Free Microsoft Certified: Azure AI Fundamentals (AI-900) Sample Questions

Correct answer: C — Object Detection

• A. Incorrect. Optical Character Recognition (OCR) is a workload designed specifically to detect and extract printed or handwritten text from images. It is not suitable for identifying people or classifying their appearance.
• B. Incorrect. Image Classification assigns a single label to an entire image. While it could classify an image as 'contains person with mask', it cannot handle scenarios with multiple people in the frame or pinpoint which individuals are wearing masks. This makes it unsuitable for a busy store entrance.
• C. Correct. Object Detection is the most appropriate workload. It identifies and localizes individual objects (in this case, people or faces) within an image, typically by drawing a bounding box around them. Each detected object can then be classified (e.g., 'mask' or 'no mask'). This allows the system to accurately assess multiple customers in a single frame.
• D. Incorrect. Semantic Segmentation involves classifying each individual pixel in an image to create a detailed map of different objects. This provides a much higher level of detail than needed for this task and is computationally more expensive and complex than object detection, making it an inefficient choice.

Correct answer: C — Clustering

• A. Incorrect. Regression is a type of supervised learning used to predict a continuous numeric value, such as predicting a user's total watch time. It is not used for grouping users into segments.
• B. Incorrect. Classification is a supervised learning technique that assigns data points to predefined, labeled categories. The scenario explicitly states there are no predefined categories, making classification unsuitable.
• C. Correct. Clustering is an unsupervised learning technique that groups similar data points together based on their characteristics without any predefined labels. This directly matches the scenario of grouping users based on viewing habits to discover natural segments.
• D. Incorrect. Reinforcement learning is a type of machine learning where an agent learns to make decisions by taking actions in an environment to maximize a cumulative reward. It is not used for unsupervised grouping of data.

Correct answer: A — Object Detection

• A. Correct. Object detection is a computer vision task that involves both identifying and locating multiple objects within an image or video feed. It provides the location of each detected object, typically by drawing a bounding box around it. This is precisely the requirement for an autonomous vehicle which must detect and locate items like pedestrians, other cars, and traffic signs.
• B. Incorrect. Image classification assigns one or more labels to an entire image (e.g., 'city street scene', 'nighttime'). It describes the image as a whole but does not provide the specific locations or bounding boxes of individual objects within the image.
• C. Incorrect. Face recognition is a specialized computer vision task that focuses exclusively on identifying or verifying a person from their facial features. It is not used for detecting a diverse range of objects like cars and traffic signs.
• D. Incorrect. Document processing is an AI workload focused on extracting text, structure, and key-value pairs from documents like PDFs, forms, or scanned images. It is unrelated to analyzing real-time video feeds from a vehicle's camera.

Correct answer: C — Humans should be accountable for how AI systems operate.

• A. Incorrect. This statement describes the principle of Reliability and Safety. This principle ensures AI systems are robust and secure, protecting them from malicious attacks, which is distinct from Accountability.
• B. Incorrect. This statement describes the principle of Transparency (also known as Explainability). While the ability of an AI system to explain its reasoning can support accountability, the core principle of accountability is about human responsibility, not the system's inherent explainability.
• C. Correct. The principle of Accountability states that the people who design, develop, and deploy AI systems are ultimately responsible for their actions and impact. It emphasizes that humans must have meaningful oversight and be answerable for how the systems operate.
• D. Incorrect. This statement describes the principle of Fairness. Fairness is focused on ensuring that AI systems treat all people equitably and do not create or reinforce societal biases.

Correct answer: B — Identify itself as an AI system to the user.

• A. Incorrect. This action is related to data collection for a transaction and involves security and privacy considerations, not transparency. The Transparency principle is about being open about the AI's nature and function, not about collecting sensitive user data.
• B. Correct. The principle of Transparency requires that AI systems be understandable. A key part of this is disclosing to users that they are interacting with an AI, not a human. This helps set appropriate expectations about the system's capabilities and limitations.
• C. Incorrect. Requesting a rating is a feedback collection mechanism used to evaluate and improve the service. While valuable, it does not inform the user about the nature of the chatbot, which is the core of the Transparency principle.
• D. Incorrect. This is a security and authentication measure to verify the user's identity. It is concerned with user privacy and system security, not with being transparent about the chatbot's identity as an AI system.

Correct answers: A, B — Implementing content filters to prevent the generation of harmful or violent text; Ensuring the model handles unexpected inputs gracefully without crashing

• A. Correct. This is a direct measure to ensure the safety of an AI system. By implementing content filters, you prevent the model from generating harmful, inappropriate, or violent content, thereby reducing the risk of real-world harm and aligning with the safety aspect of this principle.
• B. Correct. This action directly addresses the reliability of the AI system. A reliable system must be robust and resilient, capable of handling unexpected inputs or edge cases without failing or producing unsafe behavior. Ensuring graceful handling of such inputs is a core practice for building reliable AI.
• C. Incorrect. While documenting data sources is a crucial practice in responsible AI, it primarily supports the principle of Transparency, which is about being open and clear about the system's capabilities, limitations, and data.
• D. Incorrect. Assigning liability is a legal and governance measure that relates to the principle of Accountability. This principle states that people should be accountable for AI systems, but it doesn't directly implement the technical reliability or safety of the model itself.
• E. Incorrect. Making the user interface accessible supports the principle of Inclusiveness, which aims to empower everyone and ensure AI systems do not exclude or discriminate against any groups of people. It is not directly related to the system's operational reliability or safety.

Correct answer: B — False

• A. The statement is false because the solution does not meet the goal of Fairness. Using historical data that contains a known bias, such as predominantly hiring men, without any analysis or mitigation will train the AI system to learn and perpetuate this bias. This directly contradicts the principle of Fairness, which aims to treat all individuals equitably.
• B. The statement is false because training a model on biased historical data without corrective measures will reproduce and amplify gender-based disparities. Ensuring fairness requires proactive steps, such as analyzing data for bias, applying mitigation techniques like rebalancing or using fairness-aware algorithms, and validating outcomes across different demographic groups.

Correct answer: B — Online endpoint

• A. Incorrect. A batch endpoint is used for asynchronous, large-scale scoring of datasets where an immediate response is not required. It processes large volumes of data and returns results after the job is complete, making it unsuitable for an interactive mobile application needing real-time predictions.
• B. Correct. An online endpoint is specifically designed for real-time inference (also known as real-time scoring). It provides a low-latency, synchronous REST API that applications, such as a mobile app, can call to send data for a single request and receive an immediate prediction in the response.
• C. Incorrect. A pipeline endpoint is used to trigger and manage Azure Machine Learning pipelines, which are workflows for automating machine learning tasks like data preparation, training, or batch scoring. It orchestrates processes rather than serving a model for real-time inference.
• D. Incorrect. A private endpoint is a networking feature that provides secure and private connectivity to Azure services over a private network. While you can use a private endpoint to secure an online or batch endpoint, it is not an inference endpoint type itself. Its purpose is network security, not model deployment for scoring.

Correct answer: B — Regression

• A. Incorrect. Classification is used to predict a discrete category or class label (e.g., 'will it rain?', 'is this email spam?'). It is not suitable for predicting a continuous numerical value like the exact number of units sold.
• B. Correct. Regression is the correct machine learning technique for predicting a continuous numerical outcome. The goal is to predict the 'exact number' of ice cream units, which is a continuous quantity, making this a classic regression problem.
• C. Incorrect. Clustering is an unsupervised learning technique used to group similar data points into distinct clusters. It discovers patterns within the data but does not predict a specific output value based on input features.
• D. Incorrect. Semantic segmentation is a specialized computer vision task that involves classifying each pixel in an image with a specific label. It is entirely unrelated to predicting sales figures.

Correct answers: A, B, D — Classification; Regression; Time-series Forecasting

• A. Correct. Classification is a core supervised learning task natively supported by Azure Automated ML. It involves predicting a categorical label (e.g., 'spam' or 'not spam', 'cat' or 'dog'), and Automated ML automates the process of finding the best model for both binary and multi-class classification problems.
• B. Correct. Regression is another fundamental supervised learning task supported by Azure Automated ML. It is used for predicting a continuous numerical value (e.g., the price of a house, the temperature tomorrow). Automated ML handles the entire workflow, from feature engineering to model selection and tuning for regression tasks.
• C. Incorrect. Reinforcement learning is a different paradigm of machine learning where an agent learns by interacting with an environment. This requires specialized training loops and environments which are not natively supported by the standard Azure Automated ML service.
• D. Correct. Time-series forecasting is a specialized task natively supported by Azure Automated ML. It involves predicting future values based on historical time-ordered data. Automated ML provides specific capabilities for handling temporal data, such as lags and windowing, to build accurate forecasting models.
• E. Incorrect. Unsupervised image generation, often involving complex deep learning models like Generative Adversarial Networks (GANs), is not a native capability of Azure Automated ML. Such tasks require custom model development and are outside the scope of the pre-defined tasks that Automated ML supports.

Correct answers: A, B, C — The model file (or registered model); A scoring script (entry script); An environment (Docker image/conda definition)

• A. Correct. The model file (or a registered model in your Azure ML workspace) is the core component. It contains the learned parameters and architecture necessary for the endpoint to perform inference and make predictions.
• B. Correct. A scoring script, also known as an entry script (e.g., `score.py`), is required. This script contains the code to load the model and defines the logic for processing incoming data requests and returning predictions.
• C. Correct. An environment is necessary to define the runtime dependencies, such as the operating system, Python version, and specific libraries (e.g., scikit-learn, PyTorch) required to run the model and scoring script. This ensures consistent and reproducible execution on the managed compute.
• D. Incorrect. The deployment process uses Azure-managed compute resources, not your local machine. While you can choose GPU-enabled instances for your endpoint in Azure, a local GPU-based laptop is not a requirement for the deployment itself.
• E. Incorrect. The raw training dataset is used to train the model, but it is not required for deployment. The deployment process only needs the final trained model artifact to make predictions on new, unseen data.

Correct answers: A, C, D — Classification; Regression; Time-series Forecasting

• A. Correct. Azure AutoML supports classification, a supervised learning task for predicting a categorical label (e.g., 'yes' or 'no'). It automates the process of algorithm selection, feature engineering, and hyperparameter tuning to find the best classification model for the given data.
• B. Incorrect. 3D model rendering is a computer graphics and visualization task, not a machine learning task. This is handled by specialized rendering engines or services like Azure Remote Rendering, not AutoML.
• C. Correct. Azure AutoML supports regression, a supervised learning task for predicting a continuous numerical value (e.g., price, temperature). It automates preprocessing, model selection, and tuning for various regression algorithms.
• D. Correct. Azure AutoML has specialized support for time-series forecasting, which involves predicting future values based on historical time-stamped data. It automates tasks like creating time-based features, handling seasonality, and selecting appropriate forecasting models.
• E. Incorrect. Robotic Process Automation (RPA) focuses on automating repetitive, rule-based tasks, often through UI interactions. This is a different domain from machine learning and is typically handled by tools like Microsoft Power Automate, not Azure AutoML, which is designed for building and training predictive models.

Correct answers: A, B — They are trained on vast amounts of data.; They can be adapted (fine-tuned) for various downstream tasks.

• A. This is a core characteristic. Foundation models, by definition, are pre-trained on massive, broad datasets (often web-scale text, images, and code). This extensive training on vast amounts of data is what allows them to learn complex patterns and relationships, enabling their powerful generalization and emergent capabilities.
• B. This is a key advantage of foundation models. Their general-purpose knowledge can be specialized for specific, or 'downstream,' tasks. This adaptation can be done through methods like fine-tuning on a smaller, task-specific dataset or through various prompting techniques. This versatility and transferability make them highly valuable for a wide range of applications.
• C. This is incorrect. Foundation models are exceptionally versatile and excel at a wide range of tasks, particularly those involving sequential data like Natural Language Processing (text generation, summarization, translation). They are not limited to a single machine learning technique like numerical regression.
• D. This is incorrect. A key feature of models like GPT-4 is their generative nature. They are specifically designed to generate new, original content (text, code, images, etc.) that is coherent and contextually relevant, based on the patterns learned during training. They do not merely analyze or retrieve existing information.
• E. This is incorrect. The pre-training of foundation models typically uses a self-supervised learning approach, such as predicting the next token in a sequence. This is distinct from unsupervised clustering, which is a different machine learning technique focused on grouping similar data points together without predefined labels.

Correct answer: B — The pixel values of the image

• A. Incorrect. The class name (e.g., 'Cat', 'Dog') is the label or the target variable. It represents the 'ground truth' that the model is trained to predict, not the input data (features) used for making the prediction.
• B. Correct. In computer vision, an image is represented as a grid of pixels. The intensity values of these pixels (e.g., RGB values for a color image) are the raw numerical data that serve as the input features for the model. The model learns to identify patterns from these pixel values to classify the image.
• C. Incorrect. The filename is metadata that identifies the file on a storage system. It does not contain any visual information about the image's content and is therefore not used as a feature for classification.
• D. Incorrect. Resolution metadata describes the dimensions of the image (e.g., 1920x1080 pixels). While this information might be used during a data preprocessing step, like resizing all images to a standard dimension, it does not represent the visual content of the image and is not the primary feature set.

Correct answer: C — Training

• A. Incorrect. The validation dataset is used during the training process to tune the model's hyperparameters (like learning rate) and to evaluate its performance on data it hasn't been trained on, which helps in preventing overfitting and for tasks like early stopping. It is not used to directly adjust the core model parameters like weights and biases.
• B. Incorrect. The test dataset is a completely separate set of data that is held back until after the model has been fully trained and tuned. It is used for the final, unbiased evaluation of the model's performance and its ability to generalize to new, unseen data.
• C. Correct. The training dataset is the portion of data used to train the machine learning model. The learning algorithm iterates over this data to learn patterns and relationships, adjusting the model's internal parameters (weights and biases) to minimize a loss function and make accurate predictions.
• D. Incorrect. Inference is the process of using a fully trained model to make predictions on new, live data. The data used during inference is not used to fit or adjust any model parameters.

Correct answer: B — Supervised learning

• A. Incorrect. Unsupervised learning operates on datasets that do not have labels. Its primary goal is to discover hidden patterns, structures, or relationships within the input features themselves.
• B. Correct. This is the definition of supervised learning. It requires a dataset containing both input features and corresponding known output labels (or target values). The model learns the mapping between the features and labels to make predictions on new, unseen data. Common tasks include classification and regression.
• C. Incorrect. Clustering is a specific technique within unsupervised learning. It groups similar data points together based on their features, without using any predefined labels. Its goal is to discover inherent groupings in the data.
• D. Incorrect. Reinforcement learning trains an agent to make sequential decisions by interacting with an environment. The agent learns through a feedback loop of actions and consequences, receiving rewards or penalties, rather than by training on a pre-labeled dataset of feature-label pairs.

Correct answers: A, C — Features are the input variables.; Labels are the output or target variable.

• A. Correct. Features are the input variables, representing the measurable properties or attributes of the data. A machine learning model uses these features to learn patterns and make predictions.
• B. Incorrect. Labels are the output or target variable that the model learns to predict. The input variables are known as features.
• C. Correct. In supervised machine learning, the label is the known outcome or target variable. It serves as the ground truth that the model is trained to predict based on the input features.
• D. Incorrect. The model predicts the value of the label, not the features. Features are the inputs used by the model to make its prediction.
• E. Incorrect. A machine learning model typically uses multiple features to predict a single label. The relationship is not restricted to one-to-one; it is common to have many input features for one output label.

Correct answer: B — Object detection

• A. Incorrect. Image classification assigns one or more labels to an entire image (e.g., 'street scene'). It identifies what is in the image as a whole but does not provide the specific location or coordinates of individual objects within it.
• B. Correct. Object detection is the specific technique that identifies and locates multiple objects within an image. It returns a class label for each object (e.g., 'car', 'pedestrian') and a bounding box defined by coordinates, which is precisely the information a self-driving car needs to determine the position of other road users.
• C. Incorrect. Semantic segmentation classifies each pixel in an image to a specific category. While this provides detailed, pixel-level location information by creating masks, it does not typically return simple coordinates for individual object instances in the way that object detection with bounding boxes does.
• D. Incorrect. Image analysis is a broad, high-level term that encompasses many computer vision tasks, including tagging, description, and object detection. However, it is not the specific technique for providing object coordinates. Object detection is the more precise and correct answer.

Correct answers: B, C — Object detection; Optical Character Recognition (OCR)

• A. Incorrect. Image classification assigns a single label to an entire image. It is unsuitable for this scenario as it cannot locate multiple different objects (like various vehicles) within the same image, nor can it isolate a specific region like a license plate to read its text.
• B. Correct. Object detection is required to locate and classify individual vehicles (car, bus, truck) within the image. It provides bounding boxes around each detected vehicle, which is the necessary first step to identify the vehicle type and pinpoint the license plate area for further processing.
• C. Correct. Optical Character Recognition (OCR) is required to extract and read the alphanumeric characters from the license plate. After an object detection model identifies the vehicle and the license plate's location, OCR is applied to that specific region of the image to convert the text into a machine-readable format.
• D. Incorrect. Face detection is a specialized form of object detection used exclusively for identifying human faces in an image. It is not relevant for identifying vehicles or reading license plates.
• E. Incorrect. Sentiment analysis is a Natural Language Processing (NLP) task, not a computer vision task. It analyzes text to determine emotional tone and is completely unrelated to processing images of vehicles.

Correct answers: A, B — Estimating the age of a person; Detecting the emotional state (e.g., happiness, anger)

• A. Correct. Facial analysis can estimate a person's approximate age by analyzing their facial features and patterns. This is a common attribute returned by services like the Azure AI Vision Face service, though the result is a heuristic estimate, not an exact value.
• B. Correct. A core capability of facial analysis is detecting emotional states by analyzing facial expressions. The system identifies key facial cues to infer emotions like happiness, sadness, and anger, often returning confidence scores for each detected emotion.
• C. Incorrect. Facial analysis focuses on detecting attributes of a face (like age or emotion). Facial identification, a separate capability, matches a face against a pre-approved, private dataset. It does not automatically query external public or government databases to find a person's name.
• D. Incorrect. Speech translation is a function of speech and language services, such as Azure AI Speech, not computer vision. This task involves processing audio data, which is outside the scope of facial analysis.
• E. Incorrect. Facial analysis is limited to the features of the face. Determining a person's height requires a full-body view, depth information, or other contextual clues, which are beyond the capabilities of standard facial analysis.

Correct answer: B — Face verification

• A. Incorrect. Face detection is the process of locating human faces in an image and returning their bounding box coordinates. While it is a necessary first step for any face analysis, it does not perform identity comparison or decide if two faces belong to the same person.
• B. Correct. Face verification performs a one-to-one comparison between two faces to determine if they belong to the same person. This is exactly what is required for an authentication scenario like building access, where a live captured face is compared against a pre-registered employee photo.
• C. Incorrect. Face grouping is used to cluster a set of unknown faces into smaller groups based on visual similarity. It's useful for organizing large collections of photos but is not suitable for a one-to-one authentication task against a known identity.
• D. Incorrect. Face attributes analysis provides information about a detected face, such as estimated age, gender, emotion, facial hair, and glasses. This feature enriches the data about a face but does not perform identity verification.

Correct answers: A, B — The detected language of the text; The bounding box coordinates of text lines and words

• A. The Azure AI Vision Read API response includes the detected language of the extracted text, usually represented by a BCP-47 language code. This is a key feature that helps in downstream processing, such as applying language-specific parsing or translation.
• B. The response from the Read API provides the precise location of each detected line and word of text as a set of bounding box coordinates (a polygon). This allows applications to pinpoint or highlight the extracted text on the original image.
• C. Incorrect. The Azure AI Vision Read API is an Optical Character Recognition (OCR) service focused on extracting text. It does not perform semantic analysis. Determining emotional sentiment is a Natural Language Processing (NLP) task handled by services like Azure AI Language.
• D. Incorrect. Identifying the author of a text is a complex task far beyond the scope of the Read API, which only extracts visible textual content from an image or document.
• E. Incorrect. The Read API returns the raw extracted text. Creating a summary of the content is a separate, advanced NLP task that requires a service like Azure AI Language to process the text after it has been extracted.

Correct answers: A, C — Ensuring the system preserves privacy and data protection; Ensuring fairness and mitigating bias across different demographics

• A. Correct. Preserving privacy and ensuring data protection are fundamental principles of Responsible AI, especially when deploying sensitive technologies like facial recognition in public areas. This involves obtaining proper consent, minimizing data collection, ensuring secure storage, and complying with regulations such as GDPR to protect individuals' rights.
• B. Incorrect. Prioritizing processing speed at the expense of accuracy violates Responsible AI principles. In sensitive applications like facial recognition, accuracy, fairness, and safety are paramount to prevent incorrect identifications and potential harm.
• C. Correct. Facial recognition systems have historically demonstrated performance disparities across different demographic groups. The principle of Fairness requires actively working to mitigate these biases to ensure the system does not produce discriminatory or inequitable outcomes. This involves using diverse training data and rigorous evaluation.
• D. Incorrect. Collecting data without consent is a direct violation of privacy, legal, and ethical standards. Responsible AI practices mandate that data collection must have a lawful basis and be limited to what is necessary for the specified purpose (data minimization).
• E. Incorrect. Hiding the use of an AI system from the public goes against the principle of Transparency. To build trust and ensure accountability, it is crucial to be open about when and how AI systems are being used, particularly in public-facing scenarios.

Correct answer: B — Speech to text

• A. Incorrect. Text-to-speech is the capability that converts written text into spoken words. This is the opposite of what is required for captioning, which involves generating text from audio.
• B. Correct. Speech-to-text is the core capability that converts spoken language from an audio source into written text. The Azure AI Speech service provides real-time transcription, making it ideal for generating live captions for video streams.
• C. Incorrect. Speaker recognition is used to identify and verify individuals based on their unique voice characteristics. While it can be used to label who is speaking in a transcript, it does not generate the text content of the captions.
• D. Incorrect. Language detection identifies the language being spoken. Although this is often a necessary preliminary step to configure the correct speech-to-text model, it does not perform the transcription or generate the captions itself.

Correct answers: A, C, D — Language detection; Question answering; Key phrase extraction

• A. Correct. Language detection is a core feature of the Azure AI Language service. It identifies the language of a given text, which is often a crucial first step for processing multilingual content or routing text to appropriate language-specific models.
• B. Incorrect. Speech to text (or speech recognition) is a feature of the Azure AI Speech service, not the Azure AI Language service. The Speech service is dedicated to processing audio, while the Language service focuses on analyzing text.
• C. Correct. Question answering is a key feature of the Azure AI Language service. It enables developers to build systems that can find and return precise answers from structured or unstructured text sources, such as documents or knowledge bases.
• D. Correct. Key phrase extraction is a standard feature within the Azure AI Language service. It analyzes text to identify and extract the main concepts or talking points, which is useful for tasks like document summarization and content indexing.
• E. Incorrect. Text to speech (or speech synthesis) is a feature of the Azure AI Speech service. Its function is to convert written text into natural-sounding spoken audio, which is distinct from the text analysis capabilities of the Azure AI Language service.

Correct answer: A — True

• A. The statement is true because the Question Answering capability of Azure AI Language is specifically designed to create a knowledge base from various data sources, including PDF documents. It can then be used to find and return the most relevant answers to user questions based on the content of those documents. This functionality is a core component for building a chatbot that can answer questions about company policies.
• B. The statement is false because the scenario described—creating a system to answer questions based on the content of a PDF document—is a primary use case for the Question Answering feature. While a complete chatbot solution requires other components like a user interface, Question Answering provides the essential capability to solve the core problem of knowledge extraction and retrieval from the document.

Correct answer: B — Language Detection

• A. Incorrect. Sentiment Analysis determines the emotional tone of a text (e.g., positive, negative, neutral). While this could be useful for prioritizing tickets based on urgency or customer mood, it does not identify the language of the ticket, which is the primary requirement for routing.
• B. Correct. The fundamental requirement is to route tickets based on their language. Therefore, the first step must be to identify the language of each ticket. The Language Detection feature, part of the Azure AI Language service, is specifically designed for this purpose.
• C. Incorrect. Key Phrase Extraction identifies the main topics or talking points within a text. This is useful for understanding the subject of the ticket and could be used for routing to a specific product expert, but it does not determine the language of the text.
• D. Incorrect. Entity Recognition identifies and categorizes named entities such as people, organizations, locations, or products. This can provide context for handling the ticket but does not detect the language in which the ticket is written.

Correct answers: A, C — Speech-to-Text; Text-to-Speech

• A. Correct. Speech-to-text is a fundamental requirement for a voice assistant. It converts the user's spoken voice commands into written text, which the application can then process and understand.
• B. Incorrect. Speaker recognition is used to identify or verify who is speaking. While this can be a useful feature for personalization or authentication, it is not a core requirement for the assistant to simply understand and respond to a command.
• C. Correct. Text-to-speech is essential for the voice assistant to provide a spoken reply. After processing a command and generating a text-based response, this service converts that text into natural-sounding speech.
• D. Incorrect. Form Recognizer (now known as Azure AI Document Intelligence) is a service designed to extract text, key-value pairs, and tables from documents. It is not related to processing spoken commands or generating spoken replies.
• E. Incorrect. Translation is used to convert text or speech from one language to another. While this would be necessary for a multi-lingual assistant, it is not a required feature for the basic functionality of understanding and replying in a single language.

Correct answer: A — True

• A. The statement is true because the Speech Translation feature of Azure AI Speech is specifically designed for real-time speech-to-text translation. It can take a live audio stream, such as a spoken conversation, and produce a text transcription in a different target language, making it the ideal solution for this scenario.
• B. The statement is false because the primary function of the Speech Translation feature is to translate live spoken language into text in another language. Therefore, the feature does meet the specified goal.

Correct answer: A — Generative text completion

• A. Correct. Generative text completion is the core capability of large language models like those in Azure OpenAI Service. These models excel at understanding long passages of text and generating new, coherent, and contextually relevant content. This makes them ideal for abstractive summarization, where they can condense complex documents like financial reports into concise executive summaries.
• B. Incorrect. Object detection is a computer vision capability used to identify and locate objects within images or videos. It is unrelated to processing or summarizing textual information.
• C. Incorrect. Image synthesis is a generative AI capability for creating new images from text prompts or other inputs. It operates on visual data, not text, and cannot be used for document summarization.
• D. Incorrect. Speaker recognition is a capability of Azure AI Speech services used to identify or verify an individual based on their voice from audio data. It is not applicable to summarizing written reports.

Correct answer: B — Content Safety filters

• A. Incorrect. Azure Monitor is a service for collecting, analyzing, and acting on telemetry data, logs, and metrics from cloud and on-premises environments. It is used for performance monitoring and diagnostics, not for content moderation.
• B. Correct. Azure AI services, including Azure OpenAI, have built-in Content Safety filters. These are specifically designed to detect and block harmful content in prompts and generated outputs across categories like hate, sexual, violence, and self-harm, allowing you to enforce safety policies in your applications.
• C. Incorrect. Spot Instances (or Spot VMs) are a cost-saving compute option that utilizes unused Azure capacity at a significant discount. Their availability can be revoked with short notice, and they are completely unrelated to content filtering or moderation.
• D. Incorrect. Auto-scaling is a feature that automatically adjusts compute resources (like the number of virtual machines or instances) based on the current load or demand. It ensures performance and manages costs but does not provide content safety capabilities.

Correct answers: A, B — Azure OpenAI Service (for generation); Azure AI Search (with Vector Search)

• A. Correct. Azure OpenAI Service provides the large language models (LLMs) that perform the 'Generation' part of the RAG architecture. It takes the relevant context retrieved by the search component and the user's original query to generate a comprehensive, context-aware, and human-like response.
• B. Correct. Azure AI Search, specifically with its vector search capability, implements the 'Retrieval' part of the RAG architecture. It indexes and stores data (often as vector embeddings) and efficiently performs similarity searches to find and retrieve the most relevant documents or chunks of information to ground the language model.
• C. Incorrect. Azure Kinect is a developer kit and hardware sensor platform for computer vision, depth sensing, and spatial computing. It is not used for text retrieval and generation, which are the core functions of a RAG solution.
• D. Incorrect. Azure Media Services is a cloud-based platform for media workflows, such as video and audio ingesting, encoding, and streaming. It is not a core component of a text-based RAG solution.
• E. Incorrect. Azure Logic Apps is a service for automating and orchestrating workflows. While it could be used to connect the various components of a larger application that *uses* a RAG pattern, it does not perform the core functions of retrieval or generation itself.

Correct answer: B — False

• A. The statement is false. While models like GPT are well-known for text, the field of generative AI also includes models specifically designed for other modalities. For example, models like DALL-E generate images, and others can create audio, music, and video.
• B. The statement is false. A defining feature of modern generative AI is its ability to create new content across various formats, not just text. Services like Azure OpenAI provide access to multimodal models, such as DALL-E for image generation, demonstrating these broader capabilities.

Correct answer: A — True

• A. The statement is true because Azure OpenAI Service offers access to powerful, pre-trained large language models (LLMs) and image generation models from OpenAI. This allows developers to integrate advanced generative AI capabilities into applications via REST APIs without the extensive resources and expertise required to train foundational models from scratch.
• B. The statement is false. A key feature of Azure OpenAI Service is providing managed access to pre-trained models. While these models can be fine-tuned with specific data, the service is designed so that users do not have to train large-scale models from the beginning, which is one of its primary benefits.

Correct answer: A — True

• A. The statement is true because Microsoft Azure offers a robust ecosystem for generative AI. Key services include Azure OpenAI Service, which provides access to powerful pre-trained models like GPT-4, and Azure Machine Learning, which offers tools for building, training, and deploying custom generative models. Azure AI Studio further integrates these capabilities into a unified platform.
• B. The statement is false because Azure has invested heavily in supporting generative AI. The existence of dedicated services such as Azure OpenAI Service, Azure AI Content Safety, and the comprehensive tools within Azure Machine Learning and Azure AI Studio directly contradicts the idea that Azure does not support these workloads.