Courseiva

Microsoft Certified: Azure AI Apps and Agents Developer Associate (AI-103) (AI-103) (AI-103) — Questions 1–75

455 questions total · 7 pages · All types, answers revealed

Take a mock examExam hub

Page 1 of 7

Page 2
1
MCQeasy

You need to ensure that an Azure AI services resource is accessible only from a specific virtual network. Which configuration should you update in the Azure portal?

A.The Networking blade to configure Private endpoints.
B.The Resource tags to add a network security group.
C.The Authentication blade to disable local keys.
D.The Identity blade to assign a User Assigned Managed Identity.
AnswerA

Private endpoints assign a private IP from your virtual network to the Azure AI services resource, removing public internet exposure. Configuring them on the Networking blade restricts inbound access to that specific VNet, satisfying the isolation requirement.

Why this answer

To restrict an Azure AI services (Cognitive Services) resource so it is reachable only from a specific virtual network, you configure a Private Endpoint on the resource's Networking blade. A Private Endpoint creates a private IP address from your VNet, mapping the service into your network via Azure Private Link, so public internet access can be disabled entirely. This is the standard portal-based approach for VNet-level isolation of Azure AI resources.

Exam trap

AI-103 often tests the misconception that tags, keys, or managed identities provide network isolation — only Private Endpoints (or service firewalls/VNet rules) actually restrict network reachability.

How to eliminate wrong answers

Option B is wrong because Resource tags are metadata labels used for organization, cost tracking, and policy scoping — they cannot attach a network security group or enforce any network restriction. Option C is wrong because disabling local (key-based) authentication only changes the authentication model to Entra ID; it does not restrict which network can reach the resource. Option D is wrong because assigning a User Assigned Managed Identity grants the resource an identity for outbound access to other services — it has no bearing on inbound network isolation.

Full explanation →
2
MCQhard

You are designing an agentic workflow where an agent needs to maintain conversational context and remember user preferences across multiple disparate sessions over several weeks. Which architectural component should you implement?

A.Increasing the model context window to 1,000,000 tokens for every request.
B.A persistent semantic memory store integrated with the agent's context pipeline.
C.Using a stateless function with zero memory persistence.
D.Storing the entire conversation history in a local client-side cookie.
AnswerB

A persistent semantic memory store retains embeddings of user preferences and prior interactions beyond a single session, then retrieves relevant fragments into the context pipeline at each turn. This satisfies the stem's requirement for continuity across disparate sessions spanning weeks, which ephemeral in-context history alone cannot provide.

Why this answer

Maintaining conversational context and user preferences across disparate sessions over weeks requires a persistent memory layer that survives beyond a single session or context window. A persistent semantic memory store integrated with the agent's context pipeline allows the agent to retrieve relevant past interactions and preferences via embeddings, enabling long-term personalization. This is the standard architectural pattern for long-horizon agent memory.

Exam trap

AI-103 often tests the misconception that a larger context window replaces the need for persistent memory, when in fact cross-session memory requires an external store regardless of context size.

How to eliminate wrong answers

Option A is wrong because even a 1,000,000-token context window is bounded and would require resending all history every request, which is cost-prohibitive, slow, and still cannot span weeks of sessions without exceeding limits. Option C is wrong because a stateless function with zero memory persistence explicitly cannot remember anything across sessions, which is the opposite of the requirement. Option D is wrong because storing conversation history in a client-side cookie is insecure, size-limited (typically 4KB), and not accessible server-side for semantic retrieval, making it unsuitable for long-term memory.

Full explanation →
3
Multi-Selecthard

You are designing an enterprise text analysis solution with high availability and security requirements. Which THREE practices should you implement? (Choose three.)

Select 3 answers
A.Store subscription keys in plain text configuration files inside the source code repository.
B.Implement exponential backoff retry logic in client applications to handle rate limiting (429 errors).
C.Authenticate API calls using managed identities instead of hardcoded subscription keys.
D.Disable TLS encryption on HTTP client requests to improve API throughput.
E.Configure private endpoints to restrict network access to your virtual network.
AnswersB, C, E

Exponential backoff spaces retries after HTTP 429 responses, letting the service recover from throttling instead of failing requests outright. This satisfies the high-availability requirement by keeping the text analysis solution responsive under rate limits rather than dropping calls.

Why this answer

Option B is correct because Azure AI services enforce rate limits and return HTTP 429 (Too Many Requests) responses; implementing exponential backoff retry logic lets client applications gracefully retry after the indicated delay, which is essential for high availability under load. Option C is correct because using managed identities (via Microsoft Entra ID) eliminates hardcoded subscription keys, providing secure, credential-free authentication that supports key rotation and least-privilege access. Option E is correct because private endpoints (Azure Private Link) restrict traffic to your virtual network, removing public internet exposure and meeting enterprise network security requirements.

Option A is not appropriate because storing subscription keys in plain text in a source code repository exposes secrets and violates security best practices. Option D is not appropriate because disabling TLS would remove encryption in transit, degrading security and violating compliance requirements rather than improving a legitimate design goal.

Exam trap

AI-103 often tests whether candidates recognize that security best practices (managed identity, private endpoints, TLS) and resilience patterns (exponential backoff) must be combined, while tempting shortcuts like plain-text keys or disabled TLS are always wrong.

Full explanation →
4
MCQmedium

Your organization is preparing for an annual external security audit. The auditor requests proof that all Azure AI resources comply with ISO 27001 standards. Where can you download the official compliance audit reports for Azure AI services?

A.Microsoft Service Trust Portal
B.Microsoft 365 Admin Center security center
C.Visual Studio subscription benefits page
D.Azure OpenAI Studio compliance tab
AnswerA

The Microsoft Service Trust Portal hosts downloadable audit reports, ISO 27001 certificates and compliance documentation for Azure services, including Azure AI. It directly satisfies the auditor's request for official proof of ISO 27001 compliance, unlike marketing pages or the Azure portal itself.

Why this answer

The Microsoft Service Trust Portal is the official Microsoft repository for compliance documentation, including ISO 27001 audit reports, SOC reports, and other regulatory attestations for Azure and Azure AI services. Auditors accept these reports as evidence of compliance because they are produced by independent third-party assessors. Access requires signing in with a Microsoft Entra ID account and accepting the non-disclosure agreement.

Exam trap

AI-103 often tests whether candidates confuse the Service Trust Portal (compliance evidence) with the Microsoft 365 Admin Center or Azure Portal security blades, which show posture and recommendations but do not provide official audit reports.

How to eliminate wrong answers

Option B is wrong because the Microsoft 365 Admin Center security center focuses on Microsoft 365 service health, security scores, and threat intelligence—not downloadable ISO 27001 audit reports for Azure AI. Option C is wrong because the Visual Studio subscription benefits page provides software licenses, Azure credits, and support benefits, not compliance attestations. Option D is wrong because Azure OpenAI Studio is a model development and deployment interface; it has no compliance tab and does not host audit reports.

Full explanation →
5
MCQmedium

You are developing a customer support agent using Microsoft Foundry. The agent needs to call a custom weather API tool when a user asks about local conditions. How should you define this capability in the agent configuration?

A.Register a custom function tool by providing an OpenAPI schema definition that describes the weather API endpoints, parameters, and expected responses.
B.Upload the compiled C# library containing the API client code directly into the agent code interpreter workspace.
C.Add the weather API endpoint URL as a grounding data source under the vector store settings.
D.Configure a managed data connection pointing to the weather API URL within Azure AI Search.
AnswerA

Registering a custom function tool with an OpenAPI schema gives the model a machine-readable contract for the weather API's endpoints, parameters and responses, letting it invoke the tool when users ask about local conditions. This satisfies the stem's requirement to define custom tool-calling capability.

Why this answer

Custom function tools in Microsoft Foundry agents are declared using an OpenAPI schema that describes the endpoint, HTTP methods, parameters, and response shape, allowing the model to invoke the weather API through tool calling. The schema acts as the contract the orchestrator uses to generate valid function calls and parse results back into the conversation. This is the standard, supported mechanism for exposing external REST APIs to an agent.

Exam trap

AI-103 often tests the confusion between grounding data sources (retrieval) and function tools (action), causing candidates to pick vector store or Azure AI Search options for what is really a tool-calling requirement.

How to eliminate wrong answers

Option B is wrong because uploading a compiled C# library into the code interpreter workspace does not register a callable tool for the agent's orchestration layer — code interpreter executes Python for data/code tasks, not arbitrary compiled .NET assemblies as agent tools. Option C is wrong because adding an API URL as a grounding data source under vector store settings treats it as retrieval content, not as an invocable action; vector stores index documents for semantic search, they do not execute live API calls. Option D is wrong because a managed data connection in Azure AI Search is used for indexing and querying data sources, not for defining agent tool invocations against a live weather endpoint.

Full explanation →
6
MCQeasy

You need to extract text from a multi-page PDF document using Azure AI Vision Read API. How are pages represented in the JSON response structure?

A.They are organized under a 'readResult.pages' or 'pages' array in the analyze result.
B.All text from all pages is merged into a single unstructured string without page separation.
C.Pages are returned as separate JPEG files in a ZIP archive.
D.Each page requires a separate API transaction starting with a new POST request.
AnswerA

The Read API's analyze result groups extracted content by page, exposing each under a 'readResult.pages' or 'pages' array, with lines and words nested beneath. This structure satisfies the stem's multi-page PDF scenario by keeping per-page results addressable.

Why this answer

The Read API (and the newer Image Analysis/Read model) returns an analyze result that includes a pages array (or readResult.pages in older SDK shapes), where each page contains lines, words, and their bounding polygons plus angle and dimensions. This page-level structure lets you map extracted text back to specific pages of a multi-page PDF.

Exam trap

The trap is assuming the Read API requires one call per page or returns a flat string; candidates overlook the asynchronous single-call, page-array response model.

How to eliminate wrong answers

Option B is wrong because the Read API does not merge all pages into one unstructured string; it preserves page boundaries with per-page line and word arrays. Option C is wrong because the API returns JSON, not a ZIP of JPEGs; image rendering is a separate concern and not part of the Read response. Option D is wrong because a single POST to the Read/analyze endpoint processes the entire multi-page PDF asynchronously and returns all pages in one result — you do not need one transaction per page.

Full explanation →
7
MCQeasy

Your development team needs to test a new Azure AI Language service feature without incurring ongoing charges when the service is idle. Which billing model is appropriate for initial prototyping?

A.Pay-as-you-go (Consumption-based metering)
B.Dedicated hardware hosting
C.Reserved Instance 3-year term
D.Commitment tier (Pre-purchased capacity)
AnswerA

Consumption-based metering charges only for transactions and records processed, so idle periods incur no cost. This suits prototyping where usage is sporadic and unpredictable, avoiding the committed spend of provisioned throughput while still allowing full feature testing.

Why this answer

Pay-as-you-go (consumption-based metering) charges only for the transactions and features actually used, with no upfront commitment or idle cost, making it ideal for prototyping a new Azure AI Language feature. It lets the team experiment freely and stop incurring charges when the service is idle. This is the standard billing model for initial development and testing.

Exam trap

AI-103 often tests the misconception that commitment tiers or reserved instances save money in all cases — candidates forget that idle or low-volume prototyping incurs charges under pre-purchased models.

How to eliminate wrong answers

Option B is wrong because dedicated hardware hosting provisions always-on infrastructure with fixed costs regardless of usage, which is the opposite of avoiding idle charges. Option C is wrong because a 3-year Reserved Instance requires a long-term upfront commitment and is designed for steady-state production workloads, not prototyping. Option D is wrong because a commitment tier pre-purchases capacity for a fixed term, incurring charges even when the service is idle.

Full explanation →
8
MCQeasy

What is the primary purpose of the 'selection marks' detection in the layout model?

A.To OCR printed text
B.To detect checkboxes and radio buttons
C.To identify table borders
D.To find handwritten signatures
AnswerB

Selection marks detection identifies the state of boxes and radio buttons in documents, returning whether each is selected or unselected. This lets the layout model expose form-field choices as structured data rather than plain text, distinct from general OCR of characters.

Why this answer

Selection marks in Azure AI Document Intelligence's layout model detect and extract checkboxes and radio buttons (and their selected/unselected state) from documents. The model returns each mark with a confidence score and a state of 'selected' or 'unselected', which is essential for processing forms, surveys, and questionnaires. OCR, table borders, and signatures are handled by other capabilities.

Exam trap

AI-103 often tests the distinction between layout model sub-capabilities, tempting candidates to confuse selection marks with OCR text extraction or table detection — the key is that selection marks specifically return checkbox/radio state.

How to eliminate wrong answers

Option A is wrong because OCR of printed text is the general read/OCR capability, not the specific purpose of selection marks. Option C is wrong because table border detection is part of the layout model's table extraction, not selection marks. Option D is wrong because handwritten signature detection is a separate capability (and often custom-model territory), not what selection marks provide.

Full explanation →
9
Multi-Selectmedium

Your organization is planning to deploy multiple Azure AI services and needs to implement robust cost management and governance practices. Which TWO actions should you take? (Choose two.)

Select 2 answers
A.Implement Azure Policy to enforce mandatory cost-tracking tags on all AI resource deployments.
B.Deploy all development, test, and production AI resources into a single shared resource group to simplify administration.
C.Disable diagnostic logging across all AI resources to reduce Azure storage and log ingestion costs.
D.Organize AI resources into separate resource groups and subscriptions based on business units and environments.
E.Use the Free (F0) pricing tier for all production workloads to eliminate infrastructure costs.
AnswersA, D

Azure Policy enforces tag requirements at deployment, denying or auditing resources lacking cost-tracking tags. This guarantees consistent metadata across all AI services, which is the prerequisite for filtering and attributing costs in Azure Cost Management, satisfying the mandatory governance constraint.

Why this answer

Option A is correct because Azure Policy can enforce tagging requirements (such as cost center, owner, or environment tags) at deployment time, and consistent tags are the foundation for cost tracking, chargeback, and filtering in Azure Cost Management and Billing. Option D is correct because separating AI resources into distinct resource groups and subscriptions by business unit and environment provides clear cost boundaries, enables per-subscription budgets and access control, and prevents dev/test spend from being commingled with production. Option B is not appropriate because consolidating all environments into one resource group removes cost and access separation and makes it hard to attribute spend or apply environment-specific governance.

Option C is not appropriate because disabling diagnostic logging sacrifices auditability, security monitoring, and troubleshooting, and cost reduction should not come at the expense of required telemetry. Option E is not appropriate because the F0 free tier has strict call and transaction limits and no production SLA, so it cannot support production AI workloads.

Exam trap

AI-103 often tests whether candidates recognize that cost governance requires both tagging enforcement and structural separation, and it plants 'reduce cost by disabling logging' and 'use free tier in production' as tempting but harmful distractors.

Full explanation →
10
MCQhard

You are automating document processing. You want to trigger a workflow only after a Document Intelligence operation finishes. What is the standard way to implement this?

A.Poll the 'operation-location' URL
B.Use a fixed-delay thread sleep
C.Use a WebSocket
D.Wait for a push notification
AnswerA

Document Intelligence returns an operation-location URL when a long-running analysis begins. Polling that URL is the standard completion mechanism, since the service exposes no push callback. Polling satisfies the stem's requirement to trigger the workflow only after the operation finishes.

Why this answer

The Document Intelligence analyze operation is asynchronous: it returns a 202 Accepted response with an 'Operation-Location' header containing a URL. The standard way to detect completion is to poll that URL until the status field returns 'succeeded' or 'failed'. This is the documented pattern for all long-running Document Intelligence operations.

Exam trap

The trap is assuming the service pushes a notification or that a simple sleep suffices — candidates may overlook that Document Intelligence uses the standard long-running operation polling pattern.

How to eliminate wrong answers

Option B (Use a fixed-delay thread sleep) is wrong because a fixed sleep does not guarantee the operation has finished and can either waste time or proceed prematurely. Option C (Use a WebSocket) is wrong because Document Intelligence does not expose a WebSocket interface for operation completion. Option D (Wait for a push notification) is wrong because the service does not push completion events to the client — the client must poll the operation-location URL.

Full explanation →
11
MCQmedium

You are processing customer feedback emails that contain both English and Spanish paragraphs within the same message. You need to perform sentiment analysis on each paragraph separately. What should you do?

A.Use custom text classification instead of sentiment analysis.
B.Submit the entire email as a single document with the language set to auto-detect.
C.Split the email into individual paragraphs and submit them as separate documents in the request payload.
D.Translate the entire email to French before calling sentiment analysis.
AnswerC

Azure AI Language evaluates sentiment per document, so mixed-language paragraphs must be submitted as separate documents to obtain independent scores. Splitting the email into paragraph-level documents ensures each language is assessed on its own content rather than aggregated.

Why this answer

Azure AI Language sentiment analysis processes each document independently and assumes a single language per document. When an email contains mixed languages, submitting the entire email as one document with auto-detect will produce inaccurate results because the service detects one dominant language and applies that language's model to all text. Splitting the email into individual paragraphs and submitting them as separate documents in the request payload allows the service to detect and analyze the language of each paragraph independently, yielding accurate per-paragraph sentiment.

Exam trap

AI-103 often tests the misconception that auto-detect language handles mixed-language documents correctly, when in fact the service assumes one language per document and requires splitting for per-paragraph analysis.

How to eliminate wrong answers

Option A is wrong because custom text classification is used to categorize text into custom labels, not to perform sentiment analysis; it does not solve the mixed-language problem. Option B is wrong because submitting the entire email as a single document with auto-detect forces the service to choose one language for the whole document, so Spanish paragraphs may be analyzed with an English model or vice versa, producing incorrect sentiment scores. Option D is wrong because translating the entire email to French does not address the need to analyze each paragraph separately and introduces translation errors; sentiment analysis should be performed on the original text in its native language for accuracy.

Full explanation →
12
MCQmedium

You have trained a Custom Vision classification model. You need to export the model to run locally on an edge device using ONNX format. Which project domain must you select when creating the project to support export?

A.General (compact)
B.Retail
C.Food (compact) for edge deployment
D.General
AnswerA

The General (compact) domain uses a smaller exported model footprint while retaining classification accuracy, and it is the domain that supports ONNX export for edge deployment. Selecting it satisfies the stem's requirement to run the Custom Vision model locally on an edge device.

Why this answer

To export a Custom Vision classification model to ONNX for edge deployment, the project must be created with a compact domain — specifically 'General (compact)' — because only compact domains support model export. The compact domain produces a smaller, optimized model suitable for local/edge inference.

Exam trap

AI-103 often tests the distinction between full and compact domains, so candidates pick 'General' (the full domain) or invent a domain like 'Food (compact)' instead of selecting the compact variant that actually supports ONNX export.

How to eliminate wrong answers

Option B is wrong because the Retail domain is a full domain optimized for retail scenarios and does not support export to ONNX. Option C is wrong because 'Food (compact) for edge deployment' is not a valid Custom Vision project domain name; the actual compact domains are General (compact), Landmark (compact), and Retail (compact), and Food is not one of them. Option D is wrong because the standard General domain is a full domain that does not support export; only the compact variant does.

Full explanation →
13
MCQmedium

You need to ensure that an AI model deployment is resilient to regional outages. What should you implement?

A.A single large VM.
B.Azure Front Door with regional endpoints.
C.Azure Storage redundancy.
D.Local VNET peering.
AnswerB

Azure Front Door is a global layer-7 load balancer that health-probes and fails over traffic between regional endpoints, so a regional outage routes users to a surviving region. This directly satisfies the resilience requirement, unlike single-region deployment or zone redundancy alone.

Why this answer

Azure Front Door is a global Layer 7 load balancer and CDN that can route traffic across multiple regional AI endpoints, performing health probes and automatic failover if a region becomes unavailable. By fronting regional endpoints with Front Door, the deployment remains reachable even during a regional outage. This is the standard Azure pattern for cross-region resiliency of HTTP-based AI inference services.

Exam trap

AI-103 often tests whether candidates confuse data redundancy (Storage replication) with compute/endpoint resiliency, leading them to pick storage options when the question is about keeping an inference endpoint reachable during a regional outage.

How to eliminate wrong answers

Option A is wrong because a single large VM is a single point of failure — if its region or host fails, the deployment is entirely unavailable, and vertical scaling does not provide regional redundancy. Option C is wrong because Azure Storage redundancy (LRS/ZRS/GRS) protects stored data durability, not the compute endpoint serving AI inference, so it does nothing for regional outage resilience of the model. Option D is wrong because local VNET peering only connects networks within or across regions for private connectivity; it does not provide traffic distribution, health probing, or failover for an endpoint.

Full explanation →
14
MCQhard

You are using Azure AI Foundry to evaluate a generative AI assistant that summarizes customer call transcripts. Reviewers complain that some summaries include details never present in the transcript. You need an automated metric that quantifies this specific defect during evaluation runs. Which metric should you use?

A.Groundedness
B.Relevance
C.Fluency
D.F1 score
AnswerA

Groundedness measures whether the generated response is supported by the provided source context, which is exactly the defect reviewers reported when summaries contain details absent from the transcript. Using it as an automated evaluation metric will surface and quantify those unsupported statements across the test set.

Why this answer

Groundedness is designed to detect whether statements in a generated response are supported by the supplied source content. Because the reported defect is fabricated details absent from the transcript, this metric directly quantifies the problem and is the right choice for the evaluation run.

Exam trap

The trap here is selecting a general quality metric like relevance or fluency when the specific defect is unsupported content, which only groundedness measures.

Full explanation →
15
MCQmedium

You are developing a custom extraction model using Azure AI Document Intelligence Studio. Your training dataset consists of multi-page invoices with varying structures, and some fields appear only on the final page. How should you structure your tagging configuration?

A.Create a separate custom model for the final page and chain them using an orchestration model.
B.Split every multi-page invoice into individual single-page documents before uploading.
C.Define all late-appearing fields as optional metadata parameters in the training configuration JSON file.
D.Tag the fields on whatever page they appear across the sample documents in Document Intelligence Studio.
AnswerD

Tagging each field on the page where it actually occurs lets the custom neural model learn page-position context, which matters because some fields appear only on the final page of these variably structured invoices. Tagging every field on page one would misrepresent that layout and degrade extraction accuracy.

Why this answer

In Document Intelligence Studio, you tag fields on the pages where they actually appear across your sample documents — the labeling tool is page-aware and stores the page number alongside each field's bounding region. This lets the trained model learn that, for example, 'Total' appears on page 1 in some invoices and page 3 in others. You do not need to split documents or build orchestration models just because fields appear on different pages.

Exam trap

AI-103 often tests whether candidates over-engineer solutions — the trap is assuming multi-page documents with late-appearing fields require splitting or orchestration, when the Studio natively supports tagging across pages in one model.

How to eliminate wrong answers

Option A is wrong because chaining a separate final-page model via orchestration adds unnecessary complexity and cost; a single custom model can learn fields across all pages of a multi-page document. Option B is wrong because splitting multi-page invoices into single-page files destroys the document context the model uses to associate fields (e.g., matching a line item to its invoice header) and multiplies labeling effort. Option C is wrong because Document Intelligence custom models do not support declaring fields as 'optional metadata parameters' in a JSON config — fields are defined by tagging them in the Studio UI, and the model learns optionality from the training data itself.

Full explanation →
16
MCQeasy

You are deploying an Azure OpenAI resource and want to authenticate client applications using Microsoft Entra ID instead of static API keys. Which role must be assigned to the application's managed identity to permit it to invoke model deployments?

A.Storage Blob Data Contributor
B.Key Vault Secrets Officer
C.Cognitive Services OpenAI User
D.Contributor
AnswerC

Cognitive Services OpenAI User grants the managed identity data-plane access to invoke model deployments on the Azure OpenAI resource. Assigning it at the resource scope lets the application authenticate via Microsoft Entra ID tokens, removing the need for static API keys.

Why this answer

The Cognitive Services OpenAI User role is the RBAC role specifically scoped to grant data-plane access for invoking Azure OpenAI model deployments (chat completions, embeddings, etc.) without granting management permissions. Assigning it to the application's managed identity lets the app call the inference endpoints using Entra ID tokens instead of API keys. This follows the least-privilege principle for OpenAI resource consumption.

Exam trap

AI-103 often tests the distinction between management-plane roles (Contributor, Owner) and data-plane roles (Cognitive Services OpenAI User), tricking candidates into picking Contributor because it sounds powerful enough to 'do everything.'

How to eliminate wrong answers

Option A is wrong because Storage Blob Data Contributor only grants read/write access to Azure Blob Storage containers and has no relationship to Azure OpenAI inference endpoints. Option B is wrong because Key Vault Secrets Officer manages secrets in Key Vault — it does not authorize calls to model deployments. Option D is wrong because Contributor is a management-plane role that can create and manage resources but does not grant the data-plane action Microsoft.CognitiveServices/accounts/OpenAI/deployments/chat/completions/action required to invoke models.

Full explanation →
17
MCQhard

You are calling the Azure AI Language REST API for Named Entity Recognition (NER) and receive an HTTP 429 status code. How should your application handle this error?

A.Switch the API endpoint to use the Free pricing tier.
B.Regenerate the subscription key and restart the client service.
C.Immediately re-submit the request with doubled payload size.
D.Implement an exponential backoff retry policy and resend the request after a delay.
AnswerD

HTTP 429 signals throttling, meaning the request exceeded the resource's rate limit. Retrying immediately worsens the condition, so exponential backoff spaces retries progressively, letting the quota window recover before resending. This is the prescribed handling for transient rate-limit responses from the Azure AI Language API.

Why this answer

HTTP 429 indicates 'Too Many Requests', meaning you have exceeded the rate limit or quota for the Azure AI Language service. The correct handling is to implement an exponential backoff retry policy, which gradually increases the delay between retries to avoid overwhelming the service. This is a standard practice for transient throttling errors.

Exam trap

AI-103 often tests error handling, and candidates may choose to regenerate keys or switch tiers instead of implementing retry logic for 429 errors.

How to eliminate wrong answers

Option A is wrong because switching to the Free tier would not resolve throttling; it would likely have stricter limits. Option B is wrong because regenerating the subscription key is unnecessary and does not address rate limiting; it would cause service disruption. Option C is wrong because immediately re-submitting with a larger payload would exacerbate the throttling and likely result in another 429.

Full explanation →
18
MCQhard

You are deploying a model that requires significant GPU resources. You want to ensure the deployment only scales when the average CPU load exceeds 70%. Which feature of Azure AI managed endpoints do you configure?

A.The 'Load Balancer' settings in the VNET.
B.Azure Front Door health probe settings.
C.The 'Endpoints' deployment configuration 'autoscaling' settings.
D.The 'Compute' instance restart policy.
AnswerC

Autoscaling settings on the deployment let you define a metric-based rule, such as average CPU utilisation exceeding 70%, so Azure AI managed endpoints add or remove instances only when that threshold is breached. This directly satisfies the stem's requirement to scale on CPU load rather than a fixed instance count.

Why this answer

Azure AI managed online endpoints support autoscaling rules configured on the deployment, where you define a metric (such as average CPU utilization) and a threshold (for example, 70%) that triggers scale-out or scale-in. This is configured in the deployment's 'autoscaling' settings under the endpoint configuration, allowing the deployment to add or remove instances based on load.

Exam trap

AI-103 often tests whether candidates confuse network-layer features (load balancers, Front Door, NSGs) with the deployment-level autoscaling configuration that actually controls instance count based on metrics.

How to eliminate wrong answers

Option A is wrong because load balancer settings in a VNET control traffic distribution across backend pools, not the scaling trigger for an Azure AI managed endpoint. Option B is wrong because Azure Front Door health probes determine backend health for global HTTP routing, not autoscale thresholds for a model deployment. Option D is wrong because a compute instance restart policy governs recovery behavior after failures, not the metric-based scaling logic of an endpoint deployment.

Full explanation →
19
Multi-Selecthard

Which THREE fields are typically returned by the prebuilt-receipt model?

Select 3 answers
A.MerchantName
B.EmployeeID
C.TransactionDate
D.PatientName
E.Total
AnswersA, C, E

MerchantName is one of the three fields the prebuilt-receipt model extracts, satisfying the stem's requirement for typical receipt output. The model returns merchant identification alongside transaction totals and line items, drawing the vendor name directly from the receipt header. This field appears in the model's default schema without custom training or configuration.

Why this answer

The prebuilt-receipt model in Azure AI Document Intelligence is designed to extract common receipt fields, and MerchantName (A) is one of them because it identifies the vendor or business that issued the receipt. TransactionDate (C) is also returned since the model recognizes the date the purchase occurred, which is a standard receipt field. Total (E) is included as the model extracts the final amount paid, a core field on virtually every receipt.

EmployeeID (B) is not a standard receipt field and is not returned by this model, as it relates to internal staff identifiers rather than receipt data. PatientName (D) is specific to healthcare or medical documents and is not part of the prebuilt-receipt schema.

Exam trap

AI-103 often tests whether candidates can distinguish fields returned by prebuilt-receipt versus other prebuilt models like prebuilt-idDocument or prebuilt-healthInsuranceCard, so mixing up domain-specific fields is a common trap.

Full explanation →
20
MCQeasy

You are setting up budget alerts for your Azure AI services to prevent unexpected overspending. Which Azure tool should you use to create these alerts and automatically trigger an action group when the budget threshold reaches 90%?

A.Azure Monitor alerts
B.Azure Cost Management + Billing budgets
C.Microsoft Defender for Cloud pricing settings
D.Azure Advisor cost recommendations
AnswerB

Azure Cost Management + Billing budgets track spend against a scope and threshold, and can invoke action groups when usage reaches a set percentage such as 90%. This delivers the automated alerting and triggered response the scenario requires.

Why this answer

Azure Cost Management + Billing budgets are the native mechanism for defining spending thresholds (e.g., 90% of a monthly budget) and attaching action groups that fire when thresholds are crossed. Budgets support scopes like subscriptions, resource groups, and management groups, and can trigger email, Logic App, or Automation Runbook actions via action groups. This directly satisfies the requirement to alert and act at 90% of budget.

Exam trap

AI-103 often tests the confusion between cost governance (Cost Management budgets) and operational monitoring (Azure Monitor alerts) — candidates pick Azure Monitor because 'alerts' sounds right, but budgets are the only native cost-threshold mechanism.

How to eliminate wrong answers

Option A is wrong because Azure Monitor alerts operate on metrics and logs (CPU, latency, custom metrics), not on monetary budget consumption — there is no native 'cost' metric in Azure Monitor for budget thresholds. Option C is wrong because Microsoft Defender for Cloud pricing settings configure which resources are protected and at what tier, not spending alerts. Option D is wrong because Azure Advisor cost recommendations are advisory suggestions (e.g., resize underutilized VMs) and do not create threshold-based alerts or trigger action groups.

Full explanation →
21
MCQeasy

You are testing an image analysis application in Python. You want to retrieve object bounding boxes using Azure AI Vision 4.0. Which feature string should you pass to the client?

A.VisualFeatures.TAGS
B.VisualFeatures.CAPTION
C.VisualFeatures.OBJECTS
D.VisualFeatures.READ
AnswerC

VisualFeatures.OBJECTS requests object detection, returning bounding boxes with labels and confidence scores for each detected object. This directly satisfies the stem's requirement to retrieve bounding boxes from Azure AI Vision 4.0, whereas CAPTION, TAGS and READ return descriptions, classifications or text instead of spatial coordinates.

Why this answer

In the Azure AI Vision 4.0 SDK, object detection is requested by passing VisualFeatures.OBJECTS to the analyze call, which returns bounding boxes and labels for detected objects. The enum value maps to the 'objects' feature string on the wire. This is distinct from TAGS (whole-image labels), CAPTION (single-sentence description), and READ (OCR text).

Exam trap

AI-103 often tests confusion between VisualFeatures.TAGS (whole-image labels) and VisualFeatures.OBJECTS (localized bounding boxes), causing candidates to pick TAGS when boxes are required.

How to eliminate wrong answers

Option A is wrong because VisualFeatures.TAGS returns a list of content tags for the whole image without bounding boxes, so it cannot satisfy the requirement for object bounding boxes. Option B is wrong because VisualFeatures.CAPTION returns a natural-language description of the image, not localized object coordinates. Option D is wrong because VisualFeatures.READ performs optical character recognition and returns text lines and words, not object detection boxes.

Full explanation →
22
MCQeasy

You are configuring a Microsoft Foundry project to deploy a multimodal model. You need to ensure that the model deployment scales automatically based on request volume while minimizing management overhead. Which deployment type should you select?

A.Standard deployment
B.Local container deployment
C.Provisioned Throughput deployment
D.Batch deployment
AnswerA

Standard deployment supports autoscaling based on request volume, adjusting capacity automatically without manual intervention. It minimises management overhead while meeting the scaling requirement, unlike provisioned or serverless options that demand capacity planning or lack the needed flexibility.

Why this answer

A Standard deployment in Microsoft Foundry (Azure AI Foundry) for a multimodal model provides automatic scaling based on request volume while minimizing management overhead. It is a pay-as-you-go deployment type that dynamically allocates compute resources to handle traffic, making it ideal for variable workloads without the need for manual provisioning or capacity planning.

Exam trap

AI-103 often tests the confusion between deployment types that scale automatically (Standard) and those that require manual capacity management (Provisioned Throughput), causing candidates to select Provisioned Throughput for scaling scenarios.

How to eliminate wrong answers

Option B is wrong because a local container deployment runs the model on your own infrastructure (e.g., Docker or Kubernetes), requiring significant management overhead for scaling, patching, and monitoring, and it does not automatically scale based on request volume in the cloud. Option C is wrong because Provisioned Throughput deployment reserves a fixed amount of throughput capacity, which does not automatically scale with request volume and requires manual capacity planning, increasing management overhead. Option D is wrong because Batch deployment is designed for asynchronous processing of large volumes of data, not for real-time automatic scaling based on request volume; it is optimized for offline batch jobs.

Full explanation →
23
Multi-Selecthard

Your organization is establishing a governance framework for Azure AI resources. You need to ensure cost control, tagging compliance, and security posture management across all subscriptions. Which THREE Azure services or features should you deploy? Each correct answer presents part of the solution.

Select 3 answers
A.Microsoft Defender for Cloud for security posture and compliance assessments
B.Azure Bastion for database query optimization
C.Azure Policy for automated tag enforcement and region restriction
D.Azure Cost Management budgets with alerts for spending thresholds
E.Azure Traffic Manager for database scaling
AnswersA, C, D

Microsoft Defender for Cloud continuously assesses resources against security benchmarks and regulatory standards, surfacing misconfigurations and recommendations. This satisfies the stem's security posture management requirement across all subscriptions, providing the compliance assessment layer the governance framework needs.

Why this answer

Azure Cost Management budgets with alerts (D) directly address cost control by letting you set spending thresholds per subscription and trigger notifications when actual or forecasted costs exceed them. Azure Policy (C) enforces tagging compliance and region restrictions by evaluating resource properties against built-in or custom policy definitions and applying deny, audit, or modify effects automatically. Microsoft Defender for Cloud (A) provides the security posture management piece through continuous secure score assessments, regulatory compliance dashboards, and recommendations across all subscriptions.

Azure Bastion (B) is a managed jump-host service for secure RDP/SSH access to VMs, not a database query optimization tool, so it does not fit. Azure Traffic Manager (E) is a DNS-based global traffic load balancer for distributing user traffic across endpoints, not a database scaling mechanism, so it is also irrelevant here.

Exam trap

AI-103 often tests whether candidates can distinguish genuine governance pillars (policy, cost, security posture) from unrelated Azure networking or access services that sound plausible but serve entirely different purposes.

Full explanation →
24
MCQmedium

Your company requires that all Azure AI service endpoints are accessible only from within a specific Azure Virtual Network (VNet) and that public internet access is completely disabled. How should you configure the Azure AI resource networking settings?

A.Set the routing preference to Microsoft network routing and enable service endpoints on the default subnet.
B.Disable public network access on the Networking blade and create a Private Endpoint associated with your VNet and subnet.
C.Configure the firewall to allow only the corporate public IP address range and keep public network access enabled.
D.Deploy an Azure Application Gateway in front of the Azure AI multi-service account endpoint.
AnswerB

Disabling public network access removes all internet-facing connectivity, while a Private Endpoint assigns a private IP from your VNet subnet, so traffic reaches the Azure AI resource over the Microsoft backbone and satisfies the VNet-only constraint.

Why this answer

Disabling public network access and creating a Private Endpoint associated with the VNet and subnet ensures the Azure AI resource is reachable only through private IP addresses inside the virtual network. This fully removes public internet exposure, which matches the requirement.

Exam trap

AI-103 often tests the difference between service endpoints and private endpoints, so the trap is choosing service endpoints or firewall IP allowlists when the requirement is to completely disable public access.

How to eliminate wrong answers

Option A is wrong because service endpoints still allow traffic over the Azure backbone but do not disable public network access, and routing preference does not restrict exposure. Option C is wrong because allowing a corporate public IP range keeps public network access enabled, violating the requirement. Option D is wrong because an Application Gateway is a public-facing reverse proxy and does not by itself disable public access to the AI endpoint.

Full explanation →
25
MCQmedium

You need to extract data from a document that includes handwritten signatures and printed text. Which capability is required?

A.Object Detection
B.Handwritten OCR (Read)
C.Custom Template model
D.Form Recognizer 2.0 API
AnswerB

Handwritten OCR (Read) satisfies the stem's requirement to extract both handwritten signatures and printed text, since the Read model handles mixed handwriting and print within one document. Standard printed-text OCR alone cannot recognise the cursive strokes of signatures, so Read's handwriting support is the specific capability needed here.

Why this answer

Handwritten OCR (Read) is the Azure AI Vision capability that recognizes both printed and handwritten text in documents and images, making it the required feature when signatures and handwriting are present. The Read API returns extracted text with bounding boxes and confidence, supporting mixed print and handwriting. This directly satisfies the requirement to extract data from documents containing handwritten signatures and printed text.

Exam trap

AI-103 often tests the confusion between OCR/Read (text recognition including handwriting) and custom extraction models, causing candidates to pick a custom template model when the question is specifically about recognizing handwritten text.

How to eliminate wrong answers

Option A is wrong because Object Detection identifies and localizes objects (e.g., cars, people) in images; it does not perform text recognition or handwriting extraction. Option C is wrong because a Custom Template model extracts structured fields from fixed-layout forms using labeled examples, but it relies on the underlying OCR for text recognition and does not by itself provide handwriting recognition as a distinct capability. Option D is wrong because Form Recognizer 2.0 API is a legacy version with limited handwriting support and has been superseded by newer Document Intelligence models; it is not the correct capability to cite for handwritten OCR.

Full explanation →
26
MCQeasy

Which tool provides a graphical user interface for labeling documents for custom models?

A.Azure CLI
B.Azure Storage Explorer
C.Document Intelligence Studio
D.Visual Studio Code
AnswerC

Document Intelligence Studio provides a graphical interface for labelling documents, training custom extraction models, and testing them without code. This satisfies the stem's requirement for a GUI labelling tool, unlike SDKs or REST APIs that require programmatic work.

Why this answer

Document Intelligence Studio is the web-based graphical interface provided by Azure that lets you visually label documents, define fields, train custom models, and test them against sample files. It is specifically designed for the labeling and project-management workflow of custom extraction and custom classification models, so it is the correct answer for a GUI-based labeling tool.

Exam trap

The trap here is confusing general-purpose Azure tools (Azure CLI, Storage Explorer, VS Code) with the purpose-built Document Intelligence Studio — candidates who have not used the Studio may assume any Azure GUI can label documents, but only Document Intelligence Studio provides the field-labeling and training workflow.

How to eliminate wrong answers

Option A is wrong because Azure CLI is a command-line tool for managing Azure resources — it has no graphical labeling interface for documents. Option B is wrong because Azure Storage Explorer is a GUI for browsing and managing blobs, queues, and tables in storage accounts; it can upload training documents but cannot label fields or train Document Intelligence models. Option D is wrong because Visual Studio Code is a code editor — while extensions exist for Azure, it does not provide the document-labeling GUI that Document Intelligence Studio offers.

Full explanation →
27
MCQeasy

You are developing a solution that must extract structured fields such as vendor name, invoice total, and due date from standard supplier invoices. The invoices come from many vendors but follow common invoice layouts. You want the fastest path to production without building or training a model. Which Azure AI Document Intelligence capability should you use?

A.A custom neural model trained on your invoice samples
B.A custom template model trained on your invoice samples
C.The prebuilt read model with a custom post-processing parser
D.The prebuilt invoice model
AnswerD

The prebuilt invoice model is purpose-built to extract common invoice fields such as vendor name, invoice total, due date, and line items without any training. It works across many vendors and standard layouts, matching the requirement for a fast, no-training path to production for supplier invoices.

Why this answer

The prebuilt invoice model is designed to extract common invoice fields across vendors without training. It satisfies the requirement for a fast, no-training deployment for standard supplier invoices. Custom models and the read model add effort or lack semantic understanding, so they do not meet the stated goal.

Exam trap

The trap here is reaching for a custom model by default when a prebuilt model already covers the required invoice fields.

Full explanation →
28
MCQmedium

You need to extract information from a business card. Which prebuilt model should you select?

A.prebuilt-businessCard
B.prebuilt-taxDocument
C.prebuilt-idDocument
D.prebuilt-layout
AnswerA

The prebuilt-businessCard model is trained specifically to extract fields such as contact names, job titles, company names, email addresses, and phone numbers from business card images, returning structured JSON. General document or receipt models lack these card-specific field schemas.

Why this answer

Azure AI Document Intelligence provides a dedicated prebuilt-businessCard model that extracts fields such as contact names, company, job title, phone numbers, email addresses, and website from business card images. Selecting this model avoids the need to train a custom model for a very common document type. It is the purpose-built choice for the scenario.

Exam trap

AI-103 often tests prebuilt model selection by name — candidates confuse prebuilt-idDocument (passports/licenses) with prebuilt-businessCard (contact info), or pick prebuilt-layout assuming it extracts semantic fields when it only returns raw structure.

How to eliminate wrong answers

Option B is wrong because prebuilt-taxDocument (e.g., prebuilt-tax.us.1040) is designed for US tax forms and returns IRS-specific fields like filer name, SSN, and income lines — irrelevant to business cards. Option C is wrong because prebuilt-idDocument extracts identity fields (name, date of birth, document number, expiration) from passports, driver's licenses, and similar IDs, not contact details from cards. Option D is wrong because prebuilt-layout only returns the general structure — text, tables, selection marks, and reading order — without semantic field extraction, so it would not identify 'Company' or 'Job Title' as named fields.

Full explanation →
29
MCQmedium

Your development team needs to deploy an Azure AI Search service and connect it to an Azure SQL Database containing source documents. Security requirements state that credentials must not be stored in application code. Which authentication method should you use between Azure AI Search and Azure SQL Database?

A.Shared Access Signature (SAS) token
B.Hardcoded SQL administrator password in indexer configuration
C.Anonymous access connection string
D.Managed Identity
AnswerD

Managed Identity lets Azure AI Search authenticate to Azure SQL Database through Microsoft Entra ID, issuing tokens from the platform-assigned identity rather than embedding secrets. This directly satisfies the stem's constraint that credentials must not reside in application code, since no password or connection-string secret is stored or rotated by your team.

Why this answer

Managed Identity allows Azure AI Search to authenticate to Azure SQL Database without storing credentials in code. The indexer can use the search service's managed identity to access the SQL database, provided the identity has been granted appropriate permissions in SQL.

Exam trap

AI-103 often tests the misconception that SAS tokens or connection strings with passwords are acceptable for secure authentication, when in fact managed identities are the recommended approach for Azure services to authenticate without secrets.

How to eliminate wrong answers

Option A is wrong because SAS tokens are for Azure Storage, not Azure SQL Database, and they still require managing secrets. Option B is wrong because hardcoding passwords violates security requirements. Option C is wrong because anonymous access is not supported for Azure SQL Database and would be insecure.

Full explanation →
30
Multi-Selectmedium

When designing an Azure AI Foundry project architecture, which TWO resources are typically linked to the Azure AI Hub to enable full model deployment and search capabilities? (Choose TWO)

Select 2 answers
A.Azure AI Search service
B.Azure SQL Database relational server
C.Azure OpenAI service
D.Azure Stream Analytics job cluster
E.Power BI embedded analytics workspace
AnswersA, C

Azure AI Search provides the vector and keyword indexing that grounds models in your own data. Linking it to the Azure AI Hub satisfies the search capability requirement, enabling retrieval-augmented generation and agent knowledge tools within the Foundry project.

Why this answer

Option A (Azure AI Search service) is correct because an Azure AI Search resource is a standard linked dependency of an Azure AI Hub/project, providing the vector and keyword indexing backend needed for search, retrieval, and RAG-style grounding scenarios. Option C (Azure OpenAI service) is correct because linking an Azure OpenAI resource to the hub supplies the deployed GPT and embedding models that the project uses for model deployment and inference. Option B (Azure SQL Database) is not a typical hub-linked resource; relational data can be connected separately but it is not required for model deployment or search.

Option D (Azure Stream Analytics) is a real-time stream processing service unrelated to Foundry hub model/search enablement. Option E (Power BI embedded) is a visualization/analytics service and plays no role in linking model deployment or search capabilities to an AI Hub.

Exam trap

AI-103 often tests whether candidates can distinguish AI-native linked resources (Azure OpenAI, Azure AI Search) from general-purpose Azure services (SQL Database, Stream Analytics, Power BI) that are not part of the AI Hub's core resource topology.

Full explanation →
31
Multi-Selectmedium

When configuring role-based access control (RBAC) for an Azure AI Services account, which TWO built-in roles are appropriate for assigning to users who only need to invoke AI models without managing the underlying resource infrastructure? Each correct answer presents part of the solution.

Select 2 answers
A.Owner
B.Cognitive Services OpenAI User
C.Contributor
D.Cognitive Services User
E.Storage Blob Data Reader
AnswersB, D

Cognitive Services OpenAI User grants inference access to OpenAI models without permitting resource management, satisfying the stem's constraint of invoking models only. It excludes control-plane actions such as creating or deleting the Azure AI Services account.

Why this answer

Option B, Cognitive Services OpenAI User, is correct because it grants the ability to call Azure OpenAI model inference endpoints (such as chat completions and embeddings) without granting permissions to manage the account, keys, or deployments. Option D, Cognitive Services User, is correct because it allows invoking the broader set of Azure AI Services (Cognitive Services) endpoints for inference while excluding resource-management actions. Both roles satisfy the least-privilege requirement of invoking AI models without managing underlying infrastructure.

Option A, Owner, is incorrect because it grants full control including managing resources and assigning roles, far exceeding inference-only needs. Option C, Contributor, is incorrect because it allows creating and managing resources (though not role assignment), which is more than invoking models requires. Option E, Storage Blob Data Reader, is incorrect because it only permits reading blob data and has no relevance to invoking AI model endpoints.

Exam trap

AI-103 often tests the distinction between control-plane roles (Owner, Contributor) and data-plane inference roles (Cognitive Services User, Cognitive Services OpenAI User), tempting candidates to pick Contributor as 'enough' when least privilege demands a data-plane role.

Full explanation →
32
MCQeasy

You want to detect whether corporate logos appear in user-submitted images using Azure AI Vision. Which visual feature parameter should you request?

A.features=brands
B.features=domain-specific
C.features=objects
D.features=tags
AnswerA

Requesting features=brands invokes Azure AI Vision's brand detection model, which returns identified corporate logos with name and confidence score for each detected instance. This directly satisfies the stem's requirement to detect whether corporate logos appear in submitted images, rather than describing image content or extracting text.

Why this answer

Azure AI Vision's Analyze Image API supports a 'brands' visual feature that detects corporate logos and brands within an image, returning brand names and confidence scores. Requesting features=brands enables this specific detection capability, which is distinct from general object or tag detection.

Exam trap

AI-103 often tests whether candidates confuse 'brands' with 'domain-specific' or 'tags', tricking them into selecting a broader feature when the question specifically asks for corporate logo detection.

How to eliminate wrong answers

Option B is wrong because 'domain-specific' refers to specialized models like celebrities or landmarks, not brand/logo detection. Option C is wrong because 'objects' detects common objects (e.g., car, person) but not corporate logos or brands. Option D is wrong because 'tags' returns general content tags describing the image but does not specifically identify corporate logos.

Full explanation →
33
MCQmedium

Your application processes images submitted by users. You want to ensure that images containing inappropriate or explicit content are rejected before further processing. Which Azure AI Vision feature should you evaluate first?

A.Custom Vision Object Detection
B.Image Analysis Read feature
C.Analyze Image with features=adult
D.Azure AI Language Content Moderation
AnswerC

Analyze Image with features=adult returns adult and racy classification scores, letting your application reject explicit submissions before further processing. This directly satisfies the stem's requirement to filter inappropriate content at ingestion, since the adult parameter exposes the moderation flags Microsoft Entra ID-authenticated Vision calls generate for each image.

Why this answer

The Analyze Image API with features=adult returns adult and racy content scores (0-1) plus gory and suggestive flags, which is exactly what you need to reject inappropriate images. This is the built-in content moderation capability of Azure AI Vision. Custom Vision and OCR do not perform content moderation, and Azure AI Language Content Moderation is a text-based service, not an image service.

Exam trap

AI-103 often tests service boundaries — the trap is selecting Azure AI Language Content Moderation for image moderation, when that service only handles text; image moderation belongs to Azure AI Vision's adult feature.

How to eliminate wrong answers

Option A is wrong because Custom Vision Object Detection classifies objects in images based on your own trained labels — it has no built-in adult-content detection unless you train it, which is not the 'evaluate first' built-in feature. Option B is wrong because the Read feature performs OCR to extract text from images; it does not analyze visual content for appropriateness. Option D is wrong because Azure AI Language Content Moderation analyzes text, not images — it would be the right choice for moderating user-submitted text, not images.

Full explanation →
34
MCQmedium

You are building a generative AI application that summarizes long legal documents. Users complain that summaries omit important clauses located in the middle of the document due to the model's attention limitations over extremely long texts. What phenomenon does this represent, and how can you mitigate it?

A.Token overflow error; mitigate by increasing Azure AI Search semantic ranker depth
B.Lost in the middle phenomenon; mitigate using map-reduce or recursive summarization patterns
C.Embedding drift; mitigate by retraining the text-embedding-ada-002 model
D.Prompt injection; mitigate by enabling Azure AI Content Safety prompt shields
AnswerB

The lost-in-the-middle phenomenon describes degraded recall for content positioned mid-context, which explains why middle clauses are omitted. Map-reduce or recursive summarisation splits the document into smaller chunks summarised separately, then combines results, ensuring middle sections receive adequate attention rather than being skipped.

Why this answer

The 'lost in the middle' phenomenon describes how transformer models attend more reliably to information at the beginning and end of long contexts, degrading recall for content in the middle. Map-reduce or recursive summarization mitigates this by chunking the document, summarizing each chunk independently, then combining the summaries — ensuring middle content is processed with full attention.

Exam trap

AI-103 often tests the misconception that long-context models attend uniformly to all tokens — candidates must recognize the U-shaped attention curve and choose chunking-based mitigations rather than retrieval or safety features.

How to eliminate wrong answers

Option A is wrong because a token overflow error would cause an API failure (context length exceeded), not silent omission of middle clauses; semantic ranker depth in Azure AI Search affects retrieval ranking, not summarization attention. Option C is wrong because embedding drift refers to vector representations becoming stale relative to new data distributions, which affects retrieval relevance, not summarization attention over a single long input; retraining ada-002 is not a supported customer action. Option D is wrong because prompt injection is an adversarial attack where malicious input hijacks model behavior, unrelated to attention degradation over long legitimate documents; prompt shields defend against injection, not summarization quality.

Full explanation →
35
Multi-Selectmedium

You are configuring an Azure AI Language service resource in the Azure portal. Which THREE settings or blades are available for managing the resource? (Choose three.)

Select 3 answers
A.Virtual machine size selector
B.Container registry integration blade
C.Keys and Endpoint
D.Pricing tier
E.Identity
AnswersC, D, E

The Keys and Endpoint blade exposes the subscription keys and the REST endpoint URL used to authenticate and call the Language service. It is a standard management blade on the resource, satisfying the stem's requirement for available Azure portal settings.

Why this answer

Option C, Keys and Endpoint, is correct because every Azure AI Language resource exposes a Keys and Endpoint blade in the portal where you retrieve the subscription key(s) and the REST endpoint URL used to authenticate and call the service. Option D, Pricing tier, is correct because the Azure AI Language resource is created on a pricing tier (for example Free F0 or Standard S) that determines quotas, throughput, and billing, and it can be viewed or changed in the portal. Option E, Identity, is correct because the resource supports managed identities, and the Identity blade lets you enable a system-assigned identity or assign user-assigned identities for keyless authentication to other Azure services.

Option A, Virtual machine size selector, does not belong because Azure AI Language is a fully managed PaaS cognitive service and does not provision or size VMs. Option B, Container registry integration blade, does not belong because, although Azure AI Language offers Docker containers, the resource itself does not include a container registry integration blade in the portal.

Exam trap

AI-103 often tests whether candidates confuse PaaS Cognitive Service blades (Keys and Endpoint, Pricing tier, Identity) with IaaS or container-platform blades (VM size, container registry), causing them to select irrelevant options.

Full explanation →
36
Multi-Selecthard

You are troubleshooting a RAG application in Azure AI Foundry where retrieved search results are irrelevant to user queries. Which THREE actions should you take to diagnose and improve retrieval quality? (Choose THREE)

Select 3 answers
A.Evaluate whether document chunk sizes and overlaps align with query granularity
B.Inspect embedding vector distances and similarity scores of retrieved documents
C.Increase the Azure OpenAI model temperature parameter to 1.5
D.Configure Azure Key Vault access policies for the search service
E.Enable semantic ranking in Azure AI Search to re-rank results based on deep language understanding
AnswersA, B, E

Chunk size and overlap determine the granularity of retrievable units. When chunks are far larger or smaller than the query's information need, embeddings capture diluted or fragmented meaning, so irrelevant passages surface. Aligning chunking with query granularity directly addresses the stem's irrelevant-retrieval symptom.

Why this answer

Option A is correct because chunk size and overlap directly control retrieval granularity: if chunks are too large they dilute the embedding with irrelevant content, and if too small they lose context, so aligning them with typical query granularity improves relevance. Option B is correct because inspecting embedding vector distances and similarity scores (e.g., cosine similarity in Azure AI Search) reveals whether the query embedding is actually close to the retrieved documents or whether the embedding model, index, or query vectorization is mismatched. Option E is correct because enabling semantic ranking in Azure AI Search applies a deep language model re-ranker on top of the initial BM25/vector results, promoting semantically relevant documents and typically raising retrieval precision.

Option C is not appropriate because temperature controls generative randomness, not retrieval quality, and 1.5 is outside the supported range for most Azure OpenAI models. Option D is not relevant because Key Vault access policies govern secret management, not the relevance of search results.

Exam trap

AI-103 often tests whether candidates confuse generation parameters (temperature, top_p) with retrieval quality controls — the trap is picking temperature as a fix for bad retrieval, when retrieval and generation are separate stages with separate tuning knobs.

Full explanation →
37
MCQmedium

You need to export logs from your Azure AI service to an external SIEM (Security Information and Event Management) system. What is the best way to achieve this?

A.Copy the files manually from Blob Storage.
B.Configure a diagnostic setting to stream logs to an Event Hub.
C.Use an Azure Logic App to poll the REST API.
D.Enable public access on the log file store.
AnswerB

Diagnostic settings are the native mechanism for routing Azure AI service logs to external destinations. Streaming to an Event Hub provides a scalable ingestion endpoint that the SIEM consumes, satisfying the requirement without custom polling or storage intermediaries.

Why this answer

Configuring a diagnostic setting to stream logs to an Event Hub is the standard, scalable way to export Azure AI service logs to an external SIEM. Event Hubs acts as a high-throughput ingestion buffer that the SIEM (e.g., Splunk, Sentinel) consumes via its connector. This is the recommended integration pattern for near-real-time security monitoring.

Exam trap

AI-103 often tests the misconception that manual copying or API polling is a valid log-export method — the designed mechanism is diagnostic settings streaming to Event Hub, Log Analytics, or Storage.

How to eliminate wrong answers

Option A is wrong because manually copying files from Blob Storage is not automated, not real-time, and does not scale — it is an operational anti-pattern for SIEM ingestion. Option C is wrong because polling a REST API with a Logic App is inefficient, introduces latency, and is not how Azure diagnostic logs are designed to be exported. Option D is wrong because enabling public access on the log store exposes logs to unauthorized access and is a security risk, not an export mechanism.

Full explanation →
38
Multi-Selectmedium

Which TWO monitoring strategies are effective for detecting 'data drift' in a deployed AI model?

Select 2 answers
A.Tracking changes in input data distribution.
B.Analyzing model prediction accuracy against ground truth.
C.Checking the physical temperature of the CPU.
D.Monitoring total API request volume.
E.Restarting the endpoint every hour.
AnswersA, B

Tracking changes in input data distribution directly detects covariate shift, the statistical divergence between live inference inputs and the training baseline. Monitoring feature-level distributions, such as population stability index or KL divergence, satisfies the scenario's requirement to surface data drift before prediction quality degrades.

Why this answer

Option A is correct because data drift is fundamentally a shift in the statistical distribution of the model's input features relative to the training data, so tracking changes in input data distribution (e.g., via statistical tests like KL divergence or PSI on feature histograms) directly detects it. Option B is correct because monitoring prediction accuracy against ground truth labels reveals performance degradation that typically results from drift, providing a downstream signal that the input-output relationship has changed. Option C is not relevant since CPU temperature is a hardware health metric and has no bearing on statistical data drift.

Option D does not belong because total API request volume only measures traffic load, not changes in data characteristics or model behavior. Option E is incorrect because periodically restarting the endpoint is an operational action, not a monitoring strategy, and does nothing to detect drift.

Exam trap

AI-103 often tests the distinction between data drift (input distribution change) and operational metrics (latency, volume, CPU), so candidates pick traffic or infrastructure metrics that sound like monitoring but say nothing about drift.

Full explanation →
39
MCQhard

You are building an agent that needs to execute Python code safely to calculate financial projections requested by users. Which tool in Azure AI Agent Service should you enable to allow the agent to write and run code in an isolated environment?

A.Code Interpreter tool
B.Azure Container Instances dynamic scaling group
C.Azure Functions webhook integration
D.Azure AI Search vector embedding tool
AnswerA

Code Interpreter runs Python in a Microsoft-managed sandbox, so untrusted calculations execute without touching your infrastructure. It satisfies the isolation constraint directly: the agent writes and executes code in that sandboxed session, returning results and files. Other tools, such as function calling or file search, cannot execute arbitrary Python.

Why this answer

The Code Interpreter tool in Azure AI Agent Service provides a managed, sandboxed Python execution environment where the agent can write and run code to perform calculations such as financial projections. It handles file uploads, execution, and output retrieval without exposing your infrastructure. This is the native tool designed for safe code execution within an agent.

Exam trap

AI-103 often tests whether candidates substitute generic Azure compute services (ACI, Functions) for the purpose-built Code Interpreter tool, so pick the option that names the native sandboxed execution capability.

How to eliminate wrong answers

Option B is wrong because Azure Container Instances dynamic scaling groups are a general compute construct, not an agent tool, and they do not provide the managed code-execution semantics the agent needs. Option C is wrong because Azure Functions webhook integration is a serverless invocation pattern, not a built-in agent tool for running arbitrary Python in a sandbox. Option D is wrong because Azure AI Search vector embedding is a retrieval mechanism for grounding, not a code execution environment.

Full explanation →
40
MCQhard

You are designing a high-throughput, low-latency architecture for an Azure OpenAI application. Clients are distributed globally across North America, Europe, and Asia. To ensure minimal latency and high availability, you decide to deploy multiple Azure OpenAI instances in different regions. How should you route incoming client requests to the nearest optimal region?

A.Azure ExpressRoute global reach
B.Azure Bastion global routing profile
C.Azure Local Traffic Manager (LTM)
D.Azure Front Door with latency-based routing
AnswerD

Azure Front Door performs latency-based routing at the global edge, directing each client to the Azure OpenAI instance with the lowest measured latency across North America, Europe and Asia. This satisfies the minimal-latency and high-availability constraints by combining anycast edge termination with health-probed regional back ends.

Why this answer

Azure Front Door is a global, scalable entry point that uses Microsoft's global edge network to route client requests based on latency. It continuously measures latency to the origin and directs traffic to the lowest-latency healthy backend, which is ideal for globally distributed Azure OpenAI instances. This ensures minimal latency and high availability.

Exam trap

AI-103 often tests the misconception that ExpressRoute or VPN solutions are used for global client routing, when in fact Azure Front Door or Traffic Manager are the correct services for latency-based routing.

How to eliminate wrong answers

Option A is wrong because ExpressRoute Global Reach is for connecting on-premises networks to Azure via private peering, not for routing client requests to the nearest Azure region. Option B is wrong because Azure Bastion is a managed jump server service for secure RDP/SSH access to VMs, not a global traffic router. Option C is wrong because Azure Local Traffic Manager (LTM) is not a real Azure service; it might be confused with F5's LTM, which is not a native Azure global routing solution.

Full explanation →
41
Multi-Selectmedium

When planning the deployment of an Azure AI Search service, which TWO sizing and capacity parameters must you configure during creation? Each correct answer presents part of the solution.

Select 2 answers
A.Virtual CPU core count per hypervisor
B.Replicas
C.Partitions
D.Linux kernel version
E.Active Directory domain controller IP
AnswersB, C

Replicas are copies of the index that provide query throughput and high availability. Azure AI Search requires replica count at creation because it sets concurrent query capacity and redundancy, satisfying the capacity parameter the stem asks for.

Why this answer

During creation of an Azure AI Search service, you must choose the number of replicas (Option B) and partitions (Option C) as part of the pricing tier and capacity configuration. Replicas determine high availability and query throughput by creating multiple copies of the index that also enable read replicas for query load balancing. Partitions determine storage capacity and write throughput by splitting the index across multiple shards, and together replicas and partitions define the service's scaling units (SUs).

Options A, D, and E are not Azure AI Search creation parameters: virtual CPU core count per hypervisor, Linux kernel version, and Active Directory domain controller IP are infrastructure or OS-level settings that Azure manages and are not exposed when provisioning an Azure AI Search service.

Exam trap

AI-103 often tests the specific parameters required for Azure AI Search creation, and candidates may confuse them with VM or database sizing parameters, leading to incorrect selections.

Full explanation →
42
MCQhard

You are monitoring an Azure OpenAI deployment in Azure AI Foundry and notice an increase in HTTP 429 (Too Many Requests) errors. Your application needs to handle these rate limit spikes gracefully. What is the recommended architectural solution?

A.Restart the Azure OpenAI resource from the Azure Portal
B.Implement exponential backoff and jitter retry logic in the application SDK
C.Set the model temperature parameter to 0.0
D.Switch the authentication method from API keys to Managed Identity
AnswerB

HTTP 429 responses are transient throttling signals, so retrying immediately worsens contention. Exponential backoff with jitter spaces retries progressively and randomises them, preventing synchronised retry storms across clients. This satisfies the requirement to handle rate limit spikes gracefully without dropping requests.

Why this answer

HTTP 429 responses from Azure OpenAI indicate the deployment has exceeded its tokens-per-minute or requests-per-minute quota. The documented, recommended client-side mitigation is exponential backoff with jitter, which spaces retries out and randomizes them to avoid synchronized retry storms. The Azure OpenAI SDKs and Polly-style libraries implement this pattern natively.

Exam trap

AI-103 often tests whether candidates reach for infrastructure changes (restart, auth switch, parameter tuning) when the correct answer is a client-side resilience pattern — the trap is confusing quota/rate-limit handling with configuration or security changes.

How to eliminate wrong answers

Option A is wrong because restarting the Azure OpenAI resource does not change quota limits and would disrupt all consumers of the deployment without addressing the rate-limit cause. Option C is wrong because temperature controls randomness/creativity of completions, not throughput or rate limiting — it has no effect on 429 errors. Option D is wrong because switching from API keys to Managed Identity changes authentication/authorization, not the quota consumption that triggers 429s; it improves security but not rate-limit handling.

Full explanation →
43
MCQeasy

You are deploying an Azure AI Document Intelligence resource. You need to ensure that only client requests coming from your corporate public IP address range can access the endpoint. Which networking configuration should you apply?

A.Configure Azure Bastion access restriction rules.
B.Attach an Azure Application Gateway with a WAF policy.
C.Enable Microsoft Entra conditional access policies on the API keys.
D.Configure firewall virtual networks and IP address ranges in the Azure AI resource networking settings.
AnswerD

Selecting firewall virtual networks and IP address ranges restricts inbound access to specified public IP ranges, satisfying the corporate-range constraint. This network-level control blocks other clients before requests reach the Document Intelligence endpoint, unlike key-based authentication alone.

Why this answer

Azure AI resources, including Document Intelligence, provide networking settings where you can configure firewall rules to allow access only from specific virtual networks and public IP address ranges. This directly restricts endpoint access to your corporate IP range, satisfying the requirement. It is the native, intended mechanism for IP-based access control on Azure AI services.

Exam trap

AI-103 often tests the difference between network-level access control (firewall IP rules, private endpoints) and identity-based control (conditional access, API keys)—candidates may pick conditional access when the requirement is source IP restriction.

How to eliminate wrong answers

Option A is wrong because Azure Bastion is a service for secure RDP/SSH access to VMs, not for restricting access to AI service endpoints; it has no role in API endpoint firewalling. Option B is wrong because Application Gateway with WAF protects web applications from attacks but does not restrict access based on source IP for an AI service endpoint; it would require additional configuration and is not the direct method. Option C is wrong because Microsoft Entra conditional access policies apply to user authentication and cannot restrict access based on source IP for API key-based requests to an AI endpoint.

Full explanation →
44
MCQhard

You are configuring an Azure AI Document Intelligence container for deployment in an on-premises disconnected environment. Which environment variable is mandatory for the container to start successfully and report billing usage?

A.EULA=accept, billing=<billing-endpoint-uri>, apikey=<api-key>
B.AZURE_COGNITIVE_SERVICES_TENANT_ID
C.AZURE_AI_CONNECTOR_CONNECTION_STRING
D.DOCUMENT_INTELLIGENCE_MODEL_CACHE_DIR
AnswerA

Setting EULA=accept satisfies the licence-acceptance gate, while billing and apikey connect the container to Azure for usage metering. In a disconnected on-premises deployment these three variables are mandatory: without the billing endpoint and key, the container cannot start or report consumption to Azure.

Why this answer

Document Intelligence containers require three environment variables to start and report billing: EULA=accept (to acknowledge the license), billing=<billing-endpoint-uri> (the Azure endpoint that receives usage telemetry), and apikey=<api-key> (the key for that billing endpoint). Without all three, the container will not start or will not report usage, which is mandatory even in disconnected environments because billing telemetry is sent periodically when connectivity is available.

Exam trap

AI-103 often tests container prerequisites — the trap is forgetting that billing telemetry is still required for disconnected containers, or confusing container variables with unrelated Azure service connection strings.

How to eliminate wrong answers

Option B is wrong because AZURE_COGNITIVE_SERVICES_TENANT_ID is not a recognized environment variable for Document Intelligence containers — tenant ID is not part of the container billing configuration. Option C is wrong because AZURE_AI_CONNECTOR_CONNECTION_STRING is not a Document Intelligence container variable; it resembles a connector string from a different service and has no role here. Option D is wrong because DOCUMENT_INTELLIGENCE_MODEL_CACHE_DIR is not a mandatory variable — it may be used to specify a cache path for models, but the container starts without it and it does not affect billing.

Full explanation →
45
MCQmedium

You are training a custom template model in Azure AI Document Intelligence and notice that one of your key fields is consistently failing to extract on documents where the label text is vertically oriented. How should you address this issue?

A.Switch the model type from custom template to prebuilt read model.
B.Set the queryParameters property to 'orientation=vertical' in the training API request body.
C.Rotate the source PDF documents to landscape orientation prior to uploading them for training.
D.Re-examine and ensure all instances of the field across training samples are accurately tagged with correct bounding regions in Document Intelligence Studio.
AnswerD

Extraction failures on vertically oriented labels usually stem from inconsistent or missing bounding-region annotations in the training set. Retagging every occurrence of the field across samples with accurate bounding boxes gives the model the labelled examples it needs to learn that orientation.

Why this answer

The issue is likely due to inaccurate labeling of the field in the training samples, especially for vertically oriented text. Ensuring all instances of the field are accurately tagged with correct bounding regions in Document Intelligence Studio will help the model learn to extract the field regardless of orientation.

Exam trap

AI-103 often tests troubleshooting for custom model extraction failures; candidates might incorrectly think they can adjust API parameters or rotate documents, rather than ensuring training labels are accurate and consistent.

How to eliminate wrong answers

Option A is wrong because switching to prebuilt read model would not extract custom fields; it only provides OCR text. Option B is wrong because there is no queryParameters property for orientation in the training API; orientation is handled automatically by the model. Option C is wrong because rotating documents to landscape may not align with the actual orientation of the field and could introduce other issues; the model should be trained on the original orientation.

Full explanation →
46
MCQmedium

You are migrating an Azure AI service from one region to another. What is the most efficient way to migrate the configuration?

A.Restore a backup from a Recovery Services vault.
B.Use Azure Data Factory to copy the resource.
C.Manually recreate all resources in the new region.
D.Export the resource as an ARM template and deploy it to the new region.
AnswerD

ARM templates capture the resource's declarative configuration, so redeploying the exported template into the target region recreates settings without manual re-entry. This directly satisfies the stem's efficiency constraint, since every other migration route requires rebuilding or hand-copying configuration.

Why this answer

Exporting the resource as an ARM template captures its full configuration declaratively, and deploying that template to the new region recreates the resource with identical settings efficiently. This is the recommended infrastructure-as-code approach for regional migration of Azure AI services, avoiding manual error and preserving configuration fidelity. It is faster and more reliable than rebuilding resources by hand.

Exam trap

AI-103 often tests the misconception that backups or data-copy tools migrate resource configurations, when the correct approach for configuration migration is exporting and redeploying an ARM template.

How to eliminate wrong answers

Option A is wrong because Recovery Services vault backups are for VMs and certain workloads, not for Azure AI service configuration, and cannot restore a Cognitive Services resource. Option B is wrong because Azure Data Factory is a data integration service for moving data between stores, not for migrating resource configurations. Option C is wrong because manually recreating all resources is error-prone, slow, and defeats the purpose of efficient migration when a template export is available.

Full explanation →
47
MCQmedium

You are creating a Custom Vision project and need to select a domain. Your project involves classifying images of various food dishes. Which domain should you select to optimize model accuracy for food classification?

A.Landmark
B.Food
C.General
D.Retail
AnswerB

The Food domain is a pre-trained Custom Vision domain whose underlying features were learned from food imagery, so transfer learning starts from representations already tuned to dish textures and shapes. Selecting it satisfies the food-classification scenario by giving higher accuracy than the generic domain.

Why this answer

The Food domain in Custom Vision is a specialized model pretrained on food imagery, so it produces higher accuracy for classifying dishes than a generic domain. Selecting the domain that matches the image category is the recommended way to optimize model performance in Custom Vision.

Exam trap

AI-103 often tests whether candidates know that Custom Vision domains are task-specific and immutable after project creation, so picking General 'to be safe' is a common wrong-answer trap.

How to eliminate wrong answers

Option A is wrong because the Landmark domain is optimized for recognizing famous buildings and geographic landmarks, not food dishes. Option C is wrong because the General domain is a broad-purpose model that works across many categories but is less accurate than a specialized domain for a narrow task like food classification. Option D is wrong because the Retail domain is tuned for retail product recognition (e.g., shelf items), not prepared food dishes.

Full explanation →
48
MCQhard

You are optimizing a RAG pipeline in Azure AI Foundry. The model is hallucinating when retrieving documents from a large corpus. Which strategy should you implement to improve grounding accuracy?

A.Increase the temperature of the model to 1.0
B.Use a larger embedding model
C.Implement a Cross-Encoder re-ranking step
D.Increase the chunk size to include the entire document
AnswerC

A cross-encoder jointly scores each query-document pair, capturing fine-grained relevance that bi-encoder vector similarity misses. Re-ranking the retrieved candidates before generation filters out weakly relevant passages, directly improving grounding accuracy and reducing hallucination across the large corpus.

Why this answer

Cross-Encoder re-ranking is correct because it jointly encodes the query and each candidate document, producing a far more accurate relevance score than the bi-encoder cosine similarity used in initial vector retrieval. This precision filtering removes semantically similar but contextually irrelevant chunks before they reach the LLM, directly reducing hallucination caused by noisy or off-topic grounding context. In Azure AI Foundry, this is typically implemented via the semantic ranker or a custom re-ranking stage in the RAG orchestration pipeline.

Exam trap

AI-103 often tests the misconception that 'bigger embeddings' or 'bigger chunks' automatically improve RAG quality, when in fact retrieval precision via re-ranking is the lever that most directly reduces hallucination.

How to eliminate wrong answers

Option A is wrong because increasing temperature to 1.0 raises sampling randomness, which increases hallucination rather than reducing it — grounding accuracy requires lower temperatures (typically 0–0.3). Option B is wrong because a larger embedding model improves the quality of vector representations but does not fix the fundamental recall-precision gap in top-k retrieval; irrelevant chunks can still rank highly. Option D is wrong because increasing chunk size to entire documents dilutes the embedding signal, exceeds context windows, and introduces more irrelevant tokens, worsening grounding rather than improving it.

Full explanation →
49
MCQeasy

You need to view the latency of requests made to an Azure AI services endpoint. Where do you find this information?

A.The Subscription billing dashboard.
B.The 'Deployment' logs in the machine learning workspace.
C.Azure Advisor recommendations.
D.Azure Monitor metrics.
AnswerD

Azure Monitor metrics automatically capture request latency, call counts and error rates for Azure AI services endpoints, viewable in the metrics explorer or via diagnostic settings. This is the native telemetry surface for endpoint performance, requiring no custom instrumentation.

Why this answer

Azure Monitor metrics is the platform service that collects and displays numerical telemetry such as request latency, request count, and error rates for Azure AI services endpoints. You can view latency metrics in the Azure Portal under the resource's Metrics blade or via the Azure Monitor API. This is the standard place to observe performance data for Azure resources.

Exam trap

AI-103 often tests the difference between metrics (numerical performance data) and logs (detailed event records), so candidates who think latency requires log analysis pick the deployment logs option instead of Azure Monitor metrics.

How to eliminate wrong answers

Option A is wrong because the Subscription billing dashboard shows cost and usage charges, not request latency. Option B is wrong because 'Deployment' logs in a machine learning workspace relate to model deployments in Azure ML, not to Azure AI services endpoint latency telemetry. Option C is wrong because Azure Advisor provides best-practice recommendations and optimization suggestions, not real-time latency metrics.

Full explanation →
50
Multi-Selecteasy

Which THREE options are available for configuring network security in Azure?

Select 3 answers
A.Azure Private Link / Private Endpoints.
B.Azure Storage Account Keys.
C.Azure Active Directory Groups.
D.Azure Firewall.
E.Network Security Groups (NSG).
AnswersA, D, E

Private Endpoints map a service into your virtual network via a private IP, removing public internet exposure entirely. This satisfies the stem's network security requirement by enforcing private-only connectivity to Azure services, blocking exfiltration and unauthorised access at the network layer.

Why this answer

Azure Private Link / Private Endpoints (A) is correct because it provides private connectivity to Azure PaaS services over a private IP address within your virtual network, removing exposure to the public internet and thus directly configuring network-level security. Azure Firewall (D) is correct because it is a managed, cloud-based network security service that provides stateful firewall capabilities, FQDN filtering, and threat intelligence-based filtering at the network and application layers. Network Security Groups (E) are correct because they filter network traffic to and from Azure resources in a virtual network using inbound and outbound security rules based on source/destination IP, port, and protocol.

Azure Storage Account Keys (B) are access credentials for authenticating to storage, not a network security configuration mechanism. Azure Active Directory Groups (C) are used for identity and access management (authorization), not for configuring network security controls.

Exam trap

AI-103 often tests the distinction between network security controls (NSG, Azure Firewall, Private Link) and identity/access controls (storage keys, AAD groups); candidates pick identity mechanisms thinking they secure the network.

Full explanation →
51
Multi-Selecthard

Which THREE security principles should you follow when managing Azure AI resources?

Select 3 answers
A.Auditability and logging.
B.Maximum access for all administrators.
C.Defense in Depth.
D.Disable all security features to improve speed.
E.Principle of Least Privilege.
AnswersA, C, E

Auditability and logging record who accessed or modified Azure AI resources and when, producing an evidence trail. This supports detection of anomalous activity, forensic investigation after an incident, and compliance reporting, making it a core security principle for managing these resources.

Why this answer

Option A (Auditability and logging) is correct because managing Azure AI resources requires enabling diagnostic settings, Azure Monitor logs, and activity logs so that access to keys, models, and data can be traced and anomalous use detected. Option C (Defense in Depth) is correct because Azure AI workloads should be protected with layered controls such as network isolation (private endpoints, VNet integration), identity controls (Microsoft Entra ID, managed identities), encryption (Microsoft-managed or customer-managed keys), and monitoring, so no single failure exposes the resource. Option E (Principle of Least Privilege) is correct because you should grant only the minimum RBAC roles and scopes needed (for example, Cognitive Services User rather than Contributor/Owner) and use managed identities instead of embedding keys, limiting blast radius if credentials are compromised.

Option B is wrong because maximum access for all administrators violates least privilege and expands the attack surface. Option D is wrong because disabling security features to improve speed removes protections like encryption, authentication, and logging, which is never an acceptable trade-off for Azure AI resources.

Exam trap

AI-103 often tests whether candidates can reject obviously bad security options, so the trap is overthinking and selecting a plausible-sounding but wrong principle instead of the three canonical ones (defense in depth, least privilege, auditability).

Full explanation →
52
MCQmedium

You are designing a RAG solution using Azure AI Search and Azure OpenAI. Users report that when querying technical documentation, the system misses highly relevant paragraphs because the exact keywords do not match the user's conversational query. How should you configure Azure AI Search to improve retrieval accuracy?

A.Switch the index analyzer to a standard whitespace analyzer to prevent stemming issues.
B.Enable semantic ranking on the Azure AI Search index and use hybrid search combining vector embeddings and keyword BM25.
C.Increase the number of retrieved documents (top_k) from 3 to 20 without changing the query type.
D.Convert all documents to uppercase before indexing to ensure case-insensitive matching.
AnswerB

Semantic ranking re-ranks results using deep-learning language models that capture conversational intent, while hybrid search fuses BM25 keyword scores with vector similarity. Together they retrieve paragraphs whose meaning matches the query even when exact keywords differ, directly resolving the missed-relevance problem.

Why this answer

Enabling semantic ranking and using hybrid search (combining vector embeddings and BM25 keyword search) improves retrieval accuracy by leveraging both semantic understanding and keyword matching. This addresses the issue of conversational queries missing exact keywords.

Exam trap

AI-103 often tests the misconception that simply increasing top_k or changing analyzers can fix retrieval issues, when semantic ranking and hybrid search are the correct solutions.

How to eliminate wrong answers

Option A is wrong because a whitespace analyzer would worsen matching by ignoring stemming and synonyms. Option C is wrong because increasing top_k without changing query type may retrieve more irrelevant documents. Option D is wrong because uppercasing documents doesn't improve semantic matching and may harm case-sensitive content.

Full explanation →
53
Multi-Selectmedium

You are preparing a dataset for fine-tuning an Azure OpenAI model in Azure AI Foundry. Which TWO data formatting and preparation requirements must you follow? (Choose TWO)

Select 2 answers
A.Include unencrypted passwords and PII in plain text to improve model empathy
B.Upload raw unstructured PDF files directly as the fine-tuning training dataset
C.Format training data as JSONL (JSON Lines) where each line is a chat completion example
D.Ensure training examples follow the expected system, user, and assistant message roles
E.Encode all training examples in binary audio wave files (.wav)
AnswersC, D

Fine-tuning requires JSONL, with one complete chat completion example per line containing the messages array. This line-delimited structure lets Azure AI Foundry parse each training example independently, satisfying the formatting requirement for supervised fine-tuning datasets.

Why this answer

Option C is correct because Azure OpenAI fine-tuning requires the training file to be in JSONL (JSON Lines) format, where each line is a self-contained JSON object representing one training example, typically a chat completion example with a messages array. Option D is correct because for chat models the messages must use the expected role structure — system, user, and assistant — so the model learns the conversational mapping between prompts and desired responses; the assistant message provides the target completion. Option A is wrong because including unencrypted passwords or PII is a security and compliance violation and does nothing to improve model quality.

Option B is wrong because raw unstructured PDFs cannot be uploaded directly as a fine-tuning dataset; the data must first be parsed and converted into the required JSONL chat format. Option E is wrong because .wav audio files are for speech/audio models, not for fine-tuning a text-based Azure OpenAI chat model.

Exam trap

AI-103 often tests whether candidates confuse data preparation requirements with general best practices, luring them toward 'more data is better' answers like raw PDFs or PII inclusion instead of the strict JSONL chat-format requirement.

Full explanation →
54
MCQeasy

You are monitoring your Azure OpenAI deployment in Azure AI Foundry. You notice an increase in HTTP 429 status codes. What does this error indicate?

A.Internal server error at Microsoft
B.Model deployment not found
C.Rate limit or token quota exceeded (Too Many Requests)
D.Invalid API key authentication
AnswerC

HTTP 429 is returned when requests exceed the deployment's allocated rate limit or token quota within a time window. It signals throttling rather than an authentication or payload fault, so the fix is to retry with backoff, reduce request frequency, or raise the quota.

Why this answer

HTTP 429 'Too Many Requests' indicates the client has exceeded the rate limit or token quota assigned to the Azure OpenAI deployment. This occurs when requests per minute (RPM) or tokens per minute (TPM) surpass the deployment's quota, triggering throttling until the window resets.

Exam trap

AI-103 often tests the confusion between HTTP status codes — candidates must map 429 to rate limiting/throttling, 401 to authentication, 404 to missing resources, and 500 to server errors, rather than guessing based on vague descriptions.

How to eliminate wrong answers

Option A is wrong because internal server errors at Microsoft return HTTP 500 (or 503 for service unavailable), not 429. Option B is wrong because a missing deployment returns HTTP 404 Not Found. Option D is wrong because invalid API key authentication returns HTTP 401 Unauthorized (or 403 Forbidden for authorization failures).

Full explanation →
55
MCQhard

You are managing an enterprise Azure OpenAI resource that experiences bursty traffic patterns. To optimize cost without sacrificing availability during sudden workload spikes, you decide to implement a caching layer for frequent prompts and responses. Which Azure service should you integrate in front of Azure OpenAI?

A.Azure Traffic Manager routing profile
B.Azure Cache for Redis
C.Azure ExpressRoute local gateway
D.Azure Content Delivery Network (CDN)
AnswerB

Azure Cache for Redis sits in front of Azure OpenAI and stores frequent prompt-response pairs in memory, so repeated requests are served from cache rather than consuming token-based capacity. This absorbs bursty spikes without provisioning extra throughput, directly satisfying the cost-optimisation and availability constraints in the stem.

Why this answer

Azure Cache for Redis provides a low-latency, in-memory caching layer that can store frequent prompt-response pairs, reducing the number of calls to Azure OpenAI during bursty traffic. This optimizes cost by serving cached responses and maintains availability by offloading repeated requests from the OpenAI resource. It is the standard Azure service for application-level caching in front of AI endpoints.

Exam trap

AI-103 often tests the confusion between caching services (Redis) and routing/networking services (Traffic Manager, CDN, ExpressRoute)—candidates may pick CDN thinking it caches API responses, but CDN is for static content.

How to eliminate wrong answers

Option A is wrong because Azure Traffic Manager is a DNS-based global load balancer that routes traffic to different endpoints; it does not cache responses and would not reduce OpenAI calls. Option C is wrong because ExpressRoute is a private network connection between on-premises and Azure, not a caching service, and a local gateway is for routing, not caching. Option D is wrong because Azure CDN caches static content at edge locations, but it is not designed for dynamic API responses like OpenAI prompts, and it does not support the POST requests typically used for completions.

Full explanation →
56
MCQmedium

You are deploying an Azure AI multi-service account in the West US region. The finance department requires that all resource deployment costs be tracked by department code using Azure Resource Manager tags. Which Azure feature should you use to enforce the mandatory application of the department tag across all future Azure AI resource deployments?

A.Azure Cost Management budgets
B.Azure role-based access control (RBAC)
C.Azure Blueprints
D.Azure Policy
AnswerD

Azure Policy evaluates resource deployments against rules and can deny or auto-remediate non-compliant ones, so a policy with a deny effect and tag requirement forces the department tag onto every future Azure AI resource regardless of who deploys it.

Why this answer

Azure Policy is the governance service that enforces organizational standards by evaluating resources against rules and taking effects such as deny, audit, or append. A policy with an 'append' or 'deny' effect on the tag requirement enforces mandatory department tags across all future deployments.

Exam trap

AI-103 often tests the difference between cost visibility tools (budgets, Cost Management) and enforcement tools (Azure Policy) — the trap is choosing budgets because the question mentions cost tracking, when the requirement is enforcement of tags.

How to eliminate wrong answers

Option A is wrong because Azure Cost Management budgets only monitor and alert on spending; they cannot enforce tag application on deployments. Option B is wrong because RBAC controls who can perform actions on resources, not what metadata (tags) resources must carry. Option C is wrong because Azure Blueprints packages artifacts (including policies) for repeatable environments, but the enforcement mechanism itself is Azure Policy; Blueprints alone do not enforce tag compliance on arbitrary future deployments.

Full explanation →
57
Multi-Selecthard

When analyzing documents using Azure AI Document Intelligence, which THREE properties can be found within the returned JSON result object for an extracted table cell? (Choose THREE)

Select 3 answers
A.columnIndex
B.audioTranscript
C.sentimentScore
D.content
E.rowIndex
AnswersA, D, E

columnIndex is a zero-based integer giving the cell's horizontal position within the recognised table. It satisfies the stem's requirement by supplying the column coordinate in the JSON result, enabling reconstruction of the table grid alongside rowIndex and content.

Why this answer

In the Azure AI Document Intelligence analyze-result JSON, each table cell is represented with its grid position and text: option A (columnIndex) is correct because the zero-based column position of the cell within the table is returned as the columnIndex property, and option E (rowIndex) is correct because the zero-based row position is returned as the rowIndex property. Option D (content) is also correct because the cell's recognized text is returned in the content field (with spans and boundingRegions tying it to the source). Options B (audioTranscript) and C (sentimentScore) do not belong: audioTranscript relates to speech-to-text transcription and sentimentScore to language sentiment analysis, neither of which is part of the Document Intelligence table-cell schema.

Exam trap

The trap is confusing properties from other Azure AI services (like sentimentScore from Language, or audioTranscript from Speech) with Document Intelligence table cell properties; candidates must remember that Document Intelligence focuses on spatial and textual data.

Full explanation →
58
MCQhard

You are designing an agentic workflow in Azure AI Foundry that requires the agent to call an external REST API. Which component should you use to define the tool interface?

A.Azure AI Content Safety
B.LLM output parser
C.Tool definition using a Python function
D.Flow input schema
AnswerC

A Python function tool definition lets the agent invoke the external REST API through code, satisfying the requirement to define a callable tool interface. Azure AI Foundry registers the function's name, parameters and description, so the model can select and call it, with the function body performing the HTTP request.

Why this answer

In Azure AI Foundry agentic workflows, tools are defined as callable functions with a name, description, and parameter schema; a Python function definition (often decorated or registered as a tool) is the standard way to expose an external REST API to the agent. The model then emits structured arguments that invoke that function. This is the correct component for defining the tool interface.

Exam trap

AI-103 often tests the confusion between components that define tool interfaces (Python function/tool definition) and components that filter, parse, or schema-validate other parts of the flow.

How to eliminate wrong answers

Option A is wrong because Azure AI Content Safety filters harmful content; it does not define tool interfaces. Option B is wrong because an LLM output parser processes model output but does not define the tool's callable interface or schema. Option D is wrong because a flow input schema defines the inputs to the overall workflow, not the interface for an individual external API tool.

Full explanation →
59
MCQhard

You are setting up diagnostic logging for an Azure OpenAI resource. Your security team mandates that all log streams must be exported in real-time to a security information and event management (SIEM) platform outside Azure, such as Splunk. Which destination should you configure in the Azure Monitor diagnostic setting?

A.Send to a Microsoft Sentinel workbook
B.Configure an Azure Automation runbook webhook
C.Stream to an Azure Event Hub
D.Archive to an Azure Storage account container
AnswerC

Streaming to an Azure Event Hub forwards diagnostic logs in near real-time to an external consumer, letting Splunk ingest them outside Azure. Storage accounts and Log Analytics would not meet the real-time external SIEM requirement.

Why this answer

Azure Event Hub is the supported real-time streaming destination for Azure Monitor diagnostic settings and integrates natively with external SIEM platforms like Splunk via the Splunk add-on for Event Hub. It provides low-latency, high-throughput egress of logs outside Azure.

Exam trap

AI-103 often tests whether candidates confuse batch archival (Storage) or visualization (Sentinel workbooks) with real-time external streaming, so they pick a destination that cannot deliver low-latency logs to a third-party SIEM.

How to eliminate wrong answers

Option A is wrong because a Sentinel workbook is a visualization/reporting artifact, not a log export destination. Option B is wrong because an Automation runbook webhook is for triggering automation, not streaming diagnostic logs. Option D is wrong because archiving to Storage is batch/durable storage, not real-time streaming to an external SIEM.

Full explanation →
60
MCQeasy

You are writing a Python script using the Azure AI Language SDK (`azure-ai-textanalytics`). Which client class should you instantiate to connect to your language resource?

A.LanguageServiceClient
B.TextAnalyticsClient
C.TextAnalysisClient
D.CognitiveServicesClient
AnswerB

`TextAnalyticsClient` is the entry point exposed by the `azure-ai-textanalytics` package, authenticating against your language resource endpoint and key (or Microsoft Entra ID credential). It satisfies the stem's constraint of connecting a Python script to that resource, unlike authoring-plane or REST-only alternatives.

Why this answer

The `azure-ai-textanalytics` SDK (now part of the `azure-ai-language` family) exposes `TextAnalyticsClient` as the primary client class for interacting with Azure AI Language services such as sentiment analysis, key phrase extraction, and language detection. You instantiate it with your endpoint and an `AzureKeyCredential` or `DefaultAzureCredential`. The class name directly matches the package name, making it the correct entry point for this SDK version.

Exam trap

AI-103 often tests the distinction between the older `azure-ai-textanalytics` SDK and the newer `azure-ai-language` SDK, causing candidates to confuse `TextAnalyticsClient` with `TextAnalysisClient` or assume a generic `LanguageServiceClient` exists.

How to eliminate wrong answers

Option A is wrong because `LanguageServiceClient` is not a class in the `azure-ai-textanalytics` package; it may be confused with clients from other Azure Language SDKs or older Cognitive Services SDKs. Option C is wrong because `TextAnalysisClient` is the client class used in the newer `azure-ai-language` SDK (for conversational language understanding and other unified language features), not in `azure-ai-textanalytics`. Option D is wrong because `CognitiveServicesClient` is a generic, non-existent client in this context; Azure Cognitive Services are accessed via service-specific SDKs, not a single monolithic client.

Full explanation →
61
MCQmedium

You are building an agentic workflow in Azure AI Foundry using the Semantic Kernel framework. The agent needs to call external APIs dynamically based on user intent. Which component should you register with the kernel to enable this capability?

A.A system prompt template
B.A plugin containing native or semantic functions
C.A sequential planner instance
D.A custom memory store
AnswerB

Plugins expose native code and semantic functions that Semantic Kernel can invoke, letting the agent call external APIs dynamically. Registering a plugin with the kernel satisfies the dynamic external API requirement, since the planner selects functions based on inferred user intent.

Why this answer

In Semantic Kernel, plugins encapsulate native functions (code) or semantic functions (prompts) that the kernel can invoke. Registering a plugin with the kernel enables the agent to call external APIs dynamically based on user intent, as the kernel orchestrates function calls.

Exam trap

The trap is confusing planning or memory components with the actual function providers — candidates must know that plugins are the unit of external capability registration in Semantic Kernel.

How to eliminate wrong answers

Option A is wrong because a system prompt template only sets the agent's instructions or persona; it does not provide callable functions to interact with external APIs. Option C is wrong because a sequential planner instance is a planning strategy that orchestrates existing functions, but it does not itself provide the external API capability — the functions must come from a plugin. Option D is wrong because a custom memory store stores and retrieves context/embeddings, not executable functions for API calls.

Full explanation →
62
MCQeasy

You are preparing a prompt template for a summarization task. You want to provide three distinct examples of input text and expected summaries before asking the model to process the target text. What is this prompting technique called?

A.Retrieval-augmented generation
B.Chain-of-thought prompting
C.Few-shot prompting
D.Zero-shot prompting
AnswerC

Few-shot prompting supplies several worked input-output examples inside the prompt before the target request, letting the model infer the summarisation pattern. Providing three text-summary pairs matches this definition exactly, unlike zero-shot, which gives no examples.

Why this answer

Few-shot prompting provides a small number of input-output examples (typically 2-5) in the prompt to demonstrate the desired task before asking the model to perform it on new input. Three examples of input text and expected summaries is the textbook definition of few-shot summarization prompting.

Exam trap

AI-103 often tests the confusion between few-shot prompting (providing labeled examples in the prompt) and RAG or chain-of-thought — candidates must recognize that example-based task demonstration defines few-shot, regardless of task type.

How to eliminate wrong answers

Option A is wrong because retrieval-augmented generation (RAG) involves retrieving relevant documents from an external knowledge store and injecting them into the prompt — it is about grounding, not providing labeled task examples. Option B is wrong because chain-of-thought prompting encourages step-by-step reasoning by showing or requesting intermediate reasoning steps, not input-output example pairs for a summarization task. Option D is wrong because zero-shot prompting provides no examples at all — just an instruction — which is the opposite of what the question describes.

Full explanation →
63
MCQmedium

How do you handle sensitive user data when using an Azure AI Agent?

A.Send all data to the model; the model will ignore it.
B.Increase the model parameter size.
C.Store user PII in the vector index.
D.Use Data Masking/PII filtering before the data reaches the agent.
AnswerD

Data masking and PII filtering redact or tokenise sensitive identifiers before prompts reach the agent, so personal data never enters the model context or logs. This satisfies the requirement to handle sensitive user data securely at the input boundary.

Why this answer

Responsible handling of sensitive data in Azure AI Agents requires filtering or masking PII before it reaches the agent, so that personal data is not sent to the model, logged in prompts, or persisted in agent threads. Data masking/PII filtering (for example via Azure AI Language PII detection or a pre-processing step in Prompt Flow) enforces this at the boundary. This aligns with privacy-by-design and data-minimization principles required for Responsible AI compliance.

Exam trap

The trap here is the belief that the model itself provides privacy protection or that a bigger model is safer, when AI-103 expects you to know that PII must be filtered or masked before it reaches the agent.

How to eliminate wrong answers

Option A is wrong because models do not 'ignore' sensitive data — anything in the prompt is processed, may be retained in conversation history/threads, and could be surfaced in completions or logs, creating a privacy and compliance violation. Option B is wrong because increasing model parameter size changes capability, not data handling; a larger model has no built-in PII redaction and will still process whatever is sent. Option C is wrong because storing user PII in a vector index is the opposite of the required control — it persists sensitive data in a searchable store, widening the attack surface and violating data-minimization requirements.

Full explanation →
64
Multi-Selecthard

When designing a RAG-based agent, which THREE factors directly influence the quality of the retrieved information?

Select 3 answers
A.The system prompt's instructions on retrieval.
B.The use of hybrid search (combining keyword and vector search).
C.The granularity and quality of the chunked documents in the index.
D.The frequency of the model's training updates.
E.The model's temperature setting.
AnswersA, B, C

The system prompt governs whether and how the agent invokes retrieval, including query formulation and when to search. Poorly specified retrieval instructions cause missed or irrelevant lookups, directly degrading the quality of retrieved information regardless of index quality.

Why this answer

Option A is correct because the system prompt's retrieval instructions directly steer when and how the agent invokes the retriever, what queries it formulates, and how it filters or re-ranks results, all of which shape the relevance of retrieved context. Option B is correct because hybrid search combines lexical keyword matching (e.g., BM25) with dense vector similarity, improving recall for exact terms and semantic matches alike, which raises retrieval quality. Option C is correct because chunk granularity and document quality determine what units are embedded and indexed; poorly sized or noisy chunks degrade embedding fidelity and produce irrelevant or incomplete context.

Option D does not belong because training-update frequency affects the base model's parametric knowledge, not the retrieval pipeline that supplies external context in RAG. Option E does not belong because temperature only affects sampling randomness during generation and has no direct effect on which documents are retrieved.

Exam trap

AI-103 often tests whether candidates conflate generation-time parameters (temperature, training recency) with retrieval-pipeline factors — the trap is selecting model-tuning knobs when the question asks about retrieved information quality.

Full explanation →
65
Multi-Selecthard

Your enterprise architecture team is reviewing the security posture of an Azure AI deployment that handles sensitive financial documents. Which TWO actions should you implement to secure data in transit and at rest? (Choose two.)

Select 2 answers
A.Enforce a minimum TLS version of 1.2 or higher for all API requests to secure data in transit.
B.Configure customer-managed keys (CMK) using Azure Key Vault for encryption at rest.
C.Grant 'Contributor' permissions to all developer accounts to simplify key rotation.
D.Disable diagnostic logging to prevent sensitive document text from being stored in Log Analytics.
E.Use Microsoft-managed keys exclusively without integrating Azure Key Vault.
AnswersA, B

Enforcing TLS 1.2 or higher encrypts API traffic between clients and the Azure AI service, satisfying the in-transit requirement. Older versions such as TLS 1.0 and 1.1 have known weaknesses, so mandating the minimum version protects sensitive financial documents during transmission.

Why this answer

Option A is correct because enforcing a minimum TLS version of 1.2 (or higher) on all API requests protects sensitive financial documents while they travel over the network, preventing downgrade attacks and use of deprecated protocols such as TLS 1.0/1.1. Option B is correct because configuring customer-managed keys (CMK) through Azure Key Vault gives the organization control over the encryption keys used for data at rest, satisfying compliance and key-ownership requirements for sensitive financial data. Option C is not appropriate because granting 'Contributor' to all developer accounts violates least privilege and does not itself secure data in transit or at rest; key rotation should be handled through Key Vault access policies or RBAC scoped narrowly.

Option D is wrong because disabling diagnostic logging reduces auditability and incident response rather than protecting data, and logging can be configured to avoid storing sensitive content. Option E is wrong because relying exclusively on Microsoft-managed keys removes customer control over key lifecycle and does not meet the requirement to integrate Azure Key Vault for encryption at rest.

Exam trap

AI-103 often tests whether candidates pick 'more access' or 'less logging' as security measures, when the correct answers are encryption controls (TLS for transit, CMK for rest) that map directly to the stated requirement.

Full explanation →
66
MCQhard

You are managing an Azure OpenAI model deployment in an active-passive multi-region disaster recovery architecture. A regional outage occurs, and you failover traffic to the secondary region. However, clients report receiving authentication errors. What is the most likely cause?

A.Azure automatically deletes API keys during a regional failover.
B.The client application's managed identity has not been granted the required RBAC roles on the secondary Azure OpenAI resource.
C.Azure OpenAI models require re-training after any regional failover event.
D.The primary region's DNS zone must be manually deleted for failover to work.
AnswerB

Managed identities are tenant-level principals, but their RBAC role assignments are scoped per resource. Failing over to the secondary Azure OpenAI resource leaves the identity without a role assignment there, so token authorisation fails even though authentication succeeds.

Why this answer

When using managed identity for authentication, the identity must be granted the appropriate RBAC role (such as Cognitive Services OpenAI User) on each Azure OpenAI resource it accesses. During failover to a secondary resource in another region, that resource is a distinct Azure resource with its own scope, so role assignments from the primary do not automatically carry over. The authentication error is therefore caused by missing RBAC role assignments on the secondary resource.

Exam trap

AI-103 often tests the misconception that identity and RBAC configurations automatically replicate across regions or resources, when in fact each Azure resource has its own authorization scope requiring explicit role assignments.

How to eliminate wrong answers

Option A is wrong because Azure does not delete API keys during regional failover; keys persist on the resource and failover does not alter key material. Option C is wrong because Azure OpenAI models are pre-trained and deployed as-is; they do not require retraining after a failover, and retraining is not even a supported operation for base models. Option D is wrong because DNS zones are not manually deleted for failover; Azure Traffic Manager or Front Door handles routing, and DNS records are managed automatically, not by manual deletion.

Full explanation →
67
MCQmedium

You are deploying an Azure OpenAI model in Azure AI Foundry and need to restrict access so that only specific virtual networks within your enterprise can invoke the model endpoint. Which feature should you configure?

A.API key rotation policies
B.Azure Private Link with private endpoints
C.Cross-Origin Resource Sharing (CORS) headers
D.Model rate limit quotas
AnswerB

Azure Private Link with private endpoints assigns a private IP address from your virtual network to the Azure OpenAI endpoint, so traffic never traverses the public internet. This satisfies the stem's constraint of restricting invocation to specific enterprise virtual networks, since only networks with a private endpoint connection can reach the model.

Why this answer

Azure Private Link with private endpoints creates a private IP address for the Azure OpenAI endpoint inside your virtual network, so traffic never traverses the public internet. By disabling public network access and allowing only specific VNets via private endpoints, you restrict model invocation to those enterprise networks. This is the correct feature for network-level access restriction to an Azure AI Foundry model endpoint.

Exam trap

AI-103 often tests the difference between authentication controls (keys, RBAC) and network controls (Private Link, firewalls), tricking candidates into choosing API key rotation when the question explicitly asks about network restriction.

How to eliminate wrong answers

Option A is wrong because API key rotation is a credential-hygiene practice that changes the secret but does not restrict which networks can reach the endpoint — anyone with the key from any network can still invoke it. Option C is wrong because CORS headers control which browser origins can make cross-origin requests; they are a web-application security feature, not a network isolation mechanism, and are trivially bypassed by non-browser clients. Option D is wrong because rate limit quotas control throughput and prevent abuse but do not restrict access by network location — a quota-limited endpoint is still publicly reachable.

Full explanation →
68
MCQeasy

What is the primary purpose of a 'Vector Store' in a RAG-based agentic application?

A.To store embeddings for efficient semantic similarity search.
B.To generate text responses.
C.To log user conversations for analytics.
D.To fine-tune the LLM.
AnswerA

A vector store persists the embeddings produced from source content and indexes them so the agent can run approximate nearest-neighbour semantic similarity search, retrieving the most relevant chunks to ground generation in a RAG pipeline.

Why this answer

A vector store's primary purpose in a RAG-based agentic application is to persist embeddings and enable efficient semantic similarity search. When a user query arrives, it is embedded and compared against stored vectors using approximate nearest neighbor algorithms to retrieve the most relevant chunks. This retrieval step grounds the LLM's response in external knowledge, which is the core of RAG.

Exam trap

AI-103 often tests the confusion between retrieval (vector store) and generation (LLM), and between RAG and fine-tuning as ways to inject domain knowledge.

How to eliminate wrong answers

Option B is wrong because text generation is performed by the LLM itself, not by the vector store; the vector store only supplies retrieved context. Option C is wrong because logging user conversations is an observability concern handled by logging or analytics services, not by a vector store. Option D is wrong because fine-tuning modifies the LLM's weights through training, whereas a vector store is used for retrieval-augmented generation without changing model parameters.

Full explanation →
69
MCQmedium

You are configuring content safety filters in Azure AI Foundry for a public-facing generative AI application. You need to ensure that the application detects and blocks potential jailbreak attempts and protected material text. Which component should you configure?

A.Azure Key Vault access policies restricting model endpoint tokens.
B.Azure AI Content Safety text moderation service with jailbreak and protected material detection enabled.
C.Azure API Management rate-limiting policies configured to drop requests containing long prompts.
D.Azure Monitor activity log alerts for anomalous token usage.
AnswerB

Azure AI Content Safety's text moderation API includes dedicated jailbreak and protected material detectors, which general prompt shields or model-level filters do not expose as configurable categories. Enabling both satisfies the requirement to detect and block jailbreak attempts and protected material text in the public-facing application.

Why this answer

Azure AI Content Safety provides dedicated detectors for jailbreak attempts (prompt injection/manipulation) and protected material text (copyrighted content), which can be configured in Azure AI Foundry to block or flag such inputs and outputs. This is the correct component for enforcing content safety policy on a public-facing generative AI app.

Exam trap

AI-103 often tests whether candidates confuse security/operations services (Key Vault, API Management, Monitor) with content moderation, when only Azure AI Content Safety provides jailbreak and protected material detection.

How to eliminate wrong answers

Option A is wrong because Azure Key Vault access policies govern secrets and key management, not content moderation or jailbreak detection. Option C is wrong because API Management rate limiting controls request volume and throttling, not the semantic content of prompts, and long prompts are not inherently jailbreaks. Option D is wrong because Azure Monitor activity log alerts track control-plane operations and anomalies, not the semantic content of user prompts or protected material.

Full explanation →
70
Multi-Selectmedium

You are configuring prompt engineering practices for a generative AI application in Azure AI Foundry. Which THREE practices should you follow to ensure high output quality and reliability? (Choose three.)

Select 3 answers
A.Specify the desired output format explicitly (e.g., JSON schema or markdown tables).
B.Embed all system instructions inside the user prompt string without using system roles.
C.Rely entirely on vague, implicit instructions to give the model creative freedom.
D.Use delimiters such as triple backticks or XML tags to clearly separate instructions from user-provided data.
E.Incorporate few-shot examples of edge cases and expected behaviors.
AnswersA, D, E

Declaring an explicit output format constrains generation to a defined structure, such as a JSON schema or markdown table, which directly satisfies the stem's demand for reliable, high-quality output. Without it, the model chooses its own layout, producing inconsistent, harder-to-parse responses.

Why this answer

Option A is correct because explicitly specifying the desired output format (for example, a JSON schema or markdown table) constrains the model's response structure, making outputs consistent, parseable, and reliable for downstream application code. Option D is correct because using delimiters such as triple backticks or XML tags clearly separates instructions from user-supplied data, which reduces prompt injection risk and prevents the model from confusing data with directives. Option E is correct because few-shot examples of edge cases and expected behaviors ground the model in the desired response pattern, improving accuracy and consistency on ambiguous or unusual inputs.

Option B is not appropriate because embedding all system instructions in the user prompt discards the system role, which is the proper channel for persistent behavioral guidance and guardrails in Azure AI Foundry. Option C is not appropriate because vague, implicit instructions increase ambiguity and variance in model outputs, undermining the quality and reliability the scenario requires.

Exam trap

AI-103 often tests the misconception that giving the model more freedom or hiding system instructions in user prompts improves quality, when the opposite is true for reliability.

Full explanation →
71
MCQeasy

Which authentication method is recommended to prevent 'key leakage' in Azure AI Services?

A.Storing keys in a plaintext file.
B.Creating custom headers for API requests.
C.Regenerating keys every 30 days.
D.Using Microsoft Entra ID authentication.
AnswerD

Microsoft Entra ID authentication issues short-lived, role-scoped tokens, so no long-lived service keys are stored in code or configuration. That removes the static credential an attacker could exfiltrate, directly satisfying the requirement to prevent key leakage in Azure AI Services.

Why this answer

Microsoft Entra ID authentication replaces static API keys with short-lived OAuth 2.0 tokens issued to an identity, so there is no long-lived secret to leak. Access is governed by RBAC roles on the AI resource, and tokens expire automatically. This is the recommended passwordless pattern for preventing key leakage in Azure AI Services.

Exam trap

AI-103 often tests whether candidates equate key rotation with key-leak prevention, so the trap is picking 'regenerate keys' when the truly secure answer is eliminating keys via Entra ID.

How to eliminate wrong answers

Option A is wrong because storing keys in a plaintext file is the exact anti-pattern that causes key leakage — anyone with file access or a misconfigured repo exposes the key. Option B is wrong because custom headers are just a transport mechanism; they do not remove the need for a credential and can still carry a leakable key. Option C is wrong because regenerating keys every 30 days reduces the window of exposure but does not prevent leakage — the key still exists, can be copied, and rotation can break dependent apps.

Full explanation →
72
MCQmedium

You are designing an agentic workflow where an orchestrator agent delegates sub-tasks to specialized worker agents (e.g., a research agent and a coding agent). Which framework or service native to Azure AI supports multi-agent orchestration patterns?

A.Azure Data Factory pipeline authoring
B.Semantic Kernel or Microsoft AutoGen
C.Azure Stream Analytics query editor
D.Azure Logic Apps standard workflow designer
AnswerB

Semantic Kernel and Microsoft AutoGen both provide native multi-agent orchestration: AutoGen models conversable agents that exchange messages to coordinate tasks, while Semantic Kernel's Agent Framework supports group chat and delegation patterns. This directly satisfies the stem's requirement for orchestrator-to-worker delegation across specialised research and coding agents.

Why this answer

Semantic Kernel and Microsoft AutoGen are Microsoft's frameworks designed for building AI agents and multi-agent orchestration, including patterns where an orchestrator delegates to specialized worker agents. They provide abstractions for agent roles, conversation routing, and tool use that map directly to the described workflow.

Exam trap

AI-103 often tests whether candidates can distinguish AI agent orchestration frameworks from general-purpose integration or data services that merely can call AI endpoints.

How to eliminate wrong answers

Option A is wrong because Azure Data Factory is a data integration and ETL service for pipelines and data movement, not an agent orchestration framework. Option C is wrong because Azure Stream Analytics is a real-time stream processing engine using SQL-like queries, not a multi-agent AI framework. Option D is wrong because Azure Logic Apps is a workflow automation and integration service for connectors and business processes; while it can call AI services, it is not the native multi-agent orchestration framework for Azure AI.

Full explanation →
73
MCQmedium

Your development team has created a custom model using Azure Machine Learning. You need to deploy this model as a real-time web service endpoint with auto-scaling capabilities. Which deployment target should you select?

A.Azure Functions consumption plan
B.Azure Container Registry webhook
C.Azure Batch cluster
D.Azure Managed Online Endpoint
AnswerD

Managed online endpoints provide real-time scoring with integrated autoscaling, scaling rules based on metrics or schedules, and no infrastructure management. This directly satisfies the stem's requirement for a real-time web service endpoint with auto-scaling, unlike batch endpoints or AKS-based deployments.

Why this answer

Azure Managed Online Endpoints are designed for real-time inference with built-in autoscaling, authentication, and monitoring, making them the correct deployment target for a custom Azure Machine Learning model exposed as a web service. They support scale-to-zero or rule-based scaling based on metrics like CPU utilization or request rate.

Exam trap

AI-103 often tests whether candidates confuse batch-oriented or event-driven compute (Batch, Functions) with real-time managed online endpoints, which are the only option offering native autoscaling for ML web services.

How to eliminate wrong answers

Option A is wrong because Azure Functions consumption plan is a serverless event-driven compute model not designed for hosting ML model inference with GPU or model-specific scaling semantics. Option B is wrong because Azure Container Registry webhooks trigger actions on image push events; they are not a deployment target for serving models. Option C is wrong because Azure Batch is for large-scale parallel batch processing jobs, not low-latency real-time web service endpoints.

Full explanation →
74
MCQeasy

You are configuring an Azure AI Document Intelligence prebuilt model to extract key-value pairs from standard invoices. Which API operation should your application call to process the document asynchronously?

A.POST /documentintelligence/documentModels/prebuilt-invoice:train?api-version=2024-02-29-preview
B.POST /documentintelligence/models/prebuilt-invoice/analyze?api-version=2024-02-29-preview
C.POST /formrecognizer/v2.1/prebuilt/invoice/analyze
D.POST /documentintelligence/documentModels/prebuilt-invoice:analyze?api-version=2024-02-29-preview
AnswerD

The `:analyze` operation submits the invoice and returns a 202 with an `Operation-Location` header, letting the application poll for results rather than block. This satisfies the asynchronous processing constraint, unlike synchronous `:analyze` variants or extraction endpoints that return results inline.

Why this answer

The correct async analyze operation for a prebuilt model in the 2024-02-29-preview API version is POST /documentintelligence/documentModels/{modelId}:analyze, where {modelId} is prebuilt-invoice. This returns a 202 Accepted with an Operation-Location header that your app polls until the analysis completes. The :analyze suffix is the long-running-operation (LRO) pattern used by Document Intelligence.

Exam trap

AI-103 often tests API path memorization — the trap is confusing the legacy /formrecognizer/ path with the current /documentintelligence/ path, or mixing up :train and :analyze action suffixes.

How to eliminate wrong answers

Option A is wrong because :train is the operation used to train a custom model, not to analyze a document with a prebuilt model. Option B is wrong because the path /documentintelligence/models/prebuilt-invoice/analyze is not the correct route — the service uses /documentModels/ and the :analyze suffix, not a slash-separated analyze segment. Option C is wrong because /formrecognizer/v2.1/prebuilt/invoice/analyze is the legacy Form Recognizer v2.1 endpoint, which is deprecated and does not match the 2024-02-29-preview API version specified in the question.

Full explanation →
75
MCQhard

You are configuring a custom named entity recognition project in Language Studio. You have tagged your training documents and want to split your dataset into training and testing sets. What are the recommended split options available in Language Studio?

A.Splitting exclusively by alternating lines within each document.
B.External split configuration via an Azure DevOps pipeline YAML file.
C.Automatic split by percentage or manual assignment of documents to sets.
D.Command-line split using Azure PowerShell scripts only.
AnswerC

Language Studio offers automatic splitting by a chosen percentage, or manual assignment of individual documents to training and testing sets. Both options satisfy the requirement to partition tagged documents, giving either a quick proportional split or precise control over set membership.

Why this answer

Language Studio for custom NER provides two dataset-splitting methods: automatic split by a percentage (e.g., 80/20) that randomly assigns documents to training and testing sets, or manual assignment where you explicitly tag each document as training or testing. Both are configured directly in the project's data labeling interface.

Exam trap

AI-103 often tests whether candidates know that dataset splitting in Language Studio is document-level and supports only automatic percentage or manual assignment — not line-level, external YAML, or PowerShell-based splits.

How to eliminate wrong answers

Option A is wrong because splitting by alternating lines within a document is not a supported method — splitting operates at the document level, not the line level, and would corrupt entity context. Option B is wrong because Language Studio does not support external split configuration via Azure DevOps YAML; splitting is handled in the portal or via the REST API's dataset split endpoint. Option D is wrong because there is no PowerShell-only command-line split mechanism; the REST API and portal are the supported interfaces.

Full explanation →

Page 1 of 7

Page 2

All pages

1234567

Practice AI-103 by domain

Target a specific domain to shore up weak areas.

Text AnalysisPlan And Manage AN Azure AI SolutionImplement Generative AI And Agentic SolutionsComputer VisionInformation Extraction
See all domains with question counts →