Free AI-103 Practice Questions with Detailed Explanations

A backend microservice needs to call your Azure OpenAI deployment to generate chat completions. It must not be able to view, modify, or delete deployments, and it must not be able to read the resource keys. You want the least-privilege role.

Which Azure RBAC role should you assign to the microservice's managed identity?

1. A.Cognitive Services OpenAI User — grants permission to call inference endpoints but not to manage deployments or read keys
2. B.Cognitive Services OpenAI Contributor — required to call the inference endpoint at all
3. C.Contributor — necessary to use the managed identity against the resource
4. D.Cognitive Services User — sufficient for any Azure AI service including Azure OpenAI

Your application generates marketing copy with Azure OpenAI. You must block any response that contains medium or higher severity violent content, while allowing low-severity violence (e.g., metaphors like "killer deal") to pass through. All other content categories should use default Microsoft-recommended settings.

How should you configure the content filter on the deployment?

1. A.Disable the violence category entirely and rely on Azure AI Content Safety post-processing
2. B.Set the violence output filter to block at medium severity, leaving hate, self-harm, and sexual at default settings
3. C.Set the violence input filter to block at medium severity, leaving the output filter at default
4. D.Enable prompt shields for direct jailbreaks and rely on those to block violent output

Your Azure AI Foundry hub must be reachable only from a specific virtual network. Outbound traffic from the hub to Azure OpenAI and Azure AI Search must also flow over the private network, not the public internet. Microsoft Entra ID authentication is required throughout.

Which network configuration achieves this?

1. A.Enable public network access on the Foundry hub and restrict the firewall to the corporate IP range
2. B.Configure the Foundry hub for managed virtual network isolation with the "Allow Internet Outbound" preset
3. C.Configure the Foundry hub for managed virtual network isolation with the "Allow Only Approved Outbound" preset, and create private endpoints to Azure OpenAI and Azure AI Search
4. D.Use NSGs on the VNet to block outbound traffic except to Microsoft public IPs

You are building a customer-support assistant that must hold multi-turn conversations, search a knowledge base for relevant articles, execute Python to compute order refunds when needed, and persist conversation context across days for the same user.

Which Azure AI Foundry primitive should you use?

1. A.A prompt flow with an LLM node and an AI Search lookup node
2. B.A direct chat completions call with custom function-calling logic
3. C.An Azure AI Foundry agent with file search and code interpreter tools, using threads to persist conversation context
4. D.A fine-tuned model trained on past customer-support conversations

Your company has 50,000 internal policy documents. You need an assistant that answers employee questions grounded in these documents. The documents update weekly and any answer must cite the source document.

Which approach is correct?

1. A.Fine-tune a base model on all 50,000 documents and retrain weekly
2. B.Build a RAG application with Azure AI Search over the document corpus, with integrated vectorization and citation of retrieved chunks
3. C.Pass all 50,000 documents in the system prompt to ensure the model has full context
4. D.Store documents in Cosmos DB and use SQL queries to retrieve them

You are designing an Azure AI Search index for a RAG application. The source content is PDFs in Azure Blob Storage. You need to chunk the documents, generate embeddings, and store both the chunks and embeddings — with the lowest operational overhead.

Which indexing approach is correct?

1. A.Pre-process the PDFs in Azure Functions, generate embeddings via the Azure OpenAI API, and bulk upload to the index
2. B.Use an Azure AI Search indexer with integrated vectorization, configured with a Blob data source, a text-split skill for chunking, and an Azure OpenAI embedding skill
3. C.Manually chunk the PDFs in PowerShell and upload via REST
4. D.Use Document Intelligence to extract content, then call the Azure OpenAI Assistants API to embed it

Your Foundry agent must look up customer order status and check warehouse inventory in parallel when a user asks "When will my order ship?". You want both backend calls to happen concurrently for latency reasons.

Which capability supports this?

1. A.Tool choice set to "required" with two function definitions
2. B.Parallel tool calls enabled, with two function definitions (get_order_status, check_inventory)
3. C.A single function that internally calls both backends
4. D.Two separate agents, one for each backend, called sequentially

Your application extracts customer complaint details from emails and inserts them into a database. The model must return a JSON object that exactly matches your database schema — no extra fields, no missing fields, no free-text wrapping.

Which response format option enforces this?

1. A.response_format = { "type": "text" } and instruct the model to "return only JSON"
2. B.response_format = { "type": "json_object" } — the JSON mode setting
3. C.response_format = { "type": "json_schema", "json_schema": { "name": "complaint", "schema": {...}, "strict": true } } — structured outputs
4. D.Fine-tune the model on a labeled dataset of complaint JSONs

You have built a prompt flow for medical-information Q&A. Before deploying, you must verify that responses are grounded in the retrieved source documents and never invent facts. You have a ground-truth evaluation dataset of 200 question/answer pairs with cited sources.

Which evaluation metric should you prioritize?

1. A.Fluency — measures grammatical quality of the answer
2. B.Relevance — measures how relevant the answer is to the question
3. C.Groundedness — measures whether the answer is supported by the retrieved sources
4. D.Similarity — measures lexical overlap with the ground-truth answer

Your real-time conversational agent must respond within 800ms end-to-end. The use case is a customer-service chatbot answering FAQ-style questions. Cost is a major constraint — you handle 5 million requests per month.

Which Azure OpenAI model is most appropriate?

1. A.GPT-4o — best general quality regardless of cost or latency
2. B.GPT-4o-mini — strong quality at significantly lower latency and cost than GPT-4o
3. C.o1 — chain-of-thought reasoning for complex problems
4. D.GPT-3.5-Turbo — lowest cost option

A user uploads a photo of a receipt. The application must extract the merchant name, date, line items, and total in structured JSON. The receipt format varies across thousands of merchants worldwide.

Which approach is most appropriate?

1. A.Azure AI Vision Image Analysis 4.0 with OCR (Read)
2. B.Azure AI Document Intelligence prebuilt receipt model
3. C.Custom Vision object detection trained on receipt images
4. D.GPT-4o with vision input and a structured-outputs schema

You need to detect whether warehouse shelves contain a specific custom product packaging in real-time camera feeds. The packaging is unique to your company and not represented in any general object-detection model.

Which Azure AI service is correct?

1. A.Azure AI Vision Image Analysis 4.0 with object detection
2. B.Azure AI Custom Vision with an object detection project, trained on labeled images of your packaging
3. C.Azure AI Document Intelligence custom extraction model
4. D.GPT-4o vision with a few-shot prompt showing the packaging

You want to verify a person's identity by comparing a live selfie to a passport photo for a banking onboarding flow. The user must prove they are physically present and not using a photo or video replay.

Which capability do you need, and what additional step is required?

1. A.Face Detect API — no additional steps
2. B.Face Identification with person groups — no additional steps
3. C.Face Liveness Detection — requires Microsoft Limited Access approval before use
4. D.GPT-4o with vision and a chain-of-thought prompt asking it to detect a live person

You need to extract structured data (insured party name, policy number, claim amount, incident date) from PDFs of insurance claim forms that mix typed and handwritten content. The schema is best expressed as natural-language field descriptions.

Which service is most appropriate?

1. A.Azure AI Vision Image Analysis 4.0 OCR
2. B.Azure AI Content Understanding with a custom analyzer defining each field in natural language
3. C.Custom Vision object detection
4. D.Translator document translation

A support ticket system receives free-text customer messages that may contain credit card numbers, passport numbers, and email addresses. You must redact PII before storing messages in the analytics warehouse.

Which Azure AI Language capability should you use?

1. A.Custom Named Entity Recognition trained on labeled PII examples
2. B.Prebuilt PII detection and redaction with the appropriate domain (Conversation) and PII categories enabled
3. C.Custom Text Classification to classify messages as PII-containing or not
4. D.Key Phrase Extraction filtered for credit-card-shaped strings

You are building a real-time, two-way voice agent where a user speaks to an AI in natural language and hears the AI's spoken response with sub-second turn latency. The agent must handle interruptions naturally.

Which API combination is correct?

1. A.Azure AI Speech SDK speech-to-text + Azure OpenAI chat completions + Azure AI Speech SDK text-to-speech
2. B.Azure OpenAI Realtime API with audio input and audio output
3. C.Azure AI Translator with speech translation
4. D.Azure AI Speech batch transcription + GPT-4o chat completions + neural TTS

You want to analyze product reviews not just for overall sentiment, but for sentiment on specific product aspects ("battery life: negative", "screen quality: positive"). Reviews are in English and contain multiple aspects per review.

Which capability provides this aspect-level breakdown?

1. A.Sentiment Analysis with confidence scores at document and sentence level
2. B.Sentiment Analysis with Opinion Mining (aspect-based sentiment) enabled
3. C.Custom Text Classification trained on labeled aspect-sentiment pairs
4. D.Key Phrase Extraction with sentiment post-processing

A RAG application searches over technical documentation. Users phrase queries in natural language (e.g., "how do I rotate certificates without downtime?"). The current pure-vector search returns relevant but sometimes poorly-ranked results.

Which Azure AI Search configuration improves ranking quality?

1. A.Switch to pure keyword search to avoid embedding drift
2. B.Use hybrid search (keyword + vector) with semantic ranker enabled to re-rank the top 50 results
3. C.Disable vector search and rely on BM25 only
4. D.Increase the embedding dimensions to 3072 by switching to text-embedding-3-large

Your company has a proprietary contract template used across thousands of vendor agreements. You need to extract 25 specific clauses and signature blocks from new contracts as they are uploaded. The general prebuilt Contract model is missing several clauses.

Which Document Intelligence approach is correct?

1. A.Use the prebuilt Contract model and post-process the output
2. B.Train a custom extraction model on 5–15 labeled samples of your contract template
3. C.Train a custom classification model only
4. D.Use Image Analysis OCR and parse with regex

You need to build a knowledge mining pipeline over scanned-PDF invoices in Azure Blob Storage. The pipeline must extract text via OCR, detect the language of each invoice, extract named entities (organizations, monetary values), and generate vector embeddings for semantic search.

Which Azure AI Search skillset design is correct?

1. A.A single Azure OpenAI skill that does everything
2. B.A skillset chain: OCR skill → Language Detection skill → Entity Recognition skill → Azure OpenAI Embedding skill, configured in the indexer
3. C.Separate Logic Apps for each stage
4. D.A single Document Intelligence custom model trained on all behaviors