AI-103 Practice Questions and Answers

Several teams at your company are each building their own AI app, but leadership wants shared governance: one place to manage connections, network rules, and access policies across all of them. What should you set up?

1. A separate Azure OpenAI resource for each team, managed independently
2. A Foundry hub with one project per team, connected to the shared hub ✓
3. A single Azure AI Language resource that every team calls
4. Separate resource groups with no shared hub

Tip: Whenever a question mentions sharing governance or connections across teams, think Foundry hub first.

You need your application to call Azure OpenAI without ever storing an API key in code, config files, or environment variables. Which setup meets this requirement?

1. Store the key in an environment variable and reference it at runtime
2. A managed identity with an Azure RBAC role assigned to the OpenAI resource ✓
3. Rotate the API key manually every 30 days
4. Hard-code the key and restrict the source file's permissions

Tip: "Keyless" in a question stem almost always points to managed identity plus RBAC.

You're adding a text-classification feature to a mobile app. It has to run on the device itself, work without an internet connection, and respond in under 200ms. Which type of model fits this requirement?

1. A large frontier LLM accessed through the Azure OpenAI API
2. A small language model (SLM) deployed locally on the device ✓
3. A multimodal model with image and audio support
4. An agent that calls multiple external tools

Tip: Offline, on-device, and low-latency in the same sentence is the classic signal for an SLM.

You're building a chatbot that must answer questions using your company's internal documentation, and those documents change weekly. What's the right approach to keep the bot's answers current?

1. Fine-tune the base model on the full document set
2. Configure retrieval-augmented generation (RAG) with vector or hybrid search over an indexed copy of the documents ✓
3. Increase the model's max token limit
4. Rely on keyword search only, without embeddings

Tip: Frequently changing source data plus a chatbot is almost always a RAG question, not a fine-tuning one.

Your company's product photo library needs descriptive alt text generated automatically so the site meets accessibility guidelines. Which capability should you configure?

1. Object detection with bounding boxes only
2. Azure AI Vision's dense captioning and alt-text generation ✓
3. OCR to extract any text printed on the product
4. A Custom Vision image classification model

Tip: Accessibility and alt-text in a prompt point to Vision's captioning features, not object detection or classification.

You're building a media catalog app that needs to identify and describe individual segments within uploaded video files, not just single frames. What should you configure?

1. Single-frame object detection run on sampled screenshots
2. A Content Understanding video analysis pipeline for segment-level interpretation ✓
3. Frame-by-frame OCR
4. A Custom Vision classification model applied per frame

Tip: "Segments" or "over time" in a video question usually rules out single-frame approaches.

A support bot needs to route each incoming customer message to the correct department based on what the customer is actually asking for. Which service should you use to build this?

1. Sentiment analysis on the incoming message
2. Conversational Language Understanding (CLU) with a custom intent model ✓
3. PII detection and redaction
4. Azure AI Translator

Tip: "Route based on what the customer wants" is intent classification, which means CLU.

A language-learning app needs to score how accurately a learner pronounces a phrase and highlight specific words that need work. Which Azure AI Speech feature fits this?

1. Text-to-speech synthesis
2. Pronunciation assessment ✓
3. Speech translation
4. Custom neural voice

Tip: Any question about scoring or grading spoken input points straight to pronunciation assessment.

Your finance team processes thousands of standard vendor invoices every month and needs the key fields (vendor, total, due date, line items) extracted automatically. What's the fastest way to implement this?

1. Train a custom extraction model from scratch on your invoice samples
2. Use the prebuilt invoice model in Azure AI Document Intelligence ✓
3. Run plain OCR and parse the text with regular expressions
4. Use Content Understanding for image classification only

Tip: Standard, common document types (invoices, receipts, ID cards) usually map to a prebuilt Document Intelligence model, not a custom one.

You're building a RAG pipeline that needs to answer questions from scanned contracts. Before indexing, you need a clean, structured representation of each contract that preserves layout and section context. What should you use?

1. Store the raw scanned PDFs directly in the vector index
2. Use Content Understanding to produce structured markdown output, then index that ✓
3. Run plain OCR and index the raw extracted text only
4. Skip extraction and embed the scanned images directly

Tip: When a question asks for structure preserved before indexing, Content Understanding's markdown output is the answer to look for.