messages set behavior, content filtering blocks prohibited content.

Use prompt engineering with negative prompts and ignore content filtering.

Provide few-shot examples in the user message and rely on the model's training.

Fine-tune the model with brand guidelines and disable content filtering for performance.

Why: Option A is correct because using a system message allows you to embed brand guidelines directly into the conversation context, instructing the model on tone, style, and prohibited content. Azure OpenAI's content filtering provides an additional safety layer by automatically detecting and blocking harmful or policy-violating outputs, ensuring compliance with both brand and regulatory requirements.

A healthcare startup is developing a chatbot that uses Azure OpenAI to answer patient questions. They need to ensure that the chatbot only uses information from their verified medical database and does not generate unsupported medical advice. What is the best approach?

Fine-tune a model on the medical database and deploy it.

Embed the entire medical database in the system message.

Rely on Azure OpenAI's content filtering to block unsupported advice.

Use Azure AI Search with vector search to retrieve relevant documents and pass them as context.

RAG ensures responses are grounded in indexed data.

Why: Option D is correct because it uses Azure AI Search with vector search to retrieve only relevant, verified documents from the medical database and passes them as context to the Azure OpenAI model. This grounds the model's responses in authoritative data, preventing it from generating unsupported medical advice. The retrieval-augmented generation (RAG) pattern ensures the chatbot answers are based on the provided context rather than the model's internal knowledge.

A developer wants to deploy a custom generative AI model using Azure Machine Learning. Which compute target should they choose for low-latency real-time inference?

Local deployment

Azure Batch

Azure Functions

Azure Kubernetes Service (AKS)

AKS is designed for real-time inference with low latency.

Why: Azure Kubernetes Service (AKS) is the correct compute target for low-latency real-time inference because it supports horizontal pod autoscaling, GPU acceleration, and can be configured with a low-latency ingress controller (e.g., NGINX or Azure Application Gateway) to route inference requests directly to model containers. AKS also integrates with Azure Machine Learning's real-time inference endpoint, which uses a gRPC or HTTP-based scoring protocol to achieve sub-100ms response times.

A company uses Azure OpenAI to generate code snippets. They notice that the model sometimes produces code that uses deprecated APIs. They want to minimize this without retraining the model. What should they do?

Fine-tune the model on a dataset of recent code.

Set the temperature parameter to 0 to reduce randomness.

Add a system message instructing the model to use only current, non-deprecated APIs.

System messages effectively guide model behavior.

Provide a few-shot example of correct code in the prompt.

Why: Option C is correct because adding a system message in Azure OpenAI allows you to set high-level instructions that guide the model's behavior without retraining. By explicitly instructing the model to use only current, non-deprecated APIs, you leverage the system prompt's ability to influence output style and content, effectively reducing deprecated API usage in generated code snippets.

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

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

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

System messages and content filtering directly address content restrictions.

Provide few-shot examples of disclaimers in the prompt.

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

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

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

presence_penalty

frequency_penalty

temperature

Temperature controls randomness and diversity.

max_tokens

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

Want more Implement generative AI solutions practice?

All Implement agentic AI solutions questions

Domain 7: Implement agentic AI solutions

A company is building an agent that uses Azure OpenAI Service to answer customer queries by querying a SQL database. The agent must be able to handle complex multi-turn conversations and maintain context. Which approach should the team use to implement the agent?

Use a single prompt that includes the entire conversation history and the latest user question, then generate SQL.

Use a chain-of-thought prompt to generate SQL queries directly from user input, without maintaining conversation history.

Use embeddings-based retrieval to find relevant past interactions and include them in the prompt.

Use a conversational agent framework like AutoGen with a tool that executes SQL queries, and maintain conversation state.

AutoGen provides multi-turn conversation management and tool integration.

Why: Option D is correct because AutoGen is a conversational agent framework designed for multi-turn, stateful interactions. It can maintain conversation context across turns and integrate a tool to execute SQL queries, which directly meets the requirement for complex multi-turn conversations with context retention. The other options either lack state management or rely on stateless prompt engineering, which is insufficient for maintaining context in a multi-turn agent.

A healthcare company is developing an agent that processes patient records and suggests treatment plans. The agent must comply with HIPAA regulations. Which service should the team use to ensure data privacy and compliance?

Microsoft Bot Framework SDK with a custom connector.

Azure Cognitive Search with custom analyzers.

Azure OpenAI Service with data processing enabled.

Azure AI Services deployed in a private endpoint with no data leaving the network.

Private endpoint ensures data stays within the network, aiding compliance.

Why: Option D is correct because deploying Azure AI Services within a private endpoint, combined with ensuring no data leaves the network, aligns with HIPAA's requirement for data privacy and compliance. This configuration uses Azure Private Link to keep all traffic within the Microsoft backbone network, preventing exposure to the public internet and meeting the strict data residency and encryption standards mandated by HIPAA.

A company is using Azure OpenAI Service to power a customer support agent. The agent sometimes generates incorrect information when it cannot find an answer in the knowledge base. The team wants to ensure the agent only responds using information from the knowledge base and explicitly states when it does not know the answer. Which configuration should the team use?

Use a custom model fine-tuned on the knowledge base and disable content filtering.

Use a system message that says 'If you don't know, say you don't know' and rely on the model's training.

Use 'use your own data' feature with strict content filtering and set the model to only respond based on retrieved documents.

This ensures responses are grounded in the provided data.

Use prompt engineering with a system message that instructs the model to only answer from the knowledge base, with no additional filtering.

Why: Option C is correct because the 'use your own data' feature in Azure OpenAI Service allows you to restrict the model to answer only from the retrieved documents, ensuring responses are grounded in the knowledge base. Strict content filtering further prevents the model from generating unverified information, and the explicit setting to respond based solely on retrieved documents directly addresses the requirement to state when it does not know the answer.

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?

Azure Logic Apps

Azure API Management

Azure Bot Service

Azure Bot Service provides bot hosting and channel integration.

Azure Functions

Why: 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.

A company is developing an agent that uses Azure AI Language to extract entities and intents from user queries. The agent receives a query: 'Book a flight to Paris on Friday.' The agent should extract the intent as 'BookFlight' and entities as 'Paris' (destination) and 'Friday' (date). The team uses a custom entity extraction model. After testing, the model extracts 'Paris' as location but fails to extract 'Friday' as date. What should the team do to fix this?

Increase the training data for location entities.

Add a prebuilt entity component for date.

Train the entity extraction model with more examples of dates.

More training examples improve entity recognition.

Use a different intent classification model.

Why: Option C is correct because the custom entity extraction model in Azure AI Language requires sufficient labeled examples for each custom entity type to learn patterns. Since the model extracts 'Paris' (location) but fails on 'Friday' (date), the issue is specifically with the date entity's training data, not the location. Adding more diverse examples of date expressions (e.g., 'next Monday', 'tomorrow', 'March 5th') will improve the model's ability to recognize 'Friday' as a date entity.

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

Azure Cognitive Search and Azure AI Language.

Azure Bot Service and Azure Logic Apps.

Azure Machine Learning and Azure Functions.

Azure AI Content Safety and Azure OpenAI content filtering.

These services provide comprehensive content safety.

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

Want more Implement agentic AI solutions practice?

All Implement knowledge mining and document intelligence solutions questions

Domain 8: Implement knowledge mining and document intelligence solutions

A company is building a knowledge mining solution using Azure AI Search. They need to extract key phrases from a large set of documents in multiple languages. Which skill should they add to the skillset?

Key Phrase Extraction skill

Key Phrase Extraction is designed to extract key phrases from text.

Sentiment Analysis skill

Language Detection skill

Entity Recognition skill

Why: The Key Phrase Extraction skill is the correct choice because it is specifically designed to identify and extract the most important phrases from text, which directly supports the requirement to extract key phrases from documents. Azure AI Search's built-in Key Phrase Extraction skill leverages natural language processing to analyze text and return a list of key phrases, making it the appropriate skill for this knowledge mining solution.

A healthcare organization uses Azure Document Intelligence to process patient intake forms. They notice that the confidence scores for field extraction are low. What is the most likely cause?

The document resolution is too low

The document layout is not analyzed

The custom model was trained with only 10 labeled forms

Custom models require at least 5 labeled forms; more samples improve confidence.

The batch processing size is too large

Why: Custom models in Azure Document Intelligence require a minimum of five labeled forms for training, but low confidence scores typically indicate insufficient training data. With only 10 labeled forms, the model lacks enough examples to generalize well across variations in handwriting, formatting, and field values, leading to poor extraction confidence.

A company uses Azure Document Intelligence to extract data from invoices. They deploy the model to a container for on-premises processing. After deployment, they notice that the container consumes more memory than expected. What should they do to optimize memory usage?

Set the 'Memory' environment variable to a lower value in the container configuration

The container's memory usage can be controlled via the 'Memory' setting.

Use the 'Read' model instead of the 'Layout' model

Use the cloud API instead of the container

Reduce the batch size in the client application

Why: Option A is correct because Azure Document Intelligence containers expose a 'Memory' environment variable that allows you to limit the container's memory allocation. By setting this variable to a lower value, you constrain the container's memory usage, which directly addresses the issue of higher-than-expected consumption. This is the recommended approach for optimizing memory in containerized deployments.

A company builds a knowledge mining solution using Azure AI Search with a custom skillset that includes an OCR skill. They want to ensure that images embedded in PDFs are processed. What should they configure?

Set the 'defaultLanguageCode' to 'en'

Set the 'textExtractionAlgorithm' to 'printed'

Set the 'imageAction' parameter to 'generateNormalizedImages'

This parameter enables extraction of images from documents.

Set the 'lineEnding' parameter to 'space'

Why: Option C is correct because the 'imageAction' parameter in Azure AI Search's OCR skill controls whether images embedded in documents (including PDFs) are extracted and processed. Setting it to 'generateNormalizedImages' ensures that images within PDFs are normalized and passed to the OCR skill for text extraction, which is essential for processing embedded images.

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?

Label more examples of the specific field in the training set

More labeled examples improve model accuracy for that field.

Increase the batch size in the analysis request

Reduce the image resolution to 200 DPI

Use the prebuilt-tax.us model

Train a custom model using 10 similar forms

Custom models trained on similar forms yield better accuracy.

Why: Option A is correct because 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.

A company uses this skillset in an Azure AI Search enrichment pipeline. They notice that the enrichment pipeline fails when processing a document larger than 5000 characters. What is the most likely cause?

The maximum page length is too small

The default language code is not supported

The text split mode should be 'sentences'

The output field mapping is missing or incorrect

The output 'pages' must be mapped to a collection field in the index.

Why: The enrichment pipeline fails because the output field mapping is missing or incorrect. When a skillset processes documents, the output of each skill must be explicitly mapped to an index field; if this mapping is absent or misconfigured, the pipeline cannot store the enriched data and fails, especially for larger documents that produce more output data.

Want more Implement knowledge mining and document intelligence solutions practice?

All Plan and manage an Azure AI solution questions

Domain 9: Plan and manage an Azure AI solution

A company plans to deploy an Azure AI solution that processes sensitive customer data. The solution must comply with GDPR and ensure data residency within the European Union. Which Azure resource configuration should be used?

Deploy the Azure AI services in multiple regions globally and use geo-replication.

Use the Azure AI services 'Data Residency' SKU.

Use the Free tier of Azure AI services.

Create an Azure AI services resource in a specific EU region and set the data residency option to 'EU'.

Azure AI services allow you to choose a region to control data residency.

Why: Option D is correct because Azure AI services allow you to specify a data residency option when creating the resource in a specific EU region, ensuring that all customer data remains within the European Union as required by GDPR. This configuration explicitly restricts data processing and storage to the chosen geographic boundary, meeting compliance and residency mandates.

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

Azure Advisor

Azure Monitor

Azure Service Health

Azure Cost Management

Azure Cost Management allows you to set budgets and configure alerts.

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

A company is deploying a custom vision model using Azure Custom Vision. The training data contains images with varying resolutions. The model must achieve high accuracy. Which pre-processing step should be applied to the images before training?

Resize all images to the same dimensions (e.g., 224x224).

Custom Vision expects consistent image sizes for optimal performance.

Convert images to grayscale.

Normalize pixel values to a range of 0-1.

Apply data augmentation techniques like random cropping.

Why: Custom Vision models use a fixed input size (e.g., 224x224 for ResNet-based architectures). Images with varying resolutions must be resized to the same dimensions before training to ensure consistent tensor shapes for the neural network. Without this step, the model cannot process the data correctly, leading to training failures or degraded accuracy.

You are designing an Azure AI solution that uses Language Understanding (LUIS) for intent detection. The solution must handle multiple languages dynamically based on the user's locale. What should you do?

Use Azure Translator to translate user input to English before sending to LUIS.

Create separate LUIS applications for each language and route based on locale.

This is the recommended approach for multi-language support.

Train a single LUIS app with utterances in all languages.

Enable the 'Multi-Language' feature in the LUIS app.

Why: Option B is correct because LUIS does not natively support multi-language within a single application; each LUIS app is designed for a single language. To handle multiple languages dynamically, you must create separate LUIS applications for each language and route user utterances based on the detected locale, ensuring accurate intent and entity recognition per language.

A company wants to use Azure AI services to extract text from scanned PDF documents. Which Azure AI service should they use?

Azure AI Language Understanding (LUIS)

Azure AI Document Intelligence

Document Intelligence is optimized for document extraction.

Azure AI Computer Vision API

Azure Cognitive Search

Why: Azure AI Document Intelligence (formerly Form Recognizer) is the correct service because it is specifically designed for extracting text, tables, and key-value pairs from scanned PDFs and images using optical character recognition (OCR) and deep learning models. Unlike general OCR APIs, Document Intelligence can handle complex layouts and preserve document structure, making it ideal for this use case.

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

Use Azure CDN to cache the model responses in Asia.

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

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

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

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

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

Want more Plan and manage an Azure AI solution practice?

All Implement content moderation solutions questions

Domain 10: Implement content moderation solutions

A company uses Azure Content Moderator to review user-generated images in a social media app. Recently, the team noticed that images containing subtle adult content are not being flagged. What should they do to improve detection without increasing false positives?

Increase the moderation thresholds for adult content.

Disable the adult classification tier to allow all images to pass through.

Configure a human review team using the Review tool to manually inspect flagged content.

Human review can catch subtle content that automated systems miss, and it helps reduce false positives by confirming or overturning automated decisions.

Retrain the Content Moderator model with additional labeled images of adult content.

Why: Option C is correct because Azure Content Moderator is designed to work with human review teams via the Review tool to handle edge cases where automated detection fails. By configuring a human review team, flagged images can be manually inspected to catch subtle adult content that the machine learning model misses, without lowering thresholds that would increase false positives. This approach leverages human judgment to improve detection accuracy while maintaining the existing automated moderation settings.

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

Review API

Video Moderation API

Image Moderation API

Text Moderation API

The Text Moderation API screens text for profanity and personally identifiable information (PII).

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

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?

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

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

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

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

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

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

Human review can incorporate regional nuances that automated systems miss.

Why: 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.

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?

Auto-approve the image because ReviewRecommended is false.

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

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

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

Log the inconsistency and continue processing without action.

Why: 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.

You are the AI engineer at a global e-commerce company that allows users to upload product images and descriptions. You use Azure Content Moderator to automatically moderate images for adult and racy content, and text for profanity and personal data. Recently, you noticed that some product descriptions containing profanity in French are not being flagged. Your Content Moderator text moderation API call includes the language parameter set to 'eng'. The profanity list appears to be English-only. You have a requirement to support French and Spanish in addition to English. You also need to ensure that false positives for legitimate product descriptions are minimized. You cannot use a custom term list because the profanity terms are dynamic. What should you do?

Disable the language parameter so that the API defaults to all languages.

Create separate API calls for each language and specify the language code in the request.

Set the language parameter to 'auto-detect' in the text moderation API request.

Auto-detect enables Content Moderator to identify the language and apply the correct profanity detection model, supporting multiple languages dynamically.

Add French and Spanish profanity terms to a custom term list and use the list in the API call.

Why: Option C is correct because setting the language parameter to 'auto-detect' allows the Azure Content Moderator text moderation API to automatically identify the language of the input text and apply the corresponding built-in profanity list (including French and Spanish). This meets the requirement to support multiple languages without using a custom term list, and it minimizes false positives by using the appropriate language-specific moderation model rather than a generic English-only list.

You are developing a content moderation solution that uses Azure Content Moderator to review images uploaded by users. The solution must flag images for potential adult content. Which TWO actions should you take to achieve this goal?

Train a custom image classifier using Content Moderator's training API

Use the List Management API to add images to a blocklist

Configure a human review loop using the Review tool

Human reviews can provide accurate final decisions and improve the moderation system.

Call the Evaluate operation on the image endpoint

The Evaluate operation directly classifies images for adult and racy content.

Use the OCR operation to extract text from images

Why: Option C is correct because Azure Content Moderator's Review tool enables human-in-the-loop review, which is essential for accurately flagging adult content when automated confidence scores are borderline. This allows human moderators to confirm or override the automated classification, ensuring compliance with content policies. Option D is correct because the Evaluate operation on the image endpoint directly analyzes images for adult and racy content using pre-trained models, returning a confidence score that can trigger further action.

Want more Implement content moderation solutions practice?