AB-410 Exam Questions

Transparency ensures the system‘s logic is clear, often achieved through “Thought“ traces or citations in the response.

Application Insights tracks API failures and throttling, which is critical for maintaining service availability.

The On-premises Data Gateway allows Microsoft Cloud services to securely access data residing in local, private networks.

Dynamic chaining allows an agent to “reason“ about which tools it needs and in what order to complete a multi-step task.

Category Classification is the standard model for tagging and routing text-based content into predefined business buckets.

The context window is the model‘s active memory; exceeding it causes the model to “forget“ earlier parts of the chat.

Behavioral constraints like sentence length or persona are defined in the System Message.

Microsoft provides specialized dashboards to track the environmental impact, including energy and carbon, of cloud and AI services.

Hierarchical orchestration involves a lead agent that manages delegation and review of tasks performed by subordinates.

Zero-shot prompting relies on the model‘s pre-trained knowledge to fulfill a request based solely on the instruction.

The Canvas is the visual drag-and-drop environment used to orchestrate the conversational logic of an agent.

Managed Identities (part of Entra ID) provide an identity for applications to use when connecting to resources that support Entra authentication.

Grounding restricts the model‘s knowledge base to a specific, trusted set of documents to ensure accuracy and relevance.

The Vector Search Engine uses mathematical embeddings to find document chunks that match the “meaning“ of the user‘s query.

The Face API within Azure AI Vision provides face detection, recognition, and verification (1:1 matching).

Entities are used for “Slot Filling,“ helping the bot understand specific data points within a natural language sentence.

Plugins extend Copilot‘s capabilities, allowing it to “do“ things like update a CRM record or check stock in an ERP system.

Chunking is the process of breaking down large text files so they fit within the model‘s context window.

Azure OpenAI offers the same models with Microsoft’s layers of security, compliance, and private data residency.

The Orchestrator or Planner manages the logical sequence and delegation of tasks between specialized agents.

Prompt Shields are the primary defense mechanism in Content Safety to block inputs designed to bypass the model‘s safety filters.

Prompt compression removes redundant information from the context while retaining the semantic meaning for the LLM.

User-scoped variables allow the bot to remember a specific person‘s preferences even if they return to the bot days later.

Azure Cost Management allows for tagging and reporting to track AI spending across different business units.

CoT prompting encourages the model to break down complex tasks into logical steps before arriving at a conclusion.

Purview‘s AI Hub is the central location for discovering risks, governing data, and monitoring AI usage at scale.

Inclusiveness ensures that AI benefits everyone, including those with visual, auditory, or cognitive impairments.

RBAC ensures that the agent only accesses the specific tables and rows it has been authorized to see.

Document Translation is a specialized feature of the Translator service that maintains the physical layout of the file.

Plugins extend Copilot‘s capabilities, allowing it to interact with external data and execute transactional actions.

Few-shot prompting uses in-context examples to “teach“ the model the pattern or logic it should follow for a specific task.

Category Classification is used to automatically tag or label text-based content into pre-defined business categories.

Sentiment Analysis evaluates text to determine the prevailing emotional tone, which is critical for customer service routing.

Accountability involves creating a feedback loop where humans can oversee and correct the system‘s outputs.

Max Tokens sets a hard boundary on how many tokens the model can generate, directly affecting both cost and response length.

Video Indexer automatically extracts metadata, speech, and visual insights to create summaries and structural breakdowns of video content.

The search engine (like Azure AI Search) queries the vector index to find the data most semantically related to the prompt.

Power Automate flows act as the bridge between Copilot Studio and external systems via standard or custom connectors.

Document Intelligence has specialized models for high-accuracy OCR and field extraction from handwritten and printed forms.

Image Classification is used to categorize an entire image into specific labels, such as “Damaged“ or “Intact.“

RAG provides a way for the AI to query “live“ data sources, whereas fine-tuning only updates the model‘s static internal knowledge.

Azure AI Studio (and the Azure Portal) provides the necessary credentials and endpoint URLs for service integration.

The System Message defines how the model should behave, such as “You are a helpful assistant that only answers based on PDF context.“

Few-shot prompting guides the model‘s logic and formatting by demonstrating the desired behavior through examples.

Rollup columns aggregate data from related records, providing the AI with pre-calculated metrics directly from the table.

The system message acts as the foundational instruction that tells the AI who it is and the boundaries it must operate within.

Groundedness checks if the answer is factually rooted in the provided context rather than being fabricated by the model.

Prompt injection is a security threat where users try to trick the AI into ignoring safety filters or leaking sensitive data.

Reliability and Safety focus on ensuring the system handles errors gracefully and performs consistently under pressure.

Application Insights provides real-time telemetry and error tracking to ensure the operational health of the application.

SLMs offer high-quality reasoning in a smaller package, making them ideal for specific tasks or edge deployments.

Setting Temperature to 0 makes the model deterministic, ensuring consistency and precision in technical scenarios.

The Planner is the logical brain that maps out which agents and tools are needed to solve a complex request.

The Planner is the intelligence layer that determines the roadmap of actions required to achieve a high-level user goal.

Rate limiting (throttling) prevents API exhaustion and controls costs by capping the number of requests per minute.

The context window defines the memory limit for a single session; if the conversation exceeds this, the model begins to “forget“ earlier parts.

Groundedness and Accuracy measure how well the model‘s output aligns with the provided factual context or expert answers.

Global variables persist throughout the conversation, allowing data (like a User ID) to be used across multiple topics.

In-context learning allows the model to adapt to new tasks or styles simply by reading the instructions and examples in the current prompt.

MCP is an open standard that enables seamless integration between AI models and local or remote data/tools.