When interacting with the Azure OpenAI REST API, requests and responses are structured using the JSON format. According to the Microsoft Learn documentation on the Azure OpenAI REST API, JSON is the required format for sending prompts, configuration parameters (such as temperature, max_tokens, or top_p), and metadata to the API endpoint.
A typical request to an Azure OpenAI endpoint (for example, /openai/deployments/{model-name}/chat
/completions?api-version=2024-02-01) includes a JSON body similar to the following:
{
"messages": [
{"role": "system", "content": "You are an assistant."},
{"role": "user", "content": "Write a JavaScript function to add two numbers."}
],
"max_tokens": 200,
"temperature": 0.7
}
This JSON format allows for structured data exchange between the client and the server, ensuring that key- value pairs are properly parsed by the API.
* Option A (CSV): Used for storing tabular data, not suitable for API communication.
* Option C (XML): Though historically used in web services, Microsoft's modern APIs, including Azure OpenAI, rely on JSON.
* Option D (YAML): Common in configuration files but not for REST API payloads.
Therefore, per official Microsoft documentation and the AI-900 study guide, the correct and verified answer is B. JSON, as it is the standard format required when sending requests to the Azure OpenAI REST API endpoint.

In Conversational Language Understanding (CLU), a core service within Azure AI Language, intents represent the goals or purposes behind user utterances (for example, "Track my order" or "Cancel my subscription"). However, in real-world scenarios, users often provide inputs that do not match any defined intent. To handle such cases gracefully, Microsoft recommends including a "None" intent that captures out-of- scope utterances - text that doesn't belong to any other intent in your model.
According to the Microsoft Learn module: "Build a Conversational Language Understanding app", the None intent serves as a catch-all or fallback category for utterances that the model should ignore or respond to with a default message (e.g., "I'm sorry, I don't understand that."). By training the model with multiple examples of irrelevant or unrelated utterances in this intent, you improve its ability to distinguish between valid and invalid user inputs.
The other options are incorrect:
* A. Export the model: Exporting only saves or transfers the model; it does not influence how the model detects irrelevant utterances.
* B. Create a new model: A new model would not inherently solve out-of-scope detection unless properly trained with a None intent.
* D. Create a prebuilt task entity: Entities identify specific data (like dates or products) within valid intents, not irrelevant utterances.
Thus, the correct approach to ensure that your CLU model can detect utterances outside its intended scope is to add examples of unrelated or off-topic phrases to the None intent. This improves classification accuracy and prevents incorrect intent matches.
# Correct answer: C. Add utterances to the None intent

Explanation:

This question evaluates understanding of the different Azure AI Computer Vision services and their distinct functionalities, as covered in the Microsoft AI-900 study guide and Microsoft Learn modules under "Describe features of common AI workloads" and "Identify Azure services for computer vision."
* Azure AI Document Intelligence (formerly known as Form Recognizer):This service is designed to extract structured information from documents, such as forms, receipts, and invoices. It uses optical character recognition (OCR) combined with AI models to detect key-value pairs, tables, and handwritten text. This makes it ideal for automating data entry and digitizing scanned documents.
Hence, it matches "Extract information from scanned forms and invoices."
* Azure AI Vision (formerly Computer Vision):This service provides image and video analysis capabilities. It can detect objects, people, text, and scenes; generate image captions; and extract descriptive tags. It also supports OCR for printed and handwritten text within images. Therefore, it matches "Analyze images and video, and extract descriptions, tags, objects, and text."
* Azure AI Custom Vision:Custom Vision allows you to train your own image classification and object detection models using your own labeled images. Unlike the general Vision service, Custom Vision lets you build domain-specific models-for example, detecting your company's products or identifying manufacturing defects. Hence, it matches "Train custom image classification and object detection models by using your own images." These three services complement each other within Azure's computer vision ecosystem, collectively supporting both general-purpose and specialized AI solutions for visual data analysis.

Explanation:

In the Microsoft Azure AI Fundamentals (AI-900) and Azure Machine Learning (AML) learning paths, deploying a real-time inference pipeline refers to making a trained machine learning model available as a web service that can process incoming data and return predictions instantly. To achieve this, the model must be deployed to an infrastructure capable of handling continuous, low-latency requests with high reliability and scalability.
Microsoft's official guidance from Azure Machine Learning documentation specifies that:
* For testing or development, you can deploy to Azure Container Instances (ACI) because it provides a lightweight, temporary environment suitable for small-scale or non-production workloads.
* For production-grade, real-time inference, the deployment should be made to Azure Kubernetes Service (AKS).
AKS provides enterprise-level scalability, load balancing, and high availability, which are critical for serving real-time predictions to multiple consumers simultaneously. It manages containerized applications using Kubernetes orchestration, allowing the model to scale automatically based on traffic demands.
Azure Machine Learning Compute is mainly used for model training and batch inference pipelines, not real- time endpoints. A local web service is typically used only for debugging or offline testing on a developer machine and cannot be shared for external consumption.
Therefore, when deploying a real-time inference pipeline as a service for others to consume, the correct and Microsoft-verified option is Azure Kubernetes Service (AKS). This environment ensures production readiness, secure endpoint management, and scalability for live AI applications, fully aligning with best practices outlined in the Azure Machine Learning designer documentation and AI-900 exam objectives.
https://docs.microsoft.com/en-us/azure/machine-learning/concept-designer#deploy

To prevent hateful or harmful responses from being returned by a generative AI solution, the correct approach is to implement content filtering. According to the Microsoft Learn documentation for Azure OpenAI Service and the Responsible AI principles, content filtering is a built-in safety mechanism that automatically screens both user prompts (inputs) and model outputs (responses) for inappropriate, harmful, or policy-violating material.
Content filters are designed to detect and block content such as:
* Hate speech or harassment
* Sexual or explicit material
* Self-harm or violent content
* Personally identifiable information (PII) misuse
In Azure OpenAI, the content filtering system is part of Microsoft's Responsible AI standard and cannot be disabled. It ensures that generative AI models such as GPT-3.5 or GPT-4 operate safely and ethically, reducing the risk of producing offensive or discriminatory text. The filter evaluates model responses in real time and can modify, block, or flag inappropriate outputs before they reach the user.
Let's review the other options:
* A. Abuse monitoring tracks misuse after deployment but does not actively prevent hateful responses.
* C. Fine-tuning customizes a model's style or domain knowledge but does not guarantee filtering of offensive content.
* D. Prompt engineering helps steer model behavior but cannot fully prevent harmful outputs.
Therefore, to proactively prevent hateful, unsafe, or offensive responses in a generative AI system built on Azure OpenAI, the correct and Microsoft-verified approach is B. Content filtering.