To programmatically interact with a deployed generative AI model, an application requires two primary pieces of information. The model endpoint (a URL) specifies where to send requests. Authentication credentials (e.g., an API key or token) are essential to verify the application’s identity and authorize access, ensuring secure usage. These are standard requirements for consuming any web service or API, including deployed AI models. The other options are not relevant for making inference calls to a deployed model. Exported playground history is for session replay, training datasets are for model development, and the project display name does not facilitate API calls.

✅ Option A (Correct)

Reasoning: Named Entity Recognition (NER) is the precise technique for identifying and classifying predefined categories like people, organizations, and locations within unstructured text such as customer support emails. This directly addresses the requirement to automatically identify these specific entities.

❌ Why the other choices are incorrect:

• Option B is incorrect: Key phrase extraction identifies the main topics or concepts in text but does not specifically categorize named entities like people or organizations.
• Option C is incorrect: Sentiment analysis determines the emotional tone (positive, negative, neutral) of text, not specific named entities.
• Option D is incorrect: Summarization condenses text into a shorter version, which is not the goal of identifying specific entities.

✅ Option C (Correct)

Reasoning: Designing an interface to support multiple languages and screen readers directly addresses the Microsoft Responsible AI principle of inclusiveness. This principle focuses on empowering everyone and engaging people by considering the full range of human experience and diversity, including accessibility for people with disabilities.

❌ Why the other choices are incorrect:

• Option A is incorrect: Encrypting data and restricting access relates to the ‘Privacy & Security’ responsible AI principle, not inclusiveness.
• Option B is incorrect: Evaluating model outputs across demographic groups to reduce bias is primarily aligned with the ‘Fairness’ responsible AI principle.
• Option D is incorrect: Providing explanations about predictions relates to the ‘Transparency’ responsible AI principle.

✅ Option A (Correct)

Reasoning: The scenario requires building an AI solution that “produces new product images based on written descriptions provided by users.” This directly describes the functionality of image generation (also known as text-to-image generation or generative AI for images). This AI workload takes textual input (descriptions) and synthesizes corresponding visual content (new product images).

❌ Why the other choices are incorrect:

• Option B is incorrect: Image analysis involves processing and understanding existing images to extract information (e.g., classification, content moderation, feature extraction). It does not create new images.
• Option C is incorrect: Object detection is a specific form of image analysis that identifies and locates instances of objects within an image. It does not generate images from text.
• Option D is incorrect: Optical character recognition (OCR) is used to extract text from images or scanned documents. It has no role in generating new visual content.

✅ Option B (Correct)
Reasoning: Speech synthesis, also known as Text-to-Speech (TTS), is the Azure Speech capability designed to convert written text into natural-sounding spoken audio. This directly fulfills the requirement of transforming written customer notifications into speech for a phone system.

❌ Why the other choices are incorrect:

• Option A is incorrect: Speech recognition (Speech-to-Text) converts spoken audio into text, which is the opposite of the required task.
• Option C is incorrect: Speaker recognition identifies or verifies individuals based on their voice, not for converting text into speech.
• Option D is incorrect: Speech translation converts spoken audio from one language to another. While it can involve speech synthesis for the translated output, its primary purpose is translation, not solely text-to-speech in a single language.

✅ a base64-encoded image data

Reasoning: To include an image directly in an API request to a vision-enabled model, the binary image data must be converted to text. Base64 encoding is the standard method for this conversion, allowing the image to be embedded within the JSON payload sent to the API endpoint.

✅ Option A (Correct)

Reasoning: System instructions, also known as system messages or prompts, are fundamental for defining an AI agent’s persona, role, and behavioral constraints. They guide the model on how it should respond and interact with users, ensuring it adheres to a specific character or set of rules.

❌ Why the other choices are incorrect:

• Option B is incorrect: Temperature controls the randomness of the model’s output. A higher temperature results in more creative and diverse responses, while a lower temperature makes responses more deterministic. It does not define the agent’s role or behavior.
• Option C is incorrect: Tokens per minute (TPM) is a metric related to the throughput or rate limits of a language model. It indicates how many tokens the model can process or generate per minute, not its role or behavior.
• Option D is incorrect: Max completion tokens sets the maximum length of the generated response. It limits the output size but does not define the AI agent’s inherent role, personality, or behavioral guidelines.

Statement 1: System prompts can be used to authorize users. → No

Reasoning: Authorization is a security process that controls access to resources, typically managed via API keys, tokens, or Role-Based Access Control (RBAC). A system prompt is an instruction that guides a model’s behavior; it has no role in user authorization.

Statement 2: A system prompt is used to reduce tokens per minute (TPM). → No

Reasoning: Tokens per Minute (TPM) is a service quota or rate limit on an AI model deployment. A system prompt consumes tokens as part of the input and does not manage or reduce this pre-configured service limit. TPM is controlled at the resource management level.

Statement 3: A system prompt guides the behavior of a generative AI model. → Yes

Reasoning: This is the primary function of a system prompt. It sets the context, persona, instructions, and constraints for the AI model. This initial guidance shapes the model’s tone, style, and the overall nature of its subsequent responses in the conversation.

Azure AI Search, leveraging its cognitive search capabilities and skillsets, can process documents like invoices. It integrates with underlying AI services (such as Azure AI Document Intelligence via built-in or custom skills) to extract structured fields, including nested values, based on a defined schema. This enables the enrichment of documents for indexing and subsequent analysis or search.

Why the other choices are incorrect:

• Option A is incorrect: An OCR-only pipeline extracts raw text but lacks the intelligence to identify structured fields, nested values, or apply a defined schema for data extraction.
• Option B is incorrect: Azure Speech services are for audio-to-text transcription or text-to-speech, not document processing. "Foundry Tools" is not a recognized Azure service.
• Option C is incorrect: "Azure Content Understanding" and "Foundry Tools" are not standard or recognized Azure AI services for structured document extraction.

✅ Statement 1: Voice Live returns only transcribed text. —> No

Reasoning: A live voice conversational experience involves a full interaction loop. It must understand the user’s intent (reasoning) from the transcribed text and then generate a spoken response (text-to-speech). Simply returning text describes only the speech-to-text component, not the complete conversational system.

✅ Statement 2: Voice Live requires you to separately implement speech to text and text to speech services. —> No

Reasoning: Azure provides integrated solutions like the Speech SDK that bundle multiple capabilities. This SDK simplifies development by combining speech-to-text, text-to-speech, and connections to language understanding services into a single framework, eliminating the need for separate implementations.

✅ Statement 3: Voice Live combines speech to text, reasoning, and text to speech into a single conversational experience. —> Yes

Reasoning: This statement accurately defines a voice-enabled conversational AI agent. The system ingests speech, uses a reasoning engine like Conversational Language Understanding or a bot to process intent, and synthesizes a spoken reply, creating a seamless, interactive experience for the user.

✅ recognizer.recognize_once()

Reasoning: The code creates a SpeechRecognizer object, whose purpose is to convert speech to text. The recognize_once() method is called to perform a single recognition. It captures one spoken utterance, such as a command, from the default microphone and returns the transcribed text. This is the most direct and fundamental method for handling a single command.

✅ Option A (Correct)

Reasoning: Instructions often define the AI’s persona, such as ‘You are a helpful assistant’ or ‘Act as a technical expert,’ guiding its responses and overall behavior.

✅ Option E (Correct)

Reasoning: Prompts use instructions to set boundaries on output, like specifying format (e.g., ‘respond in JSON’), length (e.g., ‘keep it brief’), or content restrictions (e.g., ‘do not use slang’).

❌ Why the other choices are incorrect:

• Option B is incorrect: Model selection occurs at the API or SDK level when calling the generative AI service, not within the text instructions of a prompt.
• Option C is incorrect: The Azure region for inference is a deployment configuration choice made when setting up the AI resource, entirely separate from prompt instructions.
• Option D is incorrect: Tokens per minute (TPM) allocation is a resource management setting for a deployed model, configured at the Azure subscription or resource level, not within the prompt itself.

✅ AudioOutputConfig(filename=”output.wav”)

Reasoning: To save synthesized speech directly to a file, the SpeechSynthesizer requires an AudioOutputConfig instance. Using the filename parameter within the AudioOutputConfig constructor explicitly directs the synthesized audio stream to the specified .wav file, fulfilling the application’s requirement.

✅ system prompt

Reasoning: A system prompt, also known as a system message, provides initial instructions to a large language model. It defines the agent’s persona, role (e.g., ‘You are a helpful travel agent’), and behavioral guidelines for the conversation. This is the direct method for configuring an agent’s role and behaviors.

result

The begin_analyze method initiates an asynchronous, long-running operation (LRO) and returns a poller object. In the Azure SDK for Python, the standard way to block execution until the LRO is complete and then retrieve the final deserialized result is by calling the result() method on the poller object. This method returns the outcome of the successful operation, which in this case is the invoice analysis data.

Solution Breakdown:

The Python code requires two values to correctly connect to the Azure OpenAI service and call the specific model deployment.

✅ Step 1: resource1 matches with -> base_url
Reasoning: The base URL for an Azure OpenAI resource’s endpoint is constructed using the resource’s unique name. The standard format is https://.openai.azure.com. In this case, the resource name is given as Resource1.

✅ Step 2: my-mini-gpt matches with -> model
Reasoning: When using the Azure OpenAI SDK to make an API call, the model parameter must specify the name of the model *deployment*, not the underlying base model name (like ‘gpt-4.1-mini’). The question states the deployment is named my-mini-gpt.

Statement 1: An AI generative model is retrained before performing each user request.

✅ No

Reasoning: Generative AI models are pre-trained on vast datasets. The process of using the model for a user request is called inference, which leverages the existing training. Retraining is a separate, computationally intensive process and is not performed for each individual query.

Statement 2: An AI agent responds by copying and pasting answers stored in a database.

✅ No

Reasoning: This describes a simple information retrieval system, not an AI agent. While an agent can retrieve data from a source (like in RAG), its primary function is to reason, plan actions, and generate new, context-aware responses rather than just fetching static answers.

Statement 3: An AI agent uses a generative AI model to establish actions based on user input.

✅ Yes

Reasoning: This is a core characteristic of an AI agent. The generative model serves as the agent’s reasoning engine, interpreting the user’s intent to determine a sequence of steps, use tools, or call APIs to accomplish a specific task.

This question maps specific tasks to Microsoft’s six foundational principles for responsible AI. Each task is a direct, practical application of one of the principles. The matching is based on the core definition of each principle as defined by Microsoft.

✅ Fairness matches with -> Evaluating model outputs to ensure that decisions are NOT biased against specific demographic groups.
Reasoning: The Fairness principle aims to ensure AI systems treat all people equitably. A key practice is to actively identify and mitigate biases related to gender, ethnicity, or other demographics to prevent discriminatory or unfair outcomes.

✅ Privacy and security matches with -> Encrypting sensitive customer data and restricting system access to authorized personnel.
Reasoning: This principle covers protecting data and maintaining system integrity. Encrypting data and implementing strict access controls are fundamental security measures to safeguard user privacy and prevent data breaches.

✅ Transparency matches with -> Informing users when they are interacting with an AI system and explaining the system’s capabilities and limitations.
Reasoning: Transparency is about making AI systems understandable. Clearly disclosing when a user is interacting with AI and explaining how it works helps build trust and allows for informed use of the technology.

✅ Reliability and safety matches with -> Testing AI systems under different conditions to reduce unexpected failures.
Reasoning: This principle dictates that AI systems must operate dependably and safely. Rigorous testing under a wide range of operating conditions is essential to ensure the system behaves as expected and does not cause unintended harm.

✅ Generating a response to a user prompt occurs during the inference stage. -> Yes

Reasoning: Inference is the process of using a trained machine learning model to make predictions on new data. Generating a response to a user prompt is a classic example of inference, where the pre-trained model applies its learned knowledge to the input prompt.

✅ A generative AI model generates responses by copying stored documents directly from the model’s training data. -> No

Reasoning: Generative AI models learn statistical patterns, grammar, and concepts from training data. They do not store the data verbatim. Responses are generated by creating new combinations of tokens based on these learned patterns, not by copying and pasting existing documents.

✅ A generative AI model produces output by predicting the next token based on patterns learned from the model’s training data. -> Yes

Reasoning: This is the core mechanism of transformer-based large language models. They are autoregressive, meaning they generate responses one token (a word or part of a word) at a time, with each new token being the most probable one based on the preceding sequence.

✅ Option B (Correct)

Reasoning: Image generation is an AI workload focused on creating new, synthetic images from various inputs, such as written descriptions. This aligns directly with the requirement to produce novel product images based on user-provided text.

❌ Why the other choices are incorrect:

• Option A is incorrect: Image analysis extracts information from existing images, like classifying content or identifying attributes, rather than creating new ones.
• Option C is incorrect: Object detection identifies and locates specific objects within pre-existing images or videos. It does not generate visual content.
• Option D is incorrect: Optical Character Recognition (OCR) converts text found in images into machine-readable text. It is not involved in image creation.

✅ computer vision

Reasoning: Computer vision is the area of AI focused on processing and understanding visual information from images and videos. A primary capability within computer vision is Optical Character Recognition (OCR), which is the specific technology used to detect and extract text from scanned documents and images.

✅ Option C (Correct)

Reasoning: Max Completion Tokens directly limits the maximum number of tokens the generative AI model will produce in its response. This parameter is crucial for controlling output length, thereby minimizing inference costs (which are often token-based) and ensuring responses adhere to a defined maximum length.

❌ Why the other choices are incorrect:

• Option A is incorrect: Top P (nucleus sampling) influences the diversity and quality of the generated text by sampling from a cumulative probability distribution of tokens. It does not directly control the response length or cost.
• Option B is incorrect: Temperature controls the randomness or creativity of the model’s output. A higher temperature yields more varied results, while a lower temperature makes the output more deterministic. It does not control length or cost.
• Option D is incorrect: Model version settings refer to the specific iteration or version of the generative AI model being used. While different models may have different cost structures or performance, this setting does not directly control the length of a specific response during inference.

Statement 1: Human-in-the-loop practices provide accountability for AI-generated decisions. ✅ Yes

Reasoning: The principle of accountability in Responsible AI dictates that people are answerable for AI systems. Human-in-the-loop (HITL) practices ensure direct human oversight and the ability to intervene in AI decisions, making a person ultimately responsible for the outcomes and establishing clear accountability.

Statement 2: Deploying an AI system to a production environment eliminates the need for ongoing monitoring. ✅ No

Reasoning: AI models in production can suffer from performance degradation due to ‘model drift’ or ‘data drift’ as real-world data changes over time. Continuous monitoring is a critical MLOps practice to track performance, detect issues, and ensure the system remains reliable and fair.

Statement 3: Disclosing the team that designed and deployed an AI system provides accountability for the system’s output. ✅ Yes

Reasoning: Accountability is closely linked with transparency. Identifying the creators of an AI system is a key part of establishing who is responsible for its design, behavior, and impact. This makes the team answerable for the system’s output and for addressing any issues that may arise.

✅ accountability

Reasoning: The principle of accountability states that the people who design and deploy AI systems are responsible for how those systems operate. The scenario described, where human reviewers oversee decisions and remain responsible for the final output, is a direct implementation of this principle, ensuring human oversight and ownership of the AI’s actions.

✅ Option C (generative AI) is correct

Reasoning: Generative AI models are specifically designed to create new, original content, such as text, images, or code. Generating marketing email drafts from a given description aligns directly with the capabilities of generative AI, which can synthesize coherent and contextually relevant text.

❌ Why the other choices are incorrect:

• Option A (speech recognition) is incorrect: Speech recognition converts spoken language into text. It does not generate new text content.
• Option B (text classification) is incorrect: Text classification categorizes existing text into predefined labels or classes. It does not create new text drafts.
• Option D (computer vision) is incorrect: Computer vision enables machines to interpret and understand visual information (e.g., images, videos). It is unrelated to generating text content.

Statement 1: The Temperature parameter can be set before deploying a model.

✅ No

Reasoning: The temperature parameter controls the creativity of the model’s response and is an inference-time setting. It is specified within the body of each API request sent to the deployed endpoint, not as a fixed configuration set before or during deployment.

Statement 2: During inference, the model name is used to route requests to a specific deployment.

✅ No

Reasoning: API requests are routed using the specific deployment name, which you define when you deploy a model. While this deployment is associated with a base model, it’s the unique deployment name that is used in the API endpoint or SDK calls to direct the request.

Statement 3: After a model is deployed, both code and testing tools can be used to interact with the model.

✅ Yes

Reasoning: A deployed model exposes a REST API endpoint. This enables interaction through code (using SDKs or direct HTTP calls) and through testing tools. Azure AI Studio provides a ‘Playground’ for interactive testing, and tools like Postman can also call the endpoint.

✅ fairness

Reasoning: The Microsoft responsible AI principle of fairness specifically addresses the need to treat all people fairly. Evaluating a model’s performance across different demographic groups (like age, gender, or ethnicity) is a primary method for identifying and mitigating bias, ensuring the AI system does not unfairly impact any particular group.

✅ be explainable to users

Reasoning: The Microsoft responsible AI principle of transparency focuses on making AI systems understandable. This involves providing clear information about the system’s purpose, how it operates, and its limitations. Explainability is a core component, enabling users to comprehend the reasoning behind AI decisions, which builds trust and confidence in the system.

✅ generative AI

Reasoning: Generative AI is the specific field of artificial intelligence focused on creating new, original content—such as text, images, or code—based on user input or prompts. This directly matches the definition of an AI workload that produces new content. The other options describe analytical tasks performed on existing content.