AI-102 EXAM QUESTIONS 2024 MICROSOFT AZURE AI 102 CERTIFICATION COURSE PART-10

This video covers a series of questions related to Azure AI Document Intelligence and Azure OpenAI. The first question asks about extracting specific data like shipping address, billing address, customer ID, amount due, due date, total tax, and subtotal from scanned documents using Azure AI Document Intelligence. The answer is the "invoice" model, as it’s designed to extract these fields, minimizing development effort compared to custom extraction. The video then elaborates on custom models within Document Intelligence, differentiating between "extraction" (template/neural) and "classification" models. Template models are for documents with common visual layouts and support over 100 languages, training quickly. Neural models handle structured and unstructured documents but are English-only, take longer to train, and are region-specific. Custom classifiers identify document types. Composite models group multiple custom models for broader capabilities, supporting up to 200 models. The next question involves analyzing a PDF with table data using Azure AI Document Intelligence. The correct model to extract table data is the "pre-built layout" model, which can extract text, tables, and other layout elements. The video also touches upon the API call structure, noting that Document Intelligence is still referred to as Form Recognizer in some API contexts. Following this, the transcript discusses a theoretical question about API authentication. For Azure AI services, the standard header for authentication is "ocp-apim-subscription-key," which is used to pass the subscription key for authorization. The subsequent question focuses on identifying handwritten content in scanned PDFs. The "pre-built read" model is identified as the solution, as it can extract text, word locations, detected language, and handwritten styles. The question then delves into the confidence score threshold for recognizing handwritten content. A threshold of 0.75 is recommended as an industry-standard balance between accuracy and coverage, avoiding excessive exclusion of potentially handwritten content due to overly strict confidence requirements. The video then addresses a scenario where an app needs to recognize contract documents, including an additional contract format, while minimizing development effort. The recommended solution is to add the additional contract format to the existing training set and retrain the model, rather than creating a new custom model or lowering the confidence score, which would reduce accuracy. Next, the task is to process scanned expense claims and extract data like merchant information, transaction time and date, taxes paid, and total cost. The "pre-built receipt" model is suggested as the appropriate solution for this, as it can extract these fields and minimizes development effort. The discussion shifts to sentiment analysis for detecting negative comments in customer feedback, with the constraint that the feedback must remain on the company's internal network. The correct sequence of actions involves provisioning the language service in Azure, deploying a Docker container to an on-premise server, and then running the container and querying the prediction endpoint. This approach ensures data privacy by keeping the processing internal. The subsequent question concerns building a chatbot for a travel agent that needs to repeatedly ask for a destination until valid input is received or the user closes the conversation. The "prompt" dialog type is identified as the most suitable solution, as it can manage context, repeat questions until valid input, and handle conversation flow. Finally, the video tackles a question about providing access to an Azure OpenAI service with multiple GPT-3.5 model deployments. Apps will access the service via REST API and use specific deployments. To ensure only authorized apps can access the service and its deployments, each app needs an API key for the Azure OpenAI resource and the specific deployment name to connect to its intended workload. The transcript emphasizes the distinction between the OpenAI resource's endpoint and the deployment name. The video concludes by encouraging viewers to engage with the content and provide feedback.