Your company creates a custom Azure Machine Learning model that uses a generative AI assistant. The model initially delivers strong results. However, six months later, the model predictions become noticeably less accurate. What is a possible cause of the issue?
Correct Answer: B
This phenomenon is known as data drift. It is one of the top reasons model accuracy declines after deployment. In the context of a generative AI assistant, data drift occurs when the statistical properties or distributions of the input data (the prompts or context provided by users) change significantly from what the model was originally trained on. Key Causes of Data Drift Changing User Behavior: Users may start interacting with the assistant differently, using new slang, jargon, or evolving topics of interest. Real-World Changes: External events (like a pandemic or economic shift) can suddenly make the model's training data outdated and irrelevant. Data Pipeline Issues: Changes in how data is collected, such as updated sensors or modified web forms, can alter the format or scale of inputs. Seasonality: Patterns may fluctuate based on the time of year, such as different holiday shopping behaviors. Impact on Generative AI For a generative assistant, data drift often manifests as: Reduced Relevance: Outputs feel outdated or fail to address contemporary topics. Increased Hallucinations: When faced with unfamiliar input patterns, the model may produce factually incorrect or nonsensical responses. Loss of Quality: Outputs may become less creative, more repetitive, or exhibit biased behavior. Reference: https://nexla.com/ai-infrastructure/data-drift
Question 17
Your company stores thousands of reports and documents across multiple systems. You recommend using Azure AI Search as part of a new generative AI solution to improve information discovery. What is a key benefit of using Azure AI Search in this scenario?
Correct Answer: B
In an environment with tens of thousands of reports and documents across multiple systems, Azure AI Search (formerly Cognitive Search) significantly improves information discovery through several core mechanisms: *-> Natural Language & Semantic Search: Unlike traditional keyword search, it understands the intent and context behind queries. Users can ask conversational questions (e.g., "Find all contracts mentioning GDPR compliance in 2023") and receive relevant results even without exact keyword matches. *-> Unified Multi-System Ingestion: It uses indexers to automatically pull and unify data from diverse sources such as SharePoint, Azure Blob Storage, SQL databases, and Cosmos DB into a single searchable index. AI-Powered Content Enrichment: During indexing, it can apply cognitive skills to extract information from unstructured data. This includes: - Optical Character Recognition (OCR) to make scanned reports searchable. - Entity Recognition to identify and tag people, locations, and organizations. - Key Phrase Extraction and language detection to enhance metadata. -Hybrid Retrieval: It combines vector search (for semantic meaning) with full-text search (for specific terms like product codes or names), merging them via Reciprocal Rank Fusion (RRF) to ensure high precision and recall. Semantic Ranking: An advanced L2 ranking layer uses deep learning models from Bing to re- order the top search results, ensuring the most contextually relevant answers appear first. This setup is commonly used as the retrieval foundation for Retrieval-Augmented Generation (RAG), where search results are fed into Large Language Models (LLMs) like GPT-4 to provide grounded, human-like answers based on your enterprise data. Reference: https://azure.microsoft.com/en-us/products/ai-services/ai-search
Question 18
Your company stores hundreds of internal business reports. You need to recommend a generative AI solution that uses an agent to answer questions based on the content in the reports. What should you include in the recommendation?
Correct Answer: C
A RAG-based generative AI solution for hundreds of internal reports uses an agent to query a vector database, ensuring answers are grounded in proprietary data, minimizing hallucinations. The system parses reports into embeddings, retrieves relevant chunks via semantic search, and uses an LLM to generate precise, cited answers. Key Components & Architecture Data Ingestion & Embedding: Convert thousands of PDFs, docs, or text files into vector embeddings stored in a vector database (e.g., Pinecone, Azure AI Search, Milvus). Agentic Workflow: Implement an intelligent agent that decomposes complex user questions into sub-queries, searches multiple data sources, and refines answers. Retrieval Mechanism: Use hybrid search (combining semantic and keyword search) for high accuracy in finding relevant report snippets. Generation & Grounding: The LLM receives the prompt with retrieved content to generate answers, improving quality and reducing errors. Incorrect: [Not D] The GAN can be used to improve the quality of document embeddings or to generate realistic, synthetic training data for the agent's semantic search, particularly in scenarios where data is unstructured or sparse.
Question 19
Your company uses a generative AI solution. You need to improve the quality of responses by using grounding. Which statement accurately describes how grounding improves accuracy and relevancy?
Correct Answer: C
Question 20
Your company has a Microsoft 365 subscription and uses Microsoft 365 Copilot Chat. Some users need to build and use declarative agents that can access work data. Which type of license should you recommend for the users?
Correct Answer: A
To use declarative agents that access work data (such as SharePoint or Graph connectors), users generally need a Microsoft 365 Copilot add-on license. While Microsoft 365 Copilot Chat is included in most business subscriptions at no extra cost, its native capabilities are primarily limited to web-grounding and basic instructions. Accessing organizational data via agents typically requires one of the following licensing paths: 1. Microsoft 365 Copilot Add-on License This is the most direct method. It grants full access to: Declarative agents grounded in tenant data without additional usage fees. Copilot Studio for authoring and managing these agents. Embedded Copilot features in Word, Excel, Outlook, and Teams. 2. Pay-As-You-Go (Consumption) Model If users do not have a full Copilot add-on license, organizations can enable metered usage (consumption-based billing). Usage-based billing: Interactions with agents that access tenant data (SharePoint, connectors) consume "Copilot credits". Requirement: This requires an Azure subscription and a billing policy set up in the Microsoft 365 admin center. Reference: https://support.microsoft.com/en-gb/topic/how-copilot-chat-works-with-and-without-a-microsoft- 365-copilot-license-5810b659-fbe0-48ee-9fe6-d731fe86cdeb
