AI-102 EXAM QUESTIONS 2024 MICROSOFT AZURE AI 102 CERTIFICATION COURSE PART-7

Contoso Limited, an international accounting firm with offices in France, Portugal, and the United Kingdom, uses Azure, Microsoft 365, Dynamics 365, and Azure Active Directory (now Microsoft Entra) for securing role-based access. They employ a naming convention of "country level role" for group assignments. Their existing intellectual property includes weekly webinars stored on blob storage, blogs on WordPress, wikis on Cosmos DB, monthly conference recordings on SharePoint, and FAQs on SharePoint. Text-based content like blogs, wikis, and FAQs are currently provided in only one language per domain and are not translated across languages. Contoso has several new requirements: 1. **Document Processing Workflow**: Automatically extract information from PDFs and images of financial documents. 2. **Customer Support Chatbot**: Answer questions using FAQs. 3. **Searchable Knowledge Base**: Enable search across all intellectual properties (webinars, blogs, wikis, recordings, FAQs). Technical requirements for these solutions include: * All content must be approved before publication. * All projects must support English, French, and Portuguese, necessitating translation. * Content must be secured using role-based access control (RBAC), adhering to the principle of least privilege. * RBAC role assignments must be made to Azure Active Directory groups, not individual users. * AI solution responses must have a confidence score of 70% or greater. If the confidence score is lower than 70%, human input is required to improve the response. Chatbot requirements: * Provide answers to FAQs. * Escalate to customer service agents or members of the "management accountants" group if needed. * "Management accountants" can approve FAQs (requiring Q&A Maker Cognitive User role). * "Consultant accountants" can create and amend FAQs (requiring Q&A Maker Editor role). * Customer service agents can browse FAQs (requiring Q&A Maker Reader role). * Access for customer service agents must be managed using Omnichannel for Customer Service. * If the response confidence score is low, the chatbot must provide other response options to customers. Document processing requirements: * Documents should be less than 20 pages, formatted as PDF or JPEG. * Distinct standards (templates) are needed for each office (UK, France, Portugal). * The solution must extract tables and text from financial documents. * It must also extract information from receipt images. * Members of the "management bookkeeper" group should define how to extract tables from financial documents. * "Consultant bookkeepers" should be able to process financial documents. Knowledge base requirements: * Support searches for equivalent terms (synonyms, e.g., "automo" and "cars"). * Accurately transcribe industry-specific jargon. * Search content in different formats, including videos. * Provide relevant links to external resources for further research. **Question 1: Extract Solution for Receipt Images** To develop an extract solution for receipt images that meets document processing and technical requirements, using Form Recognizer API, and given a JSON output, which expression should trigger a manual review by a "consultant bookkeeper" if the confidence score is low? The JSON output shows a "receiptType" field with a confidence score of 0.676 and a "merchantName" field with 0.913. The general requirement is for human input if the confidence score is lower than 70% (0.7). The correct expression is `documentResults.fields.*.confidence < 0.7`. This expression checks if the confidence score of *any* detected field in the document results is less than 0.7, triggering a manual review as required. **Question 2: Knowledge Base - Searching Equivalent Terms** When developing the knowledge base using Azure Cognitive Search, to meet the requirement of searching for equivalent terms (e.g., "automo" and "cars"), what should be included in the solution? The options are: Synonym map, Suggesters, Custom analyzer, Built-in keyphrase extraction skill. The correct solution is a **Synonym map**. Synonym maps allow defining equivalent terms, ensuring that a search for one term (e.g., "car") also returns results for its synonyms (e.g., "automo"). Suggesters are for autocomplete, custom analyzers for text processing, and keyphrase extraction for identifying important phrases, none of which directly address equivalent term search. **Question 3: Knowledge Base - Processing Wiki Content** When developing the knowledge base using Azure Cognitive Search, to process wiki content and meet technical requirements (specifically multi-language support), what should be included in the solution? Wiki content is stored in Azure Cosmos DB, is text-based, and currently in only one language. The options involve indexers for Azure Blob Storage or Cosmos DB, attached to skill sets with various skills. The correct solution is **an indexer for Azure Cosmos DB attached to a skill set that contains the language detection skill and the text translation skill**. An indexer for Cosmos DB is necessary because the wiki content is stored there. The language detection skill is crucial to identify the original language of the wiki content, preventing unnecessary or incorrect translations. Once detected, the text translation skill can translate the content into English, French, and Portuguese, fulfilling the multi-language requirement. Document extraction skill is not needed as the content is already text. **Question 4: Knowledge Base - Building a Skill (Line 1)** When developing the knowledge base using Azure Cognitive Search, a skill needs to be built for indexers. Given a code snippet for a `text.v3.EntityRecognitionSkill`, which parameter should fill the blank for `categories` in Line 1, considering the `output` fields are `persons`, `locations`, and `organizations`? The `EntityRecognitionSkill` (version 3) allows specifying categories of entities to recognize. Since the desired outputs are "persons", "locations", and "organizations", these should be provided as categories. The correct options to fill the blank for `categories` are `person`, `organization`, and `location`. **Question 5: Knowledge Base - Building a Skill (Line 2)** Following the previous question, what should fill the blank for `output` in Line 2 of the skill definition? The `EntityRecognitionSkill` (version 3) is used. For version 3 of the `EntityRecognitionSkill`, the output field that encapsulates recognized entities from the specified categories (persons, locations, organizations) is `namedEntities`. The correct option is `namedEntities`. In previous versions, "entities" might have been an option, but for `v3`, it is `namedEntities`. **Question 6: Document Processing Workflow - Extracting Text from Financial Documents** To develop the document processing workflow and identify API endpoints to extract text from financial documents (not receipts), meeting document processing requirements, which two endpoints should be used? The options are: * `vision/v3.1/analyze` * `formrecognizer/v2.0/custom/models/{modelId}/analyze` * `formrecognizer/v2.0/prebuilt/receipt/analyze` * `vision/v3.1/describe` * `vision/v3.1/read/analyzeResults/{operationId}` The correct two endpoints are: 1. `vision/v3.1/analyze`: This endpoint from Computer Vision provides Optical Character Recognition (OCR) capabilities, suitable for extracting general text from images and documents, including financial documents. 2. `formrecognizer/v2.0/custom/models/{modelId}/analyze`: This Form Recognizer custom model endpoint is essential for extracting specific information, including tables and custom jargon, from financial documents with distinct standards for each office. This allows training models tailored to Contoso's unique financial document layouts. The `prebuilt/receipt/analyze` is specifically for receipts, `describe` generates image descriptions, and `read/analyzeResults` retrieves results from a prior `read` operation, not initiates extraction itself. **Question 7: Document Processing Workflow - Custom Table Extractor** To develop a solution for the "management bookkeepers" group to create a custom table extractor using the Form Recognizer sample labeling tool, which three actions should be performed in sequence? The process for training a custom model (including a table extractor) using the Form Recognizer sample labeling tool involves: 1. **Creating a new project and loading sample documents**: This sets up the environment and provides the data for training. 2. **Labeling the sample documents**: This step involves manually identifying and tagging the tables and other fields within the documents that need to be extracted. This is how the model learns what to extract. 3. **Training the custom model**: After labeling, the model is trained using the labeled data to recognize the defined fields and tables in new documents. The sequence is crucial. Creating a composite model (combining multiple custom models) is a subsequent step, not part of creating a *single* custom table extractor. The correct sequence of actions is: Create a new project and load sample documents, Label the sample documents, Train the custom model. **Question 8: Knowledge Base - Obtaining Transcripts of Webinars** To obtain transcripts of webinars using Azure Video Indexer (previously Video Analyzer for Media) and ensure the solution meets the knowledge base requirement for transcribing jargon with high accuracy, what should be done? The options are: * Create a custom language model. * Configure audio indexing for videos only. * Enable multi-language detection for videos. * Build a custom person model for webinar presentations. The knowledge base requires transcribing jargon with high accuracy. The standard language models in Azure Video Indexer might not accurately recognize industry-specific accounting jargon. Therefore, the correct action is to **create a custom language model**. A custom language model can be trained with Contoso's specific vocabulary and jargon, significantly improving the accuracy of transcription for their webinars. Configuring audio indexing only restricts content, multi-language detection is for general language identification, and a custom person model identifies speakers, none of which directly address jargon accuracy.