AI-102 Practice Test

When an agent executes a tool and that tool fails, the orchestrator should catch the exception and return a textual error report (e.g., 'Error: Database connection timed out. Please try again.') to the model. Large language models can reason over these error responses, allowing them to inform the user politely or try a fallback strategy rather than crashing the client app.

Classic search is correct because Classic search best fits because Microsoft describes it as an index-first retrieval model for predictable, low-latency queries against a predefined search index. The cited source, Introduction to Azure AI Search, supports this answer for the Implement Knowledge Mining and Information Extraction Solutions scenario rather than the adjacent distractors.

Custom Neural Voice is a limited-access feature due to responsible AI risks like deepfakes and identity impersonation. To train a voice clone, Microsoft strictly requires a recorded, verbal consent statement or signed document from the voice actor whose voice will be cloned, indicating explicit authorization.

The OCR guidance says that cropping is not necessary for text lines. Microsoft explicitly notes that you can send the whole image to the Read API and it will recognize all text in the image.

Custom Blocklists with specific terms and custom match types is correct because Custom Blocklists in Azure AI Content Safety allow developers to define specific lists of words or phrases that should be blocked. The cited source, Use custom blocklists in Azure AI Content Safety, supports this answer for the Plan and Manage an Azure AI Solution scenario rather than the adjacent distractors.

This pattern is known as few-shot learning. The standard way to implement few-shot prompts in chat completions is to include a system prompt setting the context, followed by simulated conversational turns where the 'user' provides the query and the 'assistant' provides the expected output. Finally, the actual user input is appended as the last user message.

This architecture represents a multi-agent collaboration pattern (often structured as an Orchestrator-Worker or Router-Specialists pattern). Instead of having one massive model try to solve every problem, specialized agents are designed for specific tasks, and a router agent orchestrates the handoffs and manages high-level flow state.

The Document Intelligence prebuilt-invoice model is specifically trained to recognize and extract common fields (like vendor name, billing address, dates, tax, totals, and line-item tables) from standard billing invoices out-of-the-box. Using a prebuilt model requires zero training data or labeling effort.

Custom Question Answering allows you to add a pre-curated 'Chit-chat' dataset (available in various styles like Professional, Friendly, or Witty) directly to your knowledge base. This populates your KB with standard small talk pairs, saving you from writing answers for generic greetings and off-topic conversational lines.

The Read OCR engine is the best fit because Microsoft describes it as supporting printed and handwritten text across global languages. That is precisely the kind of input diversity the field app needs to handle.