AB-410 Exam Questions

Cross-lingual embeddings map different languages into the same vector space, allowing the system to recognize that “Apple“ in English and “Manzana“ in Spanish are semantically similar.

Semantic Ranking significantly improves search relevance by understanding the intent and context behind a search query.

The Indexer is the engine that connects to data sources (like Blob Storage or SQL), extracts the data, and writes it into the index.

Embeddings models represent the semantic meaning of text as numbers, which allows a vector database to find contextually relevant matches.

The “Facetable“ attribute is specifically used to create hit counts by category, enabling the user to drill down into search results.

Azure AI Vision includes pre-trained models for domain-specific content like celebrities and landmarks.

Power Automate allows the creation of flows that can call APIs and return data back to the Copilot agent via a Plugin.

Dataverse Search provides fast, intelligent search capabilities across multiple tables, which is essential for grounding AI agents in Dataverse data.

Generative Answers allow the bot to “boost“ its knowledge by searching designated websites or SharePoint sites in real-time.

Video Indexer automatically transcribes audio, identifies speakers, and extracts key insights and timestamps from video content.

Azure OpenAI offers the same models as OpenAI but with added layers of Microsoft’s security, compliance, and data residency guarantees.

Plugins in Copilot Studio allow the agent to perform actions, such as triggering an automated workflow based on the conversation‘s context.

The GUID (Globally Unique Identifier) is the unique ID for every record in Dataverse, ensuring the AI references the exact intended row.

The Prebuilt Invoice Processing model is specifically trained to recognize and extract common fields like totals, dates, and vendor names from financial documents.

Higher temperature results in more creative/random output, while lower temperature (near 0) makes the model deterministic.

Top P is an alternative to Temperature that ensures the model only considers a subset of the most likely words, increasing focus.

Using the JSON file as a Knowledge Source (RAG) allows the bot to retrieve current data at query time without the high cost of retraining or fine-tuning.

Prompt Shields (part of Content Safety) detect and block malicious prompts designed to trick the AI into ignoring its safety filters.

Conversation Summarization is specifically tuned to extract key points and resolutions from multi-turn dialogues.

Floating Point or Decimal numbers provide the precision required to store and calculate geographic coordinates accurately.

The System Message (or System Prompt) sets the “ground rules“ for the AI, such as “You are a helpful assistant that only uses professional language.“

Fairness ensures that AI systems do not reinforce or introduce bias based on protected characteristics like gender.

Presence Penalty increases the model‘s likelihood of discussing new topics by penalizing tokens based on whether they have already appeared.

Slot Filling is the process where a bot identifies and collects specific pieces of information (entities) required to complete a task or answer a query.

Chain-of-Thought (CoT) prompting improves performance on complex tasks by forcing the model to articulate the logic leading to its conclusion.

OAuth2 with Entra ID allows for delegated permissions, enabling the bot to access user-specific data securely.

The System Message defines the persona and stylistic constraints the model must adopt during the conversation.

The Speech Translation feature within Azure AI Speech can ingest audio and provide both text-based and speech-based translation in near real-time.

Groundedness is the specific evaluation metric for checking if an answer is derived from the source material rather than hallucinated.

Document Translation is a specific feature that translates entire files while preserving the original formatting and structure.

Azure AI Vision‘s spatial analysis and object detection features are used to detect specific items like Personal Protective Equipment (PPE).

PII (Personally Identifiable Information) detection identifies sensitive data like account numbers, allowing them to be masked before the summary is generated.

Microsoft Purview provides data governance and auditing capabilities to track how sensitive data is used within AI applications.

The Prediction model analyzes historical patterns to forecast future outcomes, such as a Yes/No event like machine failure.

Negative prompting (e.g., “Do not mention competitor names“) is a constraint technique to refine the model‘s behavior.

Entities are used to “pick out“ specific information from natural language and save it for later use in the conversation.

Inclusiveness ensures that AI solutions are built to empower and serve everyone, including people with visual, hearing, or cognitive impairments.

Language Detection must happen first to identify the source language before the translation service can process it.

A higher confidence threshold reduces “false positives“ but may result in the bot triggering the Fallback topic more often if the user isn‘t precise.

The Azure AI Studio Playground allows developers to iteratively test, compare, and refine prompts across different models.

Row-level security ensures that the AI, acting on behalf of the user, only retrieves records that the specific user has permission to see.

Object Detection can be trained to recognize and locate specific products or components within an image.

Zero-shot prompting relies on the model‘s pre-trained knowledge to fulfill a request without being given any specific examples in the prompt.

The “Call an Action“ node triggers a Power Automate flow, which acts as the bridge to external APIs.

Variables act as the bot‘s short-term memory, allowing it to remember details provided by the user earlier in the session.

Rate limiting protects the infrastructure and manages costs by capping the number of requests sent to the AI service per second/minute.

Initial vector searches are fast but can be imprecise; re-ranking uses a cross-encoder model to fine-tune the relevance of the top results.

Custom Vision allows the company to train a model specifically to recognize types of vehicle damage and estimate severity.

PII (Personally Identifiable Information) detection identifies and can mask sensitive data points within a text string.

The Search Index is the optimized collection of data that the search engine queries to find relevant results quickly.

Microsoft Entra ID (formerly Azure AD) Managed Identities provide a secure way for apps to authenticate to services like Azure OpenAI using identity-based access.

Category Classification in AI Builder is designed to tag text into predefined categories based on historical examples.

The context window is the memory limit for a single prompt session; if exceeded, the model “forgets“ the beginning of the conversation.

The system message provides the “inner core“ instructions, such as “You are a professional customer support agent.“

Accountability ensures that there is human oversight and responsibility for the behavior and outcomes of AI systems.

Azure AI Health Insights (specifically Health Text Analytics) is optimized to recognize and structure complex medical entities.

The Retrieval Engine queries the vector index to find the most semantically similar data chunks to include in the AI‘s prompt context.

Video Indexer identifies faces and provides a timeline of every instance a specific person appears throughout the video duration.

Image Classification is used when you want to assign a label (like a car brand) to an entire image.

Setting the response format to json_object ensures the model generates valid JSON, which is essential for programmatic integration.