Free Microsoft Certified: Azure AI Fundamentals (AI-901) Sample Questions

Correct answer: B — Entity detection

• A. Sentiment analysis is used to determine the emotional tone or opinion expressed in a body of text, such as whether a statement is positive, negative, or neutral. It does not identify specific named entities like politicians or cities.
• B. Entity detection (also known as Named Entity Recognition) is the process of identifying and categorizing specific entities within text, such as people, organizations, locations, and dates. This is the ideal technique for tagging articles with names of politicians, cities, and organizations.
• C. Summarization involves creating a shortened version of a document while retaining its primary meaning and key points. It is not intended for extracting and tagging specific names or categories within the text.
• D. Language detection is used to identify the language in which a text is written (e.g., English, Spanish, or French). It does not extract or categorize specific entities like names or locations from the content.

Correct answer: B — Speech recognition

• A. Incorrect. Speech synthesis is the process of generating spoken audio from written text. This is the reverse of the required step in the virtual assistant's workflow.
• B. Correct. Speech recognition (also known as Speech-to-Text) is the AI capability that converts spoken language into written text, which is the first step in the scenario described.
• C. Incorrect. Text-to-speech is a synonym for speech synthesis; it involves turning text data into audible speech, not converting voice input into text.
• D. Incorrect. Speaker recognition is used to identify or verify the identity of the person speaking based on their unique vocal characteristics. It does not transcribe the content of the speech into text.

Correct answer: B — Creating new, original content such as text, images, or code

• A. Categorizing existing data into predefined classes is a classic classification task within machine learning. It focuses on assigning labels to input based on learned patterns rather than generating new material.
• B. Generative AI models use patterns and structures learned from training data to create new, original content such as text, images, audio, or code. This creative capability is the core function that distinguishes generative models from discriminative or extractive models.
• C. Extracting key phrases from a document is a Natural Language Processing (NLP) task related to information extraction. It identifies and retrieves important terms from existing content rather than generating new, original content.
• D. Translating spoken audio from one language to another is a specialized speech and language processing task. While it may utilize AI, its primary objective is translation (mapping existing communication) rather than the generation of original content.

Correct answers: A, C, D — Extracting printed text from a photograph; Identifying the presence of a specific object in an image; Generating a descriptive caption for an image

• A. Extracting printed text from a photograph is a process known as Optical Character Recognition (OCR). This is a standard Computer Vision workload that involves analyzing an image to detect, recognize, and extract text characters from visual media.
• B. Generating a summary of a long document is a Natural Language Processing (NLP) task. It focuses on the linguistic interpretation and summarization of text data rather than visual analysis.
• C. Identifying the presence of a specific object in an image is a core Computer Vision task, specifically object detection or image classification. It uses visual processing algorithms to recognize and locate entities within an image.
• D. Generating a descriptive caption for an image is a Computer Vision task called image captioning. It combines visual analysis of the image content with natural language generation to describe the scene in text.
• E. Translating spoken audio into text is a speech recognition task. It falls under Speech AI or Natural Language Processing (NLP) because it processes audio signals instead of visual data.
• F. Detecting the emotional tone of a paragraph is known as sentiment analysis. This is a text analytics workload within Natural Language Processing (NLP) and does not involve visual or image-based data.

Correct answers: A, C — It can autonomously plan and execute a sequence of actions to achieve a goal.; It can interact with external tools and APIs to gather information or perform tasks.

• A. Agentic AI is designed to autonomously plan and execute a sequence of actions to achieve a goal. Unlike traditional models that follow static rules or single-step commands, agentic systems can break down complex tasks and determine the necessary steps to reach an objective independently with minimal human input.
• B. Agentic AI is characterized by its ability to operate autonomously, specifically reducing the need for human intervention. Requiring human input for every step describes a traditional, non-agentic workflow.
• C. A core feature of Agentic AI is its ability to use external tools, browse the web, or call APIs. This 'tool-use' capability allows the AI to gather real-world data and interact with other software systems to perform tasks beyond simple text generation.
• D. Agentic AI is not strictly limited to image analysis. It is typically built on large language models that work across multiple modalities, including text and code, to perform reasoning and execution.
• E. Agentic AI goes beyond simple text generation; it can plan, call tools, and execute actions. Saying it can only generate text describes a basic generative model rather than a fully realized agentic system.

Correct answer: B — Top-p

• A. Temperature controls the randomness of the output by scaling the probability distribution (logits) before applying softmax. While it makes the model more or less predictable, it does not directly limit the token selection to a specific cumulative probability threshold.
• B. Top-p, also known as nucleus sampling, limits the model's token selection to the smallest set of tokens whose cumulative probability adds up to the value p (e.g., 0.9 for 90%). This technique ensures the model only considers the most likely candidates that make up the specified probability mass, effectively controlling randomness.
• C. Frequency penalty decreases the likelihood of the model repeating the same tokens based on how many times they have already appeared in the output. It is intended to reduce verbatim repetition rather than filtering tokens by their initial probability mass.
• D. Presence penalty penalizes tokens that have already appeared at least once in the generated text to encourage the model to move on to new topics. Like frequency penalty, it affects word choice based on prior generation rather than implementing a top-p threshold filter.

Correct answer: A — Fairness

• A. Correct. Fairness in AI ensures that systems treat all people equitably and avoid discrimination based on characteristics such as gender, race, or ethnicity. The scenario describes a clear bias where women are treated less favorably than men with similar financial backgrounds, which is a direct violation of the fairness principle.
• B. Incorrect. Reliability and safety refer to the consistency, accuracy, and dependability of AI systems, ensuring they perform as intended without causing physical or financial harm due to technical failure. The issue here is not about the technical stability of the model, but about demographic bias in its results.
• C. Incorrect. Privacy and security involve protecting sensitive user data and ensuring that AI systems are resilient to attacks. The scenario focuses on the outcomes of the decision-making process rather than the unauthorized disclosure or protection of personal information.
• D. Incorrect. Transparency involves making AI systems understandable and their decision-making processes clear to stakeholders. While transparency is a tool used to discover bias, the specific act of discriminatory treatment against a protected group is a violation of the fairness principle.

Correct answer: B — Stop sequences

• A. Max tokens is used to limit the total length of the generated response by setting a hard cap on the number of tokens produced. It truncates the output once the limit is reached but does not stop generation based on specific textual content or keywords like 'END'.
• B. Stop sequences allow you to specify one or more character strings that act as a signal for the model to stop generating further tokens. By setting 'END' as a stop sequence, the API will immediately terminate the response as soon as that specific sequence is generated.
• C. Temperature is a parameter that controls the randomness or creativity of the output. Higher values make the model more creative, while lower values make it more deterministic. It influences the selection of tokens but does not determine the termination point based on specific text.
• D. Top-p, or nucleus sampling, controls diversity by limiting the model to a subset of the most probable tokens whose cumulative probability reaches a certain threshold. While it affects the variation of the output, it does not provide a mechanism to stop generation upon reaching a specific word.

Correct answer: B — Reliability and safety

• A. Fairness focuses on ensuring that AI systems treat all people equitably and do not disadvantage specific groups based on sensitive characteristics. A dosage calculation bug is a technical safety failure, not an issue of demographic bias.
• B. Reliability and safety is the principle stating that AI systems should perform consistently, safely, and reliably under expected and unexpected conditions. A software bug that produces potentially lethal dosage recommendations directly violates the requirement for the system to operate safely and avoid causing physical harm.
• C. Inclusiveness involves designing AI systems to empower everyone and accommodate a broad range of human needs and abilities. The scenario describes a dangerous output error rather than an issue with accessibility or serving diverse populations.
• D. Privacy and security concern the protection of personal data and ensuring systems are resilient to unauthorized access. While critical in healthcare, the problem described is a malfunction in the system's logic/output consistency, not a data breach or security vulnerability.

Correct answers: A, D — GPT-3.5 Turbo; GPT-4

• A. GPT-3.5 Turbo is a powerful generative language model capable of understanding and generating human-like text. It is highly effective for text-to-text tasks, such as summarizing long articles into concise paragraphs.
• B. DALL-E 3 is a text-to-image model designed to create high-quality images from descriptive text prompts. It is not capable of performing text-to-text summarization.
• C. Whisper is an automatic speech recognition (ASR) model used for transcribing audio into text. It is a speech-to-text model and does not perform text-to-text summarization tasks.
• D. GPT-4 is an advanced large language model (LLM) that excels at complex reasoning and natural language generation. It is specifically designed for text-to-text tasks, including summarizing long-form content.
• E. Text-embedding-ada-002 is an embedding model that converts text into numerical vector representations. While useful for semantic search and similarity analysis, it cannot generate natural language summaries.

Correct answer: B — To assess and mitigate unfairness and bias in AI systems.

• A. Fairlearn is not a security or encryption tool. Data encryption is typically managed by Azure Key Vault or platform-level encryption features, whereas Fairlearn focuses on responsible AI principles.
• B. Fairlearn is an open-source toolkit integrated with Azure Machine Learning that allows developers to assess and mitigate unfairness and bias in AI systems, directly supporting the Fairness principle of Responsible AI.
• C. Generating real-time captions is a capability of AI services such as Azure AI Speech or Azure AI Vision, and is unrelated to the functions provided by Fairlearn.
• D. Version tracking and lineage for deployed models are handled by MLOps tools and the Azure Machine Learning model registry, not by fairness toolkits.

Correct answer: B — Reliability and safety

• A. Inclusiveness focuses on ensuring that AI systems are accessible and designed to work for a diverse range of people, regardless of background or physical ability. Testing for spoofing attacks like printed photographs is a matter of robustness rather than accessibility.
• B. Reliability and safety require that AI systems perform consistently as intended and are resistant to manipulation, errors, or unsafe behavior. Preventing a facial recognition system from being fooled by a printed photo (spoofing) is a critical part of ensuring the system is robust, secure, and safe for real-world application.
• C. Fairness in AI centers on treating all individuals equally and avoiding bias or discrimination based on demographic traits. This scenario focuses on the system's ability to resist deception and unauthorized access, not on its equitable treatment of different groups.
• D. Transparency involves making AI systems understandable and explainable so that users know how decisions are made. While important for trust, this scenario is specifically about technical security and the system's ability to distinguish between a live person and a photograph.

Correct answers: A, C — Pre-training on large datasets; Inferencing (generating outputs)

• A. Correct. Pre-training on large datasets is a fundamental phase where the model learns language patterns, logic, and information from massive volumes of data. This training provides the foundation required for the model to generate content.
• B. Incorrect. Defragmenting the hard drive is a system maintenance task used to optimize physical file storage on a disk and is completely unrelated to the development or execution of AI models.
• C. Correct. Inferencing (or inference) is the operational phase where the trained model is put to use. In this stage, the model processes prompts or input data to generate specific outputs, responses, or content.
• D. Incorrect. Compiling source code is a standard step in software development to convert code into a machine-readable format. While the frameworks used to build AI are compiled, it is not considered a specific phase of the Generative AI model lifecycle.
• E. Incorrect. Indexing a SQL database is a data management technique used to speed up search queries within a relational database. It is not a core phase in the creation or use of a generative model.

Correct answer: B — False

• A. The statement is false because the Whisper model is an automatic speech recognition (ASR) system designed to transcribe and translate audio speech into text, whereas image generation is the primary function of models like DALL-E.
• B. The statement is true because Whisper's core functionality is speech-to-text processing and multilingual translation, making the claim that it is a text-to-image generation model incorrect.

Correct answer: B — False

• A. The statement is false because the Privacy and security principle does not mandate an absolute ban on personal data usage; instead, it focuses on ensuring that data is handled responsibly, transparently, and in compliance with privacy laws and ethical standards.
• B. The statement is false because the Privacy and security principle allows AI systems to utilize personal data provided there is a valid purpose, informed user consent, and robust technical safeguards in place to protect the data from unauthorized access or breaches.

Correct answer: B — Sentiment Analysis

• A. Named Entity Recognition (NER) is used to identify and categorize entities in text, such as names of people, organizations, locations, and dates. It does not provide insights into the emotional tone or sentiment of the content.
• B. Sentiment Analysis is specifically designed to evaluate the emotional tone of text. It categorizes text into sentiment labels such as positive, negative, neutral, or mixed, making it the correct choice for analyzing product reviews.
• C. PII Detection is used to identify and redact personally identifiable information (PII) like phone numbers, email addresses, or credit card data for privacy and compliance purposes. It does not evaluate sentiment.
• D. Extractive Summarization identifies and extracts key sentences from a document to create a summary. While it shortens content, it does not analyze whether the feedback is positive, negative, or neutral.

Correct answer: C — PII Detection

• A. Key Phrase Extraction is used to identify the main talking points or essential concepts in a text. It is useful for summarization and indexing but does not specifically identify or redact sensitive personal data like names or phone numbers.
• B. Sentiment Analysis evaluates text to determine its emotional tone (e.g., positive, negative, or neutral). It is not designed to recognize or redact personally identifiable information for privacy compliance.
• C. PII (Personally Identifiable Information) Detection is specifically designed to identify and redact sensitive information such as names, phone numbers, and addresses. This is the appropriate choice for meeting privacy regulations and redacting patient data.
• D. Speech-to-Text converts spoken audio into written text. While it might be used to create the initial transcripts, it does not have the built-in capability to detect or redact sensitive information within that text.

Correct answer: C — Pronunciation Assessment

• A. Speech-to-Text converts spoken audio into written text. While it can transcribe what a student says, it does not provide the specific metrics for pronunciation accuracy or fluency scoring required for language learning feedback.
• B. Custom Neural Voice is a text-to-speech feature used to create unique, highly realistic synthetic voices. It is used for speech synthesis, not for analyzing or evaluating a user's spoken input.
• C. Pronunciation Assessment is specifically designed to evaluate how well a speaker pronounces a target phrase. It provides detailed scores for accuracy, fluency, and completeness, making it the ideal feature for language learning applications.
• D. Key Phrase Extraction is a feature of Azure AI Language used to identify the main concepts within a text document. It does not analyze speech patterns, pronunciation, or fluency.

Correct answer: B — False

• A. The statement is false because sentiment analysis in Azure AI Language provides a more granular classification than just binary positive or negative values; it explicitly identifies neutral and mixed sentiments.
• B. The statement is false because Azure AI Language's sentiment analysis capability is designed to return positive, negative, neutral, and mixed labels, providing confidence scores for each category to ensure a nuanced analysis of the text.

Correct answer: B — AudioConfig

• A. SpeechConfig is used to define service-level settings such as the Azure subscription key, region, and recognition language. It does not define the hardware or stream source for audio capture.
• B. AudioConfig is the class in the Azure AI Speech SDK used to define where audio input comes from (e.g., default microphone, an audio file, or a custom stream) or where output goes. For microphone input, developers typically use the AudioConfig.FromDefaultMicrophoneInput() method.
• C. SpeechSynthesizer is the object responsible for Text-to-Speech (TTS) operations. It converts text into audible speech and is not used for capturing or recognizing incoming audio from a microphone.

Correct answer: B — Code Interpreter

• A. Bing Search is used to retrieve up-to-date information from the web. It does not have the capability to analyze local CSV files, perform complex mathematical calculations, or generate visualizations.
• B. Code Interpreter is a tool that enables the agent to execute Python code in a sandboxed environment. It is the optimal choice for data analysis tasks, such as processing CSV files using libraries like pandas, performing mathematical computations, and generating charts or plots (e.g., using Matplotlib).
• C. Function Calling allows the agent to invoke external APIs or custom logic defined by the developer. While it provides extensibility, it is not the primary built-in tool for the direct analysis of uploaded files and automatic chart generation within the agent portal.
• D. Azure AI Search is a retrieval-augmented generation (RAG) tool used for indexing and searching across documents. It is intended for search and retrieval scenarios rather than performing computations or creating visualizations from raw data files.

Correct answer: B — An agent has access to tools and can autonomously decide when to use them.

• A. Incorrect. Agents are not limited to image processing, and both agents and chat completion models primarily process text. The distinction is based on functional orchestration rather than input modality.
• B. Correct. In Microsoft Foundry, an agent is a specialized wrapper around a model that is granted access to specific tools (functions, APIs, or data sources). The agent can autonomously determine which tool to invoke and how to sequence those calls to complete a multi-step task.
• C. Incorrect. Both agents and standard chat models utilize system prompts (or instructions). In fact, agents often require more complex system prompts to define their goals, persona, and the specific rules for using provided tools.
• D. Incorrect. Within the Microsoft Foundry ecosystem, agents are typically hosted and orchestrated in the cloud to leverage high-compute models and cloud-integrated tools, rather than running locally on a client device.

Correct answers: A, B — The Endpoint URL; The API Key

• A. Correct. The Endpoint URL is required so that the client object knows which specific Foundry resource, workspace, or model deployment to route the API requests to.
• B. Correct. In API Key authentication, the API Key itself is the primary credential provided to the SDK to authorize the application's requests against the service.
• C. Incorrect. The Tenant ID is a requirement for Microsoft Entra ID (Azure Active Directory) authentication flows, such as those using DefaultAzureCredential or Service Principals, but it is not used for API Key authentication.
• D. Incorrect. A Client Secret is used in conjunction with a Client ID for service principal authentication. It is not part of the API Key authentication method.

Correct answer: A — True

• A. The statement is true because the Chat Playground in the Azure AI Foundry portal provides a user-friendly, no-code graphical interface that allows users to send prompts, adjust parameters, and inspect model responses to evaluate behavior without writing application code.
• B. The statement is false because the Chat Playground is specifically designed as a no-code tool for prompt engineering and testing; it does not require developers to write scripts or interact via APIs to perform initial validation of deployed models.

Correct answer: B — A multimodal model

• A. Incorrect. A text-to-speech model is designed to convert written text into spoken audio. It lacks the computer vision capabilities required to analyze an image or the generative capabilities to summarize trends from visual data.
• B. Correct. A multimodal model is capable of processing and interpreting multiple types of input data, such as images and text, simultaneously. In this scenario, the model must interpret the visual data of the bar chart (vision) and generate a natural language summary (text), which is the primary function of multimodal models.
• C. Incorrect. Image generation models (like DALL-E) are used to create new images from textual descriptions or prompts. They are not designed to analyze existing images or provide textual summaries of data contained within those images.
• D. Incorrect. A text embedding model converts text into numerical representations (vectors) for tasks like semantic search or similarity comparison. It does not have the ability to process visual information from a bar chart or generate descriptive paragraphs.

Correct answer: B — Run

• A. Incorrect. A deployment makes a model or service available for use, but it is not the object that initiates an execution. It provides the infrastructure, whereas the run provides the execution instance.
• B. Correct. In the agent and thread workflow within Azure AI Foundry, a 'run' is the object used to execute an agent against a specific thread. It represents the active processing state where the agent evaluates the thread history and generates a response.
• C. Incorrect. A system prompt provides instructions, personality, and behavioral guidance to the agent, but it does not trigger the execution of a thread or the generation of a specific response.

Correct answer: B — 1792x1024

• A. Incorrect. This value represents a square aspect ratio (1:1). While it is a standard size supported by DALL-E 3, it does not meet the landscape requirement specified in the scenario.
• B. Correct. 1792x1024 is the standard size parameter for landscape orientation in DALL-E 3. In this configuration, the width is greater than the height, which is the definition of landscape orientation and suitable for a website banner.
• C. Incorrect. 1024x1792 represents a portrait orientation. Because the height is significantly greater than the width, it is intended for vertical images rather than horizontal banners.
• D. Incorrect. 256x256 represents a square aspect ratio. Additionally, this is a legacy size supported by DALL-E 2; DALL-E 3 generally uses higher resolution standard sizes such as 1024x1024.

Correct answer: B — GPT-4o

• A. DALL-E 3 is a text-to-image generation model. While it is excellent for creating images from descriptions, it does not have the capability to analyze user-provided illustrations or generate creative text-based stories.
• B. GPT-4o (the 'o' stands for 'Omni') is a multimodal model designed to process and generate text, audio, and images. It is the ideal choice for this scenario because it can both generate creative text and perform vision-based analysis on illustrations within a single model deployment.
• C. GPT-35-Turbo is a powerful text-based model that excels at generating creative stories; however, it lacks native image-understanding (vision) capabilities required to analyze illustrations.
• D. text-embedding-3-large is used to generate numerical vector embeddings for text, which is helpful for semantic search and classification. It cannot be used for generative storytelling or image analysis.

Correct answer: B — Scene detection

• A. Speaker diarization is used to identify and separate different speakers in an audio stream (answering "who spoke when"). It is useful for processing meetings, interviews, or podcasts but cannot identify visual scene boundaries in video content.
• B. Scene detection is specifically designed to analyze video content to identify transitions and boundaries between scenes. This feature automatically determines when one scene ends and another begins based on visual cues, making it the correct choice for segmenting promotional videos.
• C. Image classification assigns labels or categories to individual images or frames. While it could identify what is in a specific frame, it does not analyze temporal changes or transitions over time to detect scene boundaries.
• D. Document layout analysis is a feature typically found in Document Intelligence services used to extract structural components like paragraphs, tables, and headers from documents. It has no application in video analysis.

Correct answer: A — Language identification

• A. Language identification is the correct feature to use in this scenario. It is specifically designed to automatically detect the primary language spoken in an audio file or written in a text block, making it the ideal tool for categorizing international conference recordings where the language is unknown.
• B. Speaker diarization is a feature used to identify and separate different speakers within a single audio file (e.g., Speaker 1, Speaker 2). While it can attribute speech segments to individuals, it does not provide information regarding the language being spoken.
• C. Sentiment analysis is used to determine the emotional tone or opinion expressed in the content, such as whether a statement is positive, negative, or neutral. It is not capable of identifying the specific language of the recording.
• D. Text translation is used to convert text from a source language to a target language. While some translation workflows incorporate language detection, translation itself requires first knowing (or detecting) the source language, and it is not the primary feature for simply identifying the language used in audio files.

Correct answers: A, B — The image file size exceeds the maximum limit; The image is in an unsupported format (e.g., .exe or .doc)

• A. Correct. The GPT-4o API has specific limits on the file size of images (typically 20MB). If an image exceeds this maximum limit, the API will reject the request, necessitating that images be resized or compressed before being sent.
• B. Correct. The API supports specific image formats (such as JPEG, PNG, WEBP, and non-animated GIF). If a file is uploaded in an unsupported format or is not an image at all (like .exe or .doc), the input will be rejected.
• C. Incorrect. Color depth and the number of colors in an image are not validation factors that would cause the GPT-4o API to reject an input.
• D. Incorrect. The aspect ratio or specific shape of an image (such as being a perfect square) does not affect its acceptability by the API.
• E. Incorrect. The device used to capture the image, whether a mobile phone or a professional camera, is irrelevant as long as the file meets the supported size and format requirements.

Correct answers: A, B — The deployment Endpoint URL; An API Key or Entra ID token

• A. Correct. The deployment endpoint URL is essential for routing the API request to the specific Azure OpenAI resource and model deployment instance. Without this, the application does not have a destination for the request.
• B. Correct. Security and authorization are handled through either an API key or a Microsoft Entra ID (formerly Azure AD) token. These credentials verify that the requestor has the appropriate permissions to access the service.
• C. Incorrect. The physical MAC (Media Access Control) address is a hardware-level identifier for network interfaces and is not involved in higher-level cloud API authentication or routing.
• D. Incorrect. A GitHub personal access token is used to authenticate with GitHub's version control and CI/CD services; it cannot be used to authorize requests to Azure OpenAI or DALL-E 3.
• E. Incorrect. While an email might be associated with a user account, it is not an authentication credential or a routing mechanism for programmatic API calls to DALL-E 3.

Correct answer: B — Table extraction

• A. Key-value pair extraction is used to identify and extract related fields such as labels and their corresponding values (e.g., an invoice number or a date). While useful for forms, it is not the optimal choice for extracting complex structured rows and columns found in financial tables.
• B. Table extraction is specifically designed to identify and extract tabular data organized in rows and columns from documents like PDFs. This feature is the most relevant for an accounting firm's need to export structured financial data into Excel spreadsheets.
• C. Selection mark extraction is used to identify and extract selection marks, such as checkboxes or radio buttons, typically found in surveys or forms. It does not extract the content of complex tables.
• D. Image captioning generates natural language descriptions of the visual content within an image. It is not an information extraction tool for structured data or financial tables found in PDF reports.

Correct answer: A — Object detection

• A. Object detection is the correct choice because it is specifically designed to identify and locate multiple objects within an image. It provides the labels for the detected objects and their bounding box coordinates, which identifies their specific location.
• B. Optical Character Recognition (OCR) is used to extract printed or handwritten text from images. It is not suitable for identifying general physical objects or their positions in a photo.
• C. Document layout analysis is used to understand the structure of documents, such as identifying tables, paragraphs, and headings. It is not used for detecting product objects in e-commerce images.
• D. Speaker diarization is an audio processing technique used to identify and distinguish between different speakers in an audio stream. It has no relevance to image processing or object detection.

Correct answer: A — Playground

• A. Correct. The Content Understanding Playground in Azure AI Foundry is a no-code, interactive environment specifically designed for testing and visualizing document extraction. It allows users to upload documents, run analysis, and review extracted fields and structures in real-time without writing any code.
• B. Incorrect. A database is used for storing and querying data, but it does not provide visual testing or interactive analysis tools for document extraction features within Azure AI Foundry.
• C. Incorrect. A Virtual Machine is a cloud computing resource that provides an operating system environment. While it can run applications, it is not the built-in visual tool provided by Azure AI Foundry for evaluating AI model performance.