AB-410 Exam Questions

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

Deflection Rate measures the percentage of sessions the AI handles successfully without escalation to a human agent.

Cosine similarity is used to determine the similarity of two vectors by measuring the cosine of the angle between them.

Reliability and Safety focus on making sure the AI system behaves as intended and does not cause harm during failures or edge cases.

PII Redaction identifies and removes or masks sensitive data, such as names and IDs, before data is logged or processed.

Knowledge Sources are the data repositories a Copilot agent uses to retrieve information and answer user questions without manually creating topics.

The On-premises Data Gateway serves as a bridge that enables fast and secure data transfer between on-premises data and Microsoft Cloud services.

Solutions are the Power Platform transport mechanism used to move apps, flows, and Copilot agents between environments.

Tokenization converts text into tokens, a format the transformer architecture can understand and manipulate.

Inclusiveness helps ensure AI applications empower everyone and meet the needs of diverse populations, regardless of ability or background.

Groundedness is the specific evaluation metric used to check whether an answer comes from the source material instead of being hallucinated.

Hallucinations happen when an LLM produces information that is factually wrong but seems plausible.

Negative prompting, such as “Do not mention competitor names,“ is a constraint technique used to refine model behavior.

Rate limiting protects infrastructure and controls costs by limiting the number of requests sent to the AI service per second or minute.

Entity Extraction is built to identify and extract specific data points, such as dates, names, or amounts, from unstructured text.

Transparency makes sure users understand the AI’s capabilities and limitations, including that they are communicating with a bot.

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

Hybrid search combines traditional keyword matching with modern vector search to return the most accurate retrieval results.

Setting response format to json_object makes the model generate valid JSON, which is required for programmatic integration.

Object Detection identifies that an item is present in an image and also returns its exact pixel coordinates, or bounding box.

The “Call an Action“ node starts a Power Automate flow, and that flow serves as the bridge to external APIs.

A higher confidence threshold lowers “false positives,“ but it can cause the bot to trigger the Fallback topic more often when the user is not precise.

Presence Penalty pushes the model toward new topics by penalizing tokens that have already appeared in the output.

Entra ID, formerly Azure AD, manages authentication and authorization so that only permitted users or apps can access the AI.

Top P is another option instead of Temperature; it makes the model consider only a subset of the most likely words, which increases focus.

The context window sets the memory limit for one conversation or prompt session.

Power Automate lets you create flows that call APIs and return data to the Copilot agent through a Plugin.

PII, or Personally Identifiable Information, detection identifies sensitive data points in a text string and can mask them.

Prompt Shields, part of Content Safety, identify and block malicious prompts that are designed to make the AI ignore its safety filters.

Transparency centers on the explainability of AI systems, so people can understand their logic and decision-making processes.

Frequency Penalty penalizes tokens according to how often they already appear in the text, which discourages repetitive output.

Generative Answers use an LLM to search specified data sources, such as a website or SharePoint, to answer questions that have not been explicitly programmed in the bot’s topics.

Custom Translator lets you train the translation engine with your own domain-specific terminology, such as legal or medical jargon, to improve accuracy.

The Orchestrator coordinates each step: receiving the user input, querying the search index, and sending the context to the LLM.

Copilot Studio serves as the orchestration layer for building the “brains” of agents, which can then be hosted in Power Apps or other channels.

The System Prompt functions as the AI’s internal instructions, telling it how to behave, such as using only professional language.

Embedding models transform text into high-dimensional vectors, or embeddings, so semantic similarity comparisons can be made.

Azure AI Content Safety is a dedicated service that checks prompts and completions for harmful content categories.

Custom Connectors let the Power Platform communicate with any service that exposes a public or private REST API.

RAG enables the model to retrieve current information from an external data source, such as the 2026 catalog, before it generates a response.

Security Roles and Business Units define data access scope, helping ensure AI agents follow organizational boundaries and permissions.

Temperature controls how creative or deterministic the model is; 0 is very literal, while 1.0 is very creative.

Azure AI Studio is the unified platform for building, evaluating, and deploying AI models, including open-source options through the Model Catalog.

Tokens are the core text units processed by LLMs; in English, one token is roughly equal to 4 characters.

Plugins let Copilot agents take actions or retrieve data from external systems through Connectors or Power Automate flows.

Azure AI Vision is used to extract text from the image through OCR, and Azure AI Translator converts that text into the target language.

Sentiment Analysis can classify feedback as positive, negative, or neutral, allowing the flow to mark negative feedback as urgent.

Grounding keeps the model’s responses tied strictly to the provided context, meaning the legal documents, and helps prevent hallucinations.

Semantic Kernel is the SDK built to connect LLMs with “skills“ and “connectors“ so they can work with real-world data and applications.

Microsoft Purview helps ensure that data used by or produced from AI apps is governed, classified, and compliant with privacy regulations.

Dataverse Plug-ins are custom code snippets that run server-side to manage complex data validation or external integrations.

Azure AI Document Intelligence is designed specifically to extract text, key-value pairs, and structural data from documents such as invoices.

Chain-of-Thought prompting encourages the model to “think step-by-step,“ which can improve performance on complex logical or mathematical tasks.

The “Filterable“ attribute enables the field to be used in $filter queries to narrow search results using specific criteria.

Copilot Studio lets you upload files, including Excel or PDFs, directly as Knowledge Sources so the bot can answer questions using that data.

Brand Detection is a specialized Azure AI Vision feature that can identify thousands of global logos and brands in images.

To authenticate and send requests, an application requires the specific endpoint, or API Base URL, and a valid API Key or Token.

Semantic Kernel is an open-source SDK that enables developers to combine AI services with conventional programming languages to build agents.

The Prediction model evaluates historical patterns to estimate the likelihood of future outcomes, such as Yes/No or binary results.

Generative AI models are strong at understanding long-form text and reducing it into concise summaries while preserving key facts.