Microsoft Certified: Azure AI Apps and Agents Developer Associate (AI-103) (AI-103) AI-103 Questions 76–150 | Page 2/7

MCQeasy

You are managing prompt assets in Azure AI Foundry. Your team collaborates on developing, testing, and versioning system prompts. Which feature in Azure AI Foundry should you use to store and manage these prompt templates?

A.Azure Storage Blob lifecycle management policies

B.Prompt flow / Prompt management workspace assets

C.Azure App Service deployment slots

D.Azure Key Vault secret versions

AnswerB

Prompt flow provides workspace assets purpose-built for authoring, testing, and versioning prompt templates within Azure AI Foundry. It satisfies the collaboration and versioning constraint by centralising prompt storage, letting teams iterate on system prompts as managed, reusable artefacts rather than ad hoc strings.

Why this answer

Prompt flow / Prompt management workspace assets is correct because Azure AI Foundry provides a dedicated prompt asset store where teams can author, version, test, and share system prompts and prompt templates as first-class workspace artifacts. This enables collaboration, change tracking, and reproducibility across the development lifecycle, which is exactly what the scenario requires. Prompt flow integrates these assets into orchestrated LLM workflows with evaluation and deployment support.

Exam trap

AI-103 often tests whether candidates map generic Azure services (Blob, Key Vault, App Service) onto AI-specific needs, so the trap is choosing a storage or secrets service instead of the purpose-built prompt management feature.

How to eliminate wrong answers

Option A is wrong because Azure Storage Blob lifecycle management policies govern data retention and tiering (hot/cool/archive) for blobs — they have no concept of prompt versioning, testing, or collaboration. Option C is wrong because Azure App Service deployment slots are for blue/green deployment of web applications, not for storing or versioning prompt templates. Option D is wrong because Azure Key Vault secret versions manage sensitive credentials and cryptographic material, not prompt content — using Key Vault for prompts would be an abuse of a secrets store and lacks prompt-specific tooling.

Full explanation →

MCQeasy

You are writing a system prompt for an Azure OpenAI model deployed in Azure AI Foundry. You want to ensure the model responds only using the provided context and refuses to answer if the context is insufficient. What technique should you use?

A.Configure Azure AI Content Safety severity levels to high

B.Set the model temperature to 1.0 and top_p to 0.0

C.Enable semantic caching on the Azure OpenAI deployment

D.Include explicit grounding instructions and a negative constraint in the system prompt

AnswerD

Explicit grounding instructions direct the model to answer solely from supplied context, while a negative constraint tells it to refuse when that context is insufficient. Together they satisfy the requirement to prevent answers drawn from outside knowledge.

Why this answer

Explicit grounding instructions in the system prompt — such as 'Answer only using the provided context; if the context does not contain the answer, respond that you do not know' — are the standard prompt-engineering technique to constrain the model to retrieved context and refuse out-of-scope questions. This is a negative constraint that directly shapes model behavior at inference time. It is the correct, low-cost, deterministic approach for RAG grounding.

Exam trap

AI-103 often tests whether candidates confuse content safety filtering with grounding enforcement, tricking them into picking a Content Safety option when the question is about prompt-level refusal behavior.

How to eliminate wrong answers

Option A is wrong because Content Safety severity levels filter harmful categories (hate, violence, self-harm, sexual) and do not enforce grounding or refusal on insufficient context. Option B is wrong because temperature 1.0 increases randomness and top_p 0.0 is an invalid/degenerate setting that would break sampling — this combination does not enforce grounding. Option C is wrong because semantic caching stores and reuses prior responses to reduce latency and cost; it does not influence whether the model refuses to answer when context is insufficient.

Full explanation →

MCQeasy

You are designing an AI application and want to test different foundation models (e.g., GPT-4o vs. Llama 3) side-by-side using the same prompt and test dataset. Which feature in Azure AI Foundry should you use?

A.Azure Monitor workbook performance graphs

B.Azure Cosmos DB cross-region replication

C.Side-by-side evaluation in Azure AI Foundry playground

D.Azure Front Door traffic routing

AnswerC

Side-by-side evaluation runs the identical prompt and test dataset against multiple deployed models within one playground view, producing comparable metrics and outputs. This directly satisfies the stem's constraint of comparing GPT-4o against Llama 3 on the same inputs without rebuilding separate evaluation pipelines.

Why this answer

Side-by-side evaluation in the Azure AI Foundry playground lets you run the same prompt and test dataset against multiple foundation models and compare outputs, metrics, and quality. This is the native feature for model comparison during design and selection.

Exam trap

AI-103 often tests whether candidates recognize Azure AI Foundry's native evaluation and playground comparison features versus unrelated Azure services that sound like they compare or route traffic.

How to eliminate wrong answers

Option A is wrong because Azure Monitor workbooks visualize telemetry and performance metrics, not model output quality comparisons for the same prompt. Option B is wrong because Cosmos DB cross-region replication is a database availability and latency feature, unrelated to model evaluation. Option D is wrong because Azure Front Door is a global HTTP load balancer and CDN, not a model evaluation tool.

Full explanation →

MCQeasy

Which file format is NOT natively supported for training custom models in Azure AI Document Intelligence?

A.PDF

B.JPEG

C.BMP

D.DOCX

AnswerD

DOCX is not a native training format for custom models; Document Intelligence accepts PDF, JPEG, PNG, TIFF, BMP and text-based inputs. The question asks which format is unsupported, and DOCX requires conversion first, making it the exception among the listed options.

Why this answer

Azure AI Document Intelligence custom model training accepts PDF, JPEG, PNG, BMP, TIFF, and HEIF image formats, but DOCX (Word) files are not natively supported as training inputs. To train on Word content, you must first convert DOCX to PDF or render pages as images before uploading to the labeled training set. This is why DOCX is the correct answer to 'which is NOT natively supported.'

Exam trap

AI-103 often tests the assumption that 'any common document format works' — candidates pick PDF or JPEG thinking one of them is unsupported, missing that DOCX is the odd one out because Document Intelligence is image/PDF-centric.

How to eliminate wrong answers

Option A is wrong because PDF is one of the primary supported formats for custom model training (both text-based and scanned PDFs). Option B is wrong because JPEG is explicitly supported as an image input for custom template/neural model training. Option C is wrong because BMP is also a supported image format alongside PNG, TIFF, and HEIF.

Full explanation →

MCQhard

You are building an agentic application that uses function calling. The model frequently hallucinates arguments for complex tool schemas. Which technique should you apply to improve function calling reliability?

A.Convert all tool arguments into unstructured natural language text.

B.Remove all parameter descriptions from the tool schema to reduce prompt token count.

C.Set the model frequency penalty to high values.

D.Enhance the JSON schema with detailed parameter descriptions, enum constraints, and strict schema validation settings.

AnswerD

Detailed parameter descriptions, enum constraints and strict schema validation narrow the model's generation space, reducing invented arguments. This directly addresses the hallucinated-arguments constraint by making valid values explicit and rejecting malformed output before tool execution.

Why this answer

Detailed parameter descriptions, enum constraints, and strict schema validation give the model explicit guidance on expected argument formats and allowed values, reducing ambiguity that leads to hallucinated arguments. Strict schema validation (e.g., JSON Schema strict mode) rejects malformed outputs and forces the model to conform. This is the most direct and effective technique for improving function-calling reliability.

Exam trap

AI-103 often tests the misconception that reducing schema detail or token count improves function calling, when in fact richer, stricter schemas reduce hallucinated arguments.

How to eliminate wrong answers

Option A is wrong because converting arguments to unstructured natural language removes the structured contract that function calling relies on, making parsing and validation harder and increasing errors. Option B is wrong because removing parameter descriptions deprives the model of context about what each parameter means, which increases hallucination rather than reducing it. Option C is wrong because frequency penalty controls repetition in generated text, not the correctness or structure of function arguments; it does not address schema adherence.

Full explanation →

Multi-Selecteasy

Which THREE prebuilt document models are natively available out-of-the-box in Azure AI Document Intelligence? Each correct selection represents a valid prebuilt model.

Select 3 answers

A.Prebuilt W-2 tax form model

B.Prebuilt Invoice model

C.Prebuilt Video Sentiment analyzer

D.Prebuilt 3D CAD Model Analyzer

E.Prebuilt Receipt model

AnswersA, B, E

The prebuilt W-2 model is one of Document Intelligence's native US tax document models, requiring no training or custom labelling. It is available out-of-the-box via the prebuilt endpoint, satisfying the question's requirement for natively available models.

Why this answer

Azure AI Document Intelligence ships several prebuilt models for common business documents, and the prebuilt W-2 tax form model (A) is one of them, designed to extract US wage and tax statement fields such as wages, tips, and federal tax withheld. The prebuilt Invoice model (B) is also natively available and extracts key invoice fields like vendor name, invoice date, due date, line items, and amounts due. The prebuilt Receipt model (E) is likewise an out-of-the-box model that extracts merchant name, transaction date, line items, and totals from sales receipts.

Options C and D are not real Document Intelligence prebuilt models: Video Sentiment analyzer is not a document analysis capability, and 3D CAD Model Analyzer does not exist in the service's prebuilt model catalog.

Exam trap

AI-103 often tests the knowledge of which prebuilt models exist; candidates may confuse Document Intelligence capabilities with other Azure AI services like Language or Vision.

Full explanation →

MCQeasy

What is the primary benefit of using the Model Catalog in Azure AI Foundry?

A.Automated prompt engineering

B.Centralized discovery and deployment of models

C.Network security configuration

D.Automatic data labeling

AnswerB

The Model Catalog provides a single curated inventory where teams browse, evaluate, and deploy models from multiple providers into Azure AI Foundry, removing the need to source and integrate each model separately. This centralises discovery and deployment governance.

Why this answer

The Model Catalog in Azure AI Foundry is a centralized hub for discovering, comparing, and deploying foundation models from Microsoft, OpenAI, Meta, Mistral, and others. Its primary value is streamlining model selection and one-click deployment into managed endpoints, so teams can quickly move from evaluation to production. It is not a prompt engineering, networking, or data labeling tool.

Exam trap

AI-103 often tests whether candidates confuse the Model Catalog's discovery/deployment role with unrelated capabilities like prompt engineering or data labeling that belong to other Azure services.

How to eliminate wrong answers

Option A is wrong because automated prompt engineering is not a function of the Model Catalog; prompt optimization is handled through separate tooling like Prompt Flow or manual prompt engineering. Option C is wrong because network security configuration is managed through Azure networking constructs (VNets, private endpoints, NSGs), not the Model Catalog. Option D is wrong because automatic data labeling is a feature of Azure Machine Learning data labeling projects, not the Model Catalog, which focuses on model discovery and deployment.

Full explanation →

MCQhard

You are deploying a custom AI model using a container image stored in Azure Container Registry (ACR). How do you ensure the deployment can pull the image securely?

A.Use the ACR 'admin' password in the configuration.

B.Assign the 'AcrPull' role to the endpoint's managed identity.

C.Deploy the model from a local machine.

D.Make the ACR public.

AnswerB

Assigning the AcrPull role to the endpoint's managed identity grants the deployment a token-based pull permission scoped to the registry, so no credentials are stored in configuration. This satisfies the secure-pull constraint by authenticating through Microsoft Entra ID rather than registry admin keys.

Why this answer

The secure, secretless way to let a deployment pull from ACR is to give the consuming resource a managed identity and assign that identity the AcrPull role scoped to the registry. The platform then authenticates to ACR using the identity's token, so no credentials are stored in configuration. This follows the Azure best practice of eliminating long-lived secrets from deployment pipelines.

Exam trap

AI-103 often tests whether candidates default to the ACR admin account (the 'easy' answer) instead of the managed identity + AcrPull pattern, which is the secure, least-privilege approach the exam expects.

How to eliminate wrong answers

Option A is wrong because the ACR admin account is a single shared credential with full registry access, cannot be scoped, and is disabled by default — using it violates least privilege and secret-management guidance. Option C is wrong because deploying from a local machine does not solve how the remote endpoint authenticates to ACR; the endpoint still needs credentials or an identity to pull the image. Option D is wrong because making ACR public exposes all images to anonymous pulls and is a severe data-exposure risk, not a secure deployment method.

Full explanation →

MCQmedium

You are testing an agent in the Azure AI Foundry playground. You observe that the agent answers questions about its internal prompt. What is the most effective way to prevent this?

A.Delete the search index.

B.Use a higher-parameter model.

C.Update the system prompt to explicitly refuse requests about its instructions.

D.Disable the playground.

AnswerC

System prompt instructions take precedence over user input, so adding an explicit refusal directive stops the agent disclosing its internal prompt. This satisfies the playground scenario by addressing the behaviour at its source rather than filtering outputs after leakage has already occurred.

Why this answer

The most effective and direct mitigation for an agent revealing its system prompt is to harden the system prompt itself with explicit instructions to refuse any request about its instructions, configuration, or internal prompt. This is a prompt-level guardrail that addresses the observed behavior without disabling functionality or removing knowledge sources.

Exam trap

AI-103 often tests the misconception that a bigger model or removing data will fix prompt leakage — the exam expects you to recognize system-prompt hardening as the direct, proportionate control.

How to eliminate wrong answers

Option A is wrong because deleting the search index removes grounding data the agent needs for legitimate answers and does nothing to stop prompt-leakage behavior. Option B is wrong because a higher-parameter model does not inherently refuse prompt-extraction attempts — larger models can be equally susceptible to jailbreaks and may even be more articulate at leaking. Option D is wrong because disabling the playground prevents testing but does not fix the underlying vulnerability in production.

Full explanation →

MCQhard

Your enterprise application uses Custom Vision. Due to compliance policies, you must ensure that your training data and model endpoints reside within a specific Azure region with strict data residency guarantees. Which resource configuration should you use?

A.Use the Global resource endpoint for Custom Vision.

B.Use Azure AI Vision multi-service account with global routing.

C.Create a Custom Vision training and prediction resource in the targeted Azure region.

D.Store images in Azure Blob Storage with geo-replication.

AnswerC

Creating both the training and prediction resources in the targeted Azure region keeps training data and model endpoints resident there, satisfying the compliance data residency constraint. Custom Vision stores data per resource region, so resource placement determines residency.

Why this answer

Custom Vision data residency is controlled by the region of the training and prediction resources you create. To guarantee that training data and model endpoints stay in a specific Azure region, you must create both a Custom Vision training resource and a prediction resource in that target region. The service then stores and processes data only within that region.

Exam trap

AI-103 often tests the difference between the global Custom Vision portal endpoint and regional resources; candidates pick the global endpoint thinking it is region-agnostic, when it actually violates data residency.

How to eliminate wrong answers

Option A is wrong because the global Custom Vision endpoint (customvision.ai) routes requests to a shared global resource that does not guarantee data residency in a specific region. Option B is wrong because a multi-service Azure AI Vision account with global routing may process data in multiple regions and does not provide the strict regional pinning required by compliance. Option D is wrong because geo-replicating images in Blob Storage addresses storage redundancy, not the region where Custom Vision trains models or hosts prediction endpoints.

Full explanation →

MCQhard

You are configuring rate limiting and request throttling for multiple client applications consuming an Azure OpenAI endpoint. Different clients have different tier agreements (Bronze, Silver, Gold). Where should you implement these tiered rate-limiting policies?

A.Azure API Management (APIM)

B.Azure Application Gateway custom WAF rules

C.Azure OpenAI service built-in tier settings

D.Network Security Group (NSG) outbound rules

AnswerA

APIM applies subscription-scoped rate-limit and quota policies, so each Bronze, Silver or Gold client key gets its own tier thresholds. The Azure OpenAI endpoint itself exposes only uniform token-per-minute limits, which cannot differentiate clients, so tiered throttling must sit in the gateway layer.

Why this answer

Azure API Management (APIM) is the correct place to implement tiered rate-limiting and throttling policies because it sits in front of the Azure OpenAI endpoint as a gateway and supports per-subscription, per-product, and per-key rate-limit policies. You can create Bronze, Silver, and Gold products/subscriptions in APIM and apply rate-limit or rate-limit-by-key policies with different quotas (calls per minute, tokens per minute) to each tier. Azure OpenAI itself does not expose customer-configurable tiered throttling policies per client.

Exam trap

AI-103 often tests the misconception that Azure OpenAI has built-in per-customer tier settings — in reality, tiered rate limiting for multiple clients is a gateway (APIM) responsibility, not a model-service feature.

How to eliminate wrong answers

Option B is wrong because Application Gateway WAF rules inspect and block HTTP traffic based on OWASP-style signatures and custom match conditions, not consumption quotas per client tier — it has no native rate-limit-by-subscription construct. Option C is wrong because Azure OpenAI's built-in quotas (TPM/RPM) are assigned per deployment/subscription by Microsoft and cannot be partitioned into arbitrary Bronze/Silver/Gold client tiers. Option D is wrong because NSGs operate at L3/L4 and filter by IP/port/protocol; they cannot inspect HTTP requests or enforce per-client request-rate policies.

Full explanation →

MCQhard

You are implementing prompt engineering for a customer support agent. To minimize hallucinations when the model lacks information, which Prompt Flow technique is most effective?

A.Chain-of-thought prompting

B.Instructional grounding to only answer from context

C.Increasing the Top_P parameter

D.Few-shot prompting with varied examples

AnswerB

Instructing the model to answer solely from supplied context constrains generation to retrieved evidence, so unsupported tokens are suppressed rather than invented. This directly satisfies the stem's requirement to minimise hallucinations when the model lacks information.

Why this answer

Instructional grounding — explicitly instructing the model to answer only from the provided context and to say it doesn't know otherwise — is the most direct and effective prompt-engineering technique to minimize hallucinations when the model lacks information. It constrains the model's generation to the retrieved context rather than its parametric memory.

Exam trap

AI-103 often tests whether candidates confuse reasoning techniques (chain-of-thought, few-shot) with grounding techniques — the trap is picking chain-of-thought because it sounds rigorous, when the question is specifically about preventing hallucination from missing context.

How to eliminate wrong answers

Option A is wrong because chain-of-thought prompting improves reasoning on complex tasks but does not prevent the model from inventing facts when context is missing; it can even amplify hallucination by generating plausible-sounding reasoning. Option C is wrong because increasing Top_P widens the sampling distribution, making the model more likely to produce diverse and potentially fabricated tokens, which increases hallucination risk. Option D is wrong because few-shot prompting with varied examples can improve format and style adherence but does not guarantee the model stays within the provided context, especially for out-of-distribution queries.

Full explanation →

MCQeasy

You are deploying an Azure AI Speech resource using an ARM template. Which parameter defines the name of the resource instance?

A.sku

B.name

C.location

D.kind

AnswerB

The `name` parameter in the ARM template's resource block sets the resource instance name, satisfying the stem's requirement to define the Azure AI Speech resource's name. Other parameters configure location, SKU, or kind, but only `name` supplies the unique identifier used during deployment.

Why this answer

In an ARM template, the 'name' property within the resource block specifies the name of the resource instance being deployed, such as the Azure AI Speech account name. This is a required property for every ARM resource and determines the unique identifier of the deployed resource. The other properties configure SKU, region, and resource type but do not set the instance name.

Exam trap

AI-103 often tests confusion between ARM template properties, expecting candidates to know that 'name' sets the resource instance identifier while 'sku', 'location', and 'kind' configure other attributes.

How to eliminate wrong answers

Option A is wrong because 'sku' defines the pricing tier and capabilities (e.g., S0, F0) of the Speech resource, not its name. Option C is wrong because 'location' specifies the Azure region where the resource is deployed, not its name. Option D is wrong because 'kind' identifies the type or variant of the Cognitive Services resource (e.g., SpeechServices), not the instance name.

Full explanation →

MCQhard

You are designing an agentic workflow in Azure AI Agent Service where multiple specialized agents need to collaborate to solve complex, multi-step financial analysis tasks. One agent manages the workflow state and decides which sub-agent to invoke next. What architectural pattern is best suited for this implementation?

A.Orchestrator-workers pattern, where a central agent dynamically assigns sub-tasks to specialized worker agents based on intermediate results.

B.Single monolithic prompt pattern, where all instructions are combined into one massive system prompt without multi-agent handoffs.

C.Map-reduce batch processing pattern, where all sub-agents execute independently in parallel without communication.

D.Static sequential pipeline, where agents execute strictly in a predefined linear order without conditional branching.

AnswerA

The orchestrator-workers pattern satisfies the stem's requirement for a central agent that manages workflow state and dynamically selects which specialised sub-agent to invoke next. Unlike fixed sequential or parallel pipelines, the orchestrator decides routing at runtime based on intermediate financial results, enabling genuine multi-step collaboration between Microsoft Entra ID-secured agents.

Why this answer

The orchestrator-workers pattern uses a central orchestrator agent that maintains workflow state, evaluates intermediate results, and dynamically selects which specialized worker agent to invoke next. This matches the requirement that one agent 'manages the workflow state and decides which sub-agent to invoke next' for multi-step financial analysis. It supports conditional branching, iterative refinement, and handoffs, which are essential when the sequence of sub-tasks cannot be known in advance.

Exam trap

AI-103 often tests whether candidates confuse 'multi-agent' with 'parallel agents' — the trap is picking map-reduce or a static pipeline when the scenario requires dynamic, stateful delegation by a central orchestrator.

How to eliminate wrong answers

Option B is wrong because a single monolithic prompt cannot support true multi-agent handoffs, state isolation, or dynamic delegation — it collapses all reasoning into one context and defeats the purpose of specialized agents. Option C is wrong because map-reduce executes sub-agents independently in parallel with no communication or shared state, so no agent can decide the next sub-task based on intermediate results. Option D is wrong because a static sequential pipeline has no conditional branching and cannot adapt the order of agent invocations to intermediate findings, which the scenario explicitly requires.

Full explanation →

MCQhard

You are troubleshooting a performance issue where your application's calls to Azure AI Vision experience high latency. Your application is hosted in Azure West US, but your Azure AI Vision resource was accidentally created in Azure Southeast Asia. What is the recommended remediation?

A.Increase the timeout threshold in the HTTP client configuration.

B.Recreate the Azure AI Vision resource in the West US region close to your application hosting.

C.Enable Azure ExpressRoute between West US and Southeast Asia.

D.Switch from REST API to gRPC protocol.

AnswerB

Latency stems from the geographic distance between the West US-hosted application and the Southeast Asia resource, since cross-region calls add network round-trip time. Recreating the Azure AI Vision resource in West US colocates it with the application, eliminating that transit delay and satisfying the latency constraint.

Why this answer

Azure AI Vision is a regional service, and network latency between the application and the service endpoint is dominated by geographic distance. Recreating the resource in West US colocates the service with the application, minimizing round-trip time and eliminating cross-region egress. This is the standard remediation for latency caused by region mismatch.

Exam trap

AI-103 often tests whether candidates recognize that latency from region mismatch cannot be fixed by timeouts, ExpressRoute, or protocol changes — the only real fix is colocating the resource with the application.

How to eliminate wrong answers

Option A is wrong because increasing the timeout only masks the latency — it does not reduce it and can make the app feel slower under load. Option C is wrong because ExpressRoute provides dedicated private connectivity but does not eliminate the speed-of-light latency between West US and Southeast Asia; it improves reliability and bandwidth, not geographic latency. Option D is wrong because switching from REST to gRPC reduces serialization overhead marginally but cannot overcome the fundamental cross-region round-trip time, which is the dominant factor here.

Full explanation →

Multi-Selectmedium

You are configuring text summarization using the Azure AI Language REST API. Which THREE parameters or configuration options are associated with summarization requests? (Choose three.)

Select 3 answers

A.maxSentenceCount parameter

B.Audio sample rate frequency

C.Abstractive summarization task definition

D.Image resolution DPI setting

E.Extractive summarization task definition

AnswersA, C, E

maxSentenceCount caps how many sentences the extractive summarization result returns, controlling output length. It satisfies the stem's requirement for a summarization request parameter in the Azure AI Language REST API, applying specifically to extractive tasks.

Why this answer

Option A (maxSentenceCount parameter) is correct because the extractive summarization task in the Azure AI Language REST API accepts a maxSentenceCount value in its parameters object to limit how many sentences the summary returns. Option C (Abstractive summarization task definition) is correct because the API supports an abstractiveSummarization task kind, which generates new natural-language summary text rather than selecting existing sentences. Option E (Extractive summarization task definition) is correct because the API also supports an extractiveSummarization task kind, which ranks and returns the most salient sentences from the input document.

Option B (Audio sample rate frequency) is incorrect because sample rate is a speech/audio processing setting, not a text summarization parameter. Option D (Image resolution DPI setting) is incorrect because DPI relates to image processing and has no role in the text summarization REST API.

Exam trap

AI-103 often tests whether candidates can identify which parameters belong to summarization versus speech or vision services, since the exam mixes options from multiple Azure AI modalities in a single question.

Full explanation →

MCQmedium

You are integrating Document Intelligence with Azure Blob Storage. You need to provide the model with a set of documents for training. How should you prepare the data?

A.Convert all files to CSV

B.Upload files and corresponding .labels.json and .ocr.json files

C.Store files in an Azure SQL database

D.Upload raw files only

AnswerB

Custom Document Intelligence models require each training document accompanied by a .labels.json file defining fields, plus an .ocr.json file providing the OCR layout. Uploading all three together to Blob Storage lets the model learn field positions and content.

Why this answer

To train a Document Intelligence model with documents in Azure Blob Storage, you must upload the documents along with their corresponding .labels.json and .ocr.json files. The .labels.json file contains the labeled data (field names and values), and the .ocr.json file contains the OCR results for the document. These files are generated using the Document Intelligence Studio or SDK and are required for training a custom model.

Exam trap

The trap is assuming that raw documents alone are sufficient for training; candidates must remember that labeled data and OCR files are required for custom extraction models.

How to eliminate wrong answers

Option A is wrong because converting files to CSV is not a supported format for Document Intelligence training; it requires the original document formats (PDF, JPEG, etc.) with accompanying JSON files. Option C is wrong because storing files in Azure SQL database is not a supported data source for Document Intelligence training; it requires Azure Blob Storage. Option D is wrong because uploading raw files only is insufficient; the labeled data and OCR files are necessary for supervised training.

Full explanation →

MCQeasy

You are deploying an Azure OpenAI model in Azure AI Foundry and want to evaluate its performance and safety against a validation dataset before moving to production. Which tool should you use?

A.Azure Monitor Log Analytics query packs

B.Azure Advisor cost recommendations

C.Model evaluation in Azure AI Foundry

D.Azure API Management developer portal

AnswerC

Model evaluation in Azure AI Foundry runs a validation dataset against the deployed model, computing performance metrics alongside safety evaluations such as harmful content and jailbreak risk. This provides the pre-production assessment the scenario requires before promoting the model.

Why this answer

Model evaluation in Azure AI Foundry provides a built-in evaluation framework where you can run your deployed model against a validation dataset and compute metrics like groundedness, relevance, fluency, coherence, and safety metrics (hate, violence, sexual, self-harm). It is the native tool for pre-production performance and safety assessment. Azure Monitor and Advisor serve different purposes (observability and cost), and API Management is for API governance.

Exam trap

AI-103 often tests whether candidates pick generic Azure monitoring/governance tools (Monitor, Advisor, APIM) for model evaluation when the correct answer is the purpose-built evaluation capability in Azure AI Foundry.

How to eliminate wrong answers

Option A is wrong because Azure Monitor Log Analytics query packs are for operational telemetry and log analysis, not for structured model quality/safety evaluation against a validation dataset. Option B is wrong because Azure Advisor cost recommendations focus on cost optimization, not model performance or safety evaluation. Option D is wrong because Azure API Management developer portal is for API documentation, subscription management, and developer onboarding — it has no model evaluation capability.

Full explanation →

MCQhard

You are troubleshooting an Azure AI Language asynchronous summarization job. You submitted a batch of documents using the `/analyze` endpoint with long-running operations (LRO). How should your application check the status and retrieve the results of the job?

A.Listen for incoming webhook notifications on an Azure Event Grid subscription attached to the text analytics resource.

B.Query the Azure Monitor metrics API for job completion status.

C.Poll the URL provided in the 'operation-location' response header using a GET request until the status is 'succeeded'.

D.Repeatedly send synchronous POST requests to the same document endpoint until a 200 OK is returned.

AnswerC

The /analyze LRO returns an operation-location header containing a job status URL. Polling that URL with GET returns the job state, and once it reports 'succeeded' the response body carries the summarization results, satisfying the asynchronous retrieval requirement.

Why this answer

Azure AI Language asynchronous (LRO) jobs return an 'operation-location' header in the initial 202 Accepted response. The application must poll that URL with GET requests until the status field reads 'succeeded', then retrieve results from the same endpoint. This is the standard Azure long-running operation pattern.

Exam trap

AI-103 often tests the LRO polling pattern; candidates incorrectly assume webhook/Event Grid push notifications or Azure Monitor metrics provide job completion status for asynchronous AI Language jobs.

How to eliminate wrong answers

Option A is wrong because Azure AI Language LROs do not push Event Grid webhook notifications for job completion; Event Grid integration is not the documented status mechanism for these analyze jobs. Option B is wrong because Azure Monitor metrics expose telemetry (latency, call counts), not per-job completion status or results. Option D is wrong because repeatedly sending synchronous POSTs to the document endpoint is not how LROs work and would resubmit jobs rather than retrieve status.

Full explanation →

MCQmedium

Your development team requires secure, passwordless authentication from an Azure Function to an Azure AI Vision resource. Which authentication mechanism should you implement to adhere to security best practices?

A.Shared Access Signature (SAS) tokens with read-only permissions

B.Basic authentication using service administrator username and password

C.Primary API Key passed in the Ocp-Apim-Subscription-Key header

D.Microsoft Entra managed identity

AnswerD

A managed identity gives the Function an automatically rotated credential issued by Microsoft Entra ID, so no secrets or passwords are stored in code or configuration. This satisfies the passwordless, secure-access constraint for reaching Azure AI Vision.

Why this answer

Microsoft Entra managed identity lets the Azure Function authenticate to the Azure AI Vision resource without storing any secret — the platform issues and rotates the token automatically, and you grant the function's identity the appropriate RBAC role (e.g., Cognitive Services User) on the target resource. This is the passwordless, least-privilege pattern Microsoft recommends for service-to-service auth. Keys and SAS tokens are shared secrets that can leak; managed identity eliminates that class of risk entirely.

Exam trap

AI-103 often tests whether candidates recognize that API keys and SAS tokens are still secrets, so the 'most secure' answer is always the passwordless managed identity option, not a key-based one.

How to eliminate wrong answers

Option A is wrong because SAS tokens are still bearer credentials that must be stored, transmitted, and rotated — they can be leaked or over-scoped, so they are not passwordless. Option B is wrong because basic authentication with an administrator username and password is the weakest possible pattern: credentials are long-lived, broadly privileged, and trivially exposed in config or logs. Option C is wrong because the Ocp-Apim-Subscription-Key header carries the primary API key, a static shared secret that, if leaked, grants full access until manually regenerated — exactly the key-leakage risk managed identity avoids.

Full explanation →

MCQeasy

What is the purpose of 'Azure Advisor' in an AI solution?

A.To manage user identities.

B.To automatically patch the AI models.

C.To provide best practice recommendations for resources.

D.To host the web frontend for the AI application.

AnswerC

Azure Advisor analyses deployed resources and surfaces best-practice recommendations across cost, security, reliability, operational excellence and performance. For an AI solution, it satisfies the stem's requirement by continuously evaluating configuration telemetry against Microsoft's recommended practices, flagging concrete improvements rather than merely reporting raw metrics or enforcing policy.

Why this answer

Azure Advisor is a personalized cloud consultant that analyzes your deployed resources and telemetry to surface best-practice recommendations across five categories: Cost, Security, Reliability, Operational Excellence, and Performance. In an AI solution, it reviews resources such as Azure OpenAI, Cognitive Services, and supporting compute/storage and flags misconfigurations like missing redundancy or underutilized capacity. It is advisory only — it never modifies resources on your behalf, so it cannot patch or manage anything.

Exam trap

AI-103 often tests whether candidates confuse Azure Advisor (advisory recommendations only) with services that actually enforce or remediate configuration, such as Azure Policy or Microsoft Defender for Cloud.

How to eliminate wrong answers

Option A is wrong because identity management is the role of Microsoft Entra ID (formerly Azure AD), not Azure Advisor, which has no authentication or directory function. Option B is wrong because Azure Advisor is a read-only recommendation engine — it does not deploy updates, patch models, or perform any remediation; model/OS patching is handled by update management tools or the service platform itself. Option D is wrong because hosting a web frontend is a compute/App Service or Static Web Apps responsibility; Advisor only analyzes existing resources and offers guidance.

Full explanation →

MCQhard

You are troubleshooting a Custom Vision object detection model that has low precision and high false-positive rates for a specific object class. During data analysis, you notice that negative images (images without the object) lack background variety. What is the recommended remediation step?

A.Increase the probability threshold in the prediction API call.

B.Remove all tagged bounding boxes and retrain as a multi-class classification model.

C.Apply aggressive data augmentation to duplicate existing positive images.

D.Add diverse negative images containing background environments similar to where false positives occur.

AnswerD

Adding negative images that mirror the deployment backgrounds teaches the model to distinguish the target class from similar-looking surroundings. This directly addresses the lack of background variety causing false positives, improving precision without altering positive samples.

Why this answer

Low precision with high false positives on a specific class almost always means the model has not learned to distinguish the target object from background contexts it mistakenly associates with that class. Adding diverse negative images that mirror the environments where false positives occur teaches the model that those backgrounds alone do not constitute the object, directly reducing false positives and improving precision.

Exam trap

AI-103 often tests the misconception that tuning the prediction threshold is a fix for precision problems, when the real remediation is improving training data with representative negative images.

How to eliminate wrong answers

Option A is wrong because raising the probability threshold only suppresses low-confidence detections globally; it does not teach the model to stop firing on specific background contexts and can hurt recall without fixing the underlying data gap. Option B is wrong because converting to multi-class classification discards bounding-box localization and does not address the false-positive root cause, which is insufficient negative/background diversity. Option C is wrong because duplicating existing positive images with augmentation reinforces the same backgrounds the model already overfits to and does nothing to introduce the missing negative variety.

Full explanation →

MCQmedium

Your application uses the Azure AI Document Intelligence Prebuilt Receipt model to process customer receipts. A receipt contains a custom tip amount that is not explicitly categorized by the default prebuilt schema. How should you extract this custom value?

A.Call the Azure OpenAI service directly with the raw image to guess the tip amount.

B.Pass a custom JSON schema parameter in the Prebuilt Receipt API request body to extend its extraction fields.

C.Modify the prebuilt receipt model schema directly in the Azure Portal to add a new required field.

D.Use the Prebuilt Layout model alongside custom regex rules or train a Custom model to capture the custom tip field.

AnswerD

The Prebuilt Receipt model returns only its fixed schema, so an uncategorised tip field cannot be extracted directly. Combining the Prebuilt Layout model with regex rules, or training a Custom model on labelled receipts, captures that field, satisfying the stem's requirement for the custom tip amount.

Why this answer

Prebuilt models expose a fixed schema and cannot be extended via request parameters or portal edits. To capture a field outside that schema (like a custom tip), you either use the Layout model to get raw text/lines and apply your own extraction logic (e.g., regex), or train a Custom extraction model with labeled samples that include the tip field. Option D describes both valid paths.

Exam trap

AI-103 often tests the misconception that prebuilt models can be 'extended' with a schema parameter or portal edit — in reality, prebuilt schemas are immutable and custom fields require Layout+logic or a Custom model.

How to eliminate wrong answers

Option A is wrong because calling Azure OpenAI directly bypasses Document Intelligence's structured extraction, introduces nondeterminism, and is not the supported mechanism for extending a prebuilt model's schema. Option B is wrong because the Prebuilt Receipt API does not accept a custom JSON schema parameter — prebuilt schemas are immutable and defined by Microsoft. Option C is wrong because prebuilt models cannot be modified in the Azure Portal; you cannot add required fields to a Microsoft-managed model.

Full explanation →

MCQmedium

You need to ensure that your agent's response is formatted as valid JSON. How should you approach this?

A.Set the temperature to 0.5.

B.Use an external library to convert text to JSON.

C.Ask the user to request JSON.

D.Define the JSON schema in the system prompt instructions.

AnswerD

Embedding the JSON schema within the system prompt constrains the model's output structure at generation time, satisfying the requirement for valid JSON. The system prompt carries higher priority than user turns, so the schema reliably governs formatting across responses.

Why this answer

Defining the JSON schema in the system prompt instructions guides the model to generate output that conforms to the specified structure. This approach leverages the model's instruction-following capabilities to produce valid JSON without external parsing. It is the most direct and reliable method for ensuring format compliance.

Exam trap

AI-103 often tests the misconception that temperature or external libraries can enforce JSON formatting, when the primary method is instructing the model via the system prompt.

How to eliminate wrong answers

Option A is wrong because setting the temperature to 0.5 affects randomness and creativity, not output format; it does not guarantee valid JSON. Option B is wrong because using an external library to convert text to JSON is a post-processing step that may fail if the model's output is not structured, and it does not ensure the model generates JSON natively. Option C is wrong because asking the user to request JSON shifts the responsibility and does not ensure the agent's response is formatted correctly by default.

Full explanation →

100

MCQhard

You are implementing an agentic workflow where an agent needs to search a vector database, inspect the retrieved results, and if the results are insufficient, reformulate the query and search again before answering. What agent design pattern are you implementing?

A.Static zero-shot generation pattern

B.Direct REST API proxy pattern

C.Database indexing batch pattern

D.Self-Correction / Reflection pattern

AnswerD

Self-correction/reflection fits because the agent evaluates retrieved results and reformulates its query when they prove insufficient, then searches again before answering. This iterative critique-and-retry loop, rather than a single retrieval pass, satisfies the stem's requirement for repeated searching until results suffice.

Why this answer

The Self-Correction / Reflection pattern is an agentic design where the agent evaluates its own intermediate output (retrieved results), determines whether it is sufficient, and if not, reformulates its approach and retries before producing a final answer. This iterative evaluate-and-retry loop is the defining characteristic of reflection-based agents. It enables the agent to improve answer quality without human intervention.

Exam trap

AI-103 often tests agentic patterns by describing a loop behavior — candidates pick a generic 'RAG' or 'API' answer instead of recognizing the evaluate-and-retry cycle as Reflection/Self-Correction.

How to eliminate wrong answers

Option A is wrong because static zero-shot generation produces a single response with no evaluation or retry loop — it cannot inspect results and reformulate. Option B is wrong because a direct REST API proxy pattern is a simple pass-through integration with no reasoning or self-evaluation component. Option C is wrong because database indexing batch pattern refers to offline data preparation/indexing, not runtime agent reasoning or query reformulation.

Full explanation →

101

Multi-Selectmedium

When designing prompt engineering strategies for complex reasoning tasks in Azure OpenAI, which TWO techniques help reduce hallucinations and improve output reliability? (Choose TWO)

Select 2 answers

A.Few-shot prompting providing high-quality input-output demonstration examples

B.Chain-of-thought (CoT) prompting instructing step-by-step reasoning

C.Disabling system prompts entirely to maximize model freedom

D.Setting the top_p parameter to 0.0 while setting temperature to 2.0

E.Encoding all prompt variables in hexadecimal format

AnswersA, B

Few-shot prompting supplies concrete input-output pairs that demonstrate the desired format and reasoning pattern, anchoring the model to expected responses. This reduces hallucination by constraining generation to patterns evidenced in the demonstrations, improving reliability on complex reasoning tasks.

Why this answer

Few-shot prompting (A) is correct because supplying several high-quality input-output demonstration examples in the prompt conditions the model on the desired format, tone, and reasoning pattern, which measurably reduces hallucination and improves consistency on complex reasoning tasks. Chain-of-thought prompting (B) is correct because explicitly instructing the model to reason step by step decomposes a hard problem into intermediate steps, improving accuracy on multi-step reasoning and making errors easier to detect. The unmarked options do not belong: removing system prompts (C) strips grounding and behavioral constraints, which increases hallucination risk; setting top_p to 0.0 with temperature 2.0 (D) is a contradictory, degenerate sampling configuration that does not improve reliability; and hex-encoding prompt variables (E) only obscures text and adds no reasoning benefit.

Exam trap

The trap is selecting parameter tweaks (temperature, top_p) or format changes as reliability techniques — the exam expects you to recognize prompt engineering methods like few-shot and CoT.

Full explanation →

102

MCQeasy

Which Azure feature allows you to control which users can modify or delete your Azure AI resources?

A.Azure RBAC.

B.Conditional Access policies.

C.Azure Firewall.

D.Resource Locks.

AnswerA

Azure role-based access control assigns roles scoped to Azure AI resources, determining which users may modify or delete them. This directly satisfies the stem's access-control requirement, unlike keys or network rules, which govern authentication and connectivity rather than user permissions.

Why this answer

Azure RBAC (Role-Based Access Control) is the mechanism that governs who can perform management-plane actions such as modifying or deleting Azure AI resources. You assign roles like Contributor, Owner, or Cognitive Services Contributor to users, groups, or service principals at a scope (management group, subscription, resource group, or resource), and those assignments determine whether modify/delete operations are permitted.

Exam trap

AI-103 often tests the distinction between RBAC (who can do what), Conditional Access (under what conditions they can sign in), and Resource Locks (what actions are blocked for everyone) — candidates frequently pick Resource Locks when the question asks about controlling user permissions.

How to eliminate wrong answers

Option B is wrong because Conditional Access policies govern the conditions under which authentication succeeds (device compliance, location, MFA), not what actions an authenticated user can perform on a resource. Option C is wrong because Azure Firewall is a network security service that filters traffic, not an authorization system for resource management operations. Option D is wrong because Resource Locks (CanNotDelete, ReadOnly) prevent deletion or modification regardless of RBAC, but they are a guardrail, not a user-permission mechanism — they do not control which users can modify resources, they block the action for everyone.

Full explanation →

103

MCQeasy

What is the primary function of the 'layout' model in Azure AI Document Intelligence?

A.To translate text into different languages

B.To classify document types

C.To extract key-value pairs from receipts

D.To extract text, tables, and document structure

AnswerD

The layout model performs document structure analysis, returning extracted text, tables, selection marks and their bounding regions without applying domain-specific field schemas. It provides the general-purpose OCR and structural foundation that custom or prebuilt models build upon.

Why this answer

The layout model in Azure AI Document Intelligence extracts text, tables, selection marks, and document structure (such as paragraphs, headings, and reading order) without requiring a trained custom model. It is the general-purpose model for understanding a document's structural layout, making it the correct answer.

Exam trap

AI-103 often tests model-purpose confusion — candidates pick the receipt or custom model for key-value extraction, forgetting that the layout model is specifically about text, tables, and structural elements, not field extraction.

How to eliminate wrong answers

Option A is wrong because translation is handled by Azure AI Translator, not by any Document Intelligence model — Document Intelligence extracts content but does not translate it. Option B is wrong because document type classification is performed by the custom classification model (or the prebuilt/invoice models for specific types), not by the layout model, which does not categorize documents. Option C is wrong because key-value pair extraction from receipts is the job of the prebuilt receipt model (or a custom extraction model), whereas the layout model returns structure and content, not field-specific key-value pairs.

Full explanation →

104

Multi-Selectmedium

Which TWO metrics are reported by Document Intelligence Studio to evaluate model performance?

Select 2 answers

A.Recall

B.Throughput

C.Latencey

D.Precision

E.Confidence score

AnswersA, D

Recall measures the proportion of genuine labelled fields the model successfully extracted, directly quantifying missed detections. Document Intelligence Studio reports it alongside precision, so it satisfies the stem's requirement for a performance metric rather than a training setting or confidence threshold.

Why this answer

Document Intelligence Studio reports Recall (A) and Precision (D) as the two core metrics for evaluating a custom model's extraction quality. Recall measures the proportion of actual ground-truth field values that the model correctly extracted, indicating how many true items were found. Precision measures the proportion of the model's extracted values that are actually correct, indicating how many of its predictions were accurate.

These two metrics together characterize the model's accuracy on labeled test data in the Studio's evaluation view. Throughput (B) and Latency (C) are performance/operational characteristics rather than model accuracy metrics, and Confidence score (E) is a per-prediction value returned by the model, not an aggregate evaluation metric reported by the Studio.

Exam trap

AI-103 often tests the confusion between evaluation metrics (precision, recall, F1) and runtime attributes (confidence score, latency, throughput) — candidates pick confidence score because it sounds like a quality measure.

Full explanation →

105

MCQhard

You are analyzing multi-language documents containing mixed English and Japanese text using Azure AI Vision Read API. How should you specify the language parameter to ensure optimal OCR accuracy?

A.Use the legacy OCR v3.0 endpoint with multi-language flag.

B.You must submit English and Japanese pages separately in distinct API calls.

C.Set language=en-US exclusively.

D.Omit the language parameter or set language=unk to enable automatic language detection across supported languages.

AnswerD

Omitting the parameter, or passing language=unk, triggers Azure AI Vision's automatic language identification, which detects the script per text line. This satisfies the mixed English and Japanese constraint, since forcing a single language code would degrade recognition of the other script.

Why this answer

The Read API supports automatic language detection when the language parameter is omitted or set to unk, allowing it to identify and process multiple languages within the same document, including mixed English and Japanese text. This is the recommended approach for multi-language documents because forcing a single language can degrade recognition of the other language's characters.

Exam trap

AI-103 often tests the misconception that you must specify a single language for OCR, when the Read API's automatic detection via omitted or unk language is the correct multi-language approach.

How to eliminate wrong answers

Option A is wrong because the legacy OCR v3.0 endpoint is deprecated and does not provide the multi-language detection capability of the Read API; it also lacks the modern layout and handwriting features. Option B is wrong because splitting pages into separate calls is unnecessary and operationally inefficient; the Read API handles mixed-language pages in a single call. Option C is wrong because setting language=en-US exclusively forces English recognition and will misrecognize Japanese characters, hurting accuracy.

Full explanation →

106

MCQeasy

You are provisioning an Azure AI services account and need to select the kind of resource. Which kind parameter value should you select for an Azure OpenAI deployment?

A.CognitiveSearch

B.MachineLearning

C.OpenAI

D.BotService

AnswerC

The kind parameter value OpenAI provisions an Azure OpenAI resource, exposing the OpenAI model deployments and completions API rather than the general multi-service cognitive endpoints. Selecting this kind is what enables GPT and embedding model deployments within the Azure AI services account.

Why this answer

For an Azure OpenAI deployment, the resource kind must be 'OpenAI'. This kind provisions an Azure OpenAI resource that provides access to OpenAI models like GPT-4, GPT-3.5, DALL-E, and Whisper through Azure's infrastructure. The 'OpenAI' kind is specifically designed for Azure OpenAI Service, enabling features such as model deployment, fine-tuning, and content filtering.

Other kinds correspond to different Azure AI services and do not support OpenAI model deployments.

Exam trap

AI-103 often tests the confusion between different Azure AI resource kinds, especially when candidates assume that any AI-related service can host OpenAI models, leading them to select 'MachineLearning' or 'CognitiveSearch' incorrectly.

How to eliminate wrong answers

Option A is wrong because 'CognitiveSearch' is the kind for Azure Cognitive Search, which provides search capabilities over content, not OpenAI model hosting. Option B is wrong because 'MachineLearning' is the kind for Azure Machine Learning workspaces, used for building, training, and deploying custom machine learning models, not for accessing OpenAI's pre-trained models. Option D is wrong because 'BotService' is the kind for Azure Bot Service, which is used to develop conversational AI bots, not for deploying OpenAI models.

Full explanation →

107

MCQmedium

Your organization has deployed an Azure AI Custom Vision model. You need to automate the retraining pipeline whenever new training images are uploaded to an Azure Blob Storage container. Which Azure service should you use to trigger the retraining pipeline?

A.Azure ExpressRoute monitor

B.Azure Backup service

C.Azure Event Grid

D.Azure Monitor alert rules

AnswerC

Azure Event Grid delivers blob-created events from the storage account to a subscriber, so an upload immediately triggers the retraining pipeline. It provides native, near-real-time event-driven invocation, unlike polling or scheduled triggers, satisfying the automation requirement.

Why this answer

Azure Event Grid is a fully managed event-routing service that natively integrates with Blob Storage. When a blob is created in a container, Event Grid publishes a BlobCreated event that can trigger an Azure Function, Logic App, or webhook to invoke the Custom Vision retraining pipeline.

Exam trap

AI-103 often tests the confusion between Azure Monitor alerts (metric/log-based, not per-object) and Event Grid (event-driven, per-blob), luring candidates toward Monitor for storage-triggered automation.

How to eliminate wrong answers

Option A is wrong because ExpressRoute Monitor is a network-monitoring feature for ExpressRoute circuits and has no event-triggering capability for storage. Option B is wrong because Azure Backup is a data-protection service for backing up VMs, files, and databases — it does not emit blob-creation events or trigger pipelines. Option D is wrong because Azure Monitor alert rules react to metrics and logs (e.g., CPU thresholds) and are not designed for near-real-time, per-object event routing; they also lack native Blob Storage event subscriptions.

Full explanation →

108

Multi-Selectmedium

You are designing an image analysis solution using Azure AI Vision 4.0. Which TWO visual features can be requested in a single Image Analysis API analyze request? (Choose two.)

Select 2 answers

A.tags

B.Custom Vision Classification

C.read

D.Azure AI Face Analyze

E.Azure AI Document Intelligence Layout

AnswersA, C

Tags is one of the visual features supported by the Azure AI Vision 4.0 Image Analysis analyze request. Multiple features can be combined in one call, so tags can be requested alongside another feature such as read.

Why this answer

Option A (tags) is correct because the Image Analysis 4.0 analyze request accepts a 'visualFeatures' parameter that can include 'tags' to return content tags describing the image. Option C (read) is also correct because 'read' is a valid visual feature in the same 'visualFeatures' list, enabling OCR text extraction within the same analyze call. Both features belong to the Azure AI Vision Image Analysis 4.0 API surface and can be combined in one request, e.g., visualFeatures=tags,read.

Option B (Custom Vision Classification) is not part of the Image Analysis API; it is a separate Custom Vision service with its own prediction endpoint. Option D (Azure AI Face Analyze) is a distinct Face service API, not a visual feature of Image Analysis. Option E (Azure AI Document Intelligence Layout) is a separate Document Intelligence model/API and cannot be requested through the Image Analysis analyze operation.

Exam trap

AI-103 often tests service boundaries — the trap is selecting features from adjacent Azure services (Custom Vision, Face, Document Intelligence) as if they were Image Analysis features, when only native visual features like tags and read belong in the analyze request.

Full explanation →

109

MCQeasy

When creating a new project in the Custom Vision portal, you need to choose a classification type. You want to tag an image with multiple independent tags (e.g., 'outdoor', 'sunset', 'mountain'). Which classification type should you select?

A.Multilabel

B.Object Detection

C.Binary Classification

D.Multiclass

AnswerA

Multilabel classification applies independent sigmoid outputs per tag, so each tag is scored separately rather than competing in a single softmax set. This satisfies the stem's requirement that an image carry multiple independent tags simultaneously, such as 'outdoor', 'sunset' and 'mountain' together.

Why this answer

Multilabel classification lets a single image carry multiple independent tags simultaneously, which is exactly what is needed when an image can be 'outdoor', 'sunset', and 'mountain' at the same time. Each tag is evaluated independently, so the model outputs a probability per tag rather than forcing a single winning class. This matches the requirement of applying multiple non-exclusive labels.

Exam trap

AI-103 often tests the distinction between Multilabel (multiple non-exclusive tags) and Multiclass (one exclusive label), and candidates pick Multiclass because it sounds like it handles 'multiple' classes.

How to eliminate wrong answers

Option B is wrong because Object Detection returns bounding boxes and labels for located objects, not image-level tags — it answers 'where is the object' rather than 'what tags apply to this image'. Option C is wrong because Binary Classification only decides between two mutually exclusive classes (A or B), so it cannot represent three independent tags. Option D is wrong because Multiclass assigns exactly one label from a set of mutually exclusive classes, meaning an image could be tagged 'sunset' OR 'mountain' but never both.

Full explanation →

110

MCQeasy

You are reviewing the pricing details for Azure AI Language. How is billing calculated for pre-built text analysis features such as sentiment analysis and key phrase extraction?

A.Hourly uptime of the resource endpoint

B.Number of unique client application connections

C.Total gigabytes of training data uploaded

D.Number of text records (documents) processed

AnswerD

Pre-built Azure AI Language features bill per text record submitted, not per character, API call or compute hour. Sentiment analysis and key phrase extraction each charge on the document count processed, so the constraint in the stem is satisfied by counting text records.

Why this answer

Azure AI Language pre-built features such as sentiment analysis and key phrase extraction are consumption-based services billed per text record (document) processed. Each API call that submits a document for analysis counts as one billable transaction, regardless of document length within the per-document size limit. This pay-per-use model aligns cost directly with the volume of text analyzed rather than with infrastructure uptime or training activity.

Exam trap

AI-103 often tests the misconception that Azure AI Language bills on resource uptime or data storage like IaaS services, when in fact pre-built cognitive features are metered per text record processed.

How to eliminate wrong answers

Option A is wrong because Azure AI Language is a serverless, consumption-based cognitive service — you are not billed for endpoint uptime or provisioned hours; that model applies to always-on resources like Azure App Service or provisioned throughput units, not standard pre-built Language API calls. Option B is wrong because billing is not tied to the number of client applications or connections; the service meters transactions per document submitted, so ten apps or one app sending the same volume incur identical charges. Option C is wrong because training data volume is irrelevant to pre-built features — no training occurs for sentiment analysis or key phrase extraction, and even for custom models, billing is based on text records and training hours, not gigabytes uploaded.

Full explanation →

111

MCQeasy

You are planning to deploy an Azure AI solution that includes multiple Azure Cognitive Services resources. You need to ensure that the resources can be managed together and that costs are consolidated. What should you use?

A.Deploy each resource in a separate subscription and use Azure Cost Management to aggregate costs.

B.Use an Azure Cognitive Services multi-service resource.

C.Configure each resource to send diagnostic logs to a central Log Analytics workspace.

D.Create each resource in the same resource group and subscription.

AnswerB

A multi-service resource allows you to access multiple Azure Cognitive Services APIs (such as Vision, Language, Speech) with a single key and endpoint. It consolidates billing and management, and you can use one resource to call several services. This simplifies deployment and cost tracking for solutions that use multiple AI services.

Why this answer

An Azure Cognitive Services multi-service resource provides a single Azure resource that can be used to access multiple cognitive services, such as Computer Vision, Text Analytics, and Speech. It simplifies management, offers a single key and endpoint, and consolidates billing for those services. This is the recommended approach when you need to use several AI services together.

Exam trap

The trap here is thinking that resource group or subscription placement alone achieves consolidated management, when a multi-service resource is specifically designed for that purpose.

Full explanation →

112

MCQmedium

You are developing a multi-agent orchestration solution using Azure AI Agent Service and want to ensure that user inputs are safely evaluated before they are passed to the model. Which component should you configure to inspect and filter incoming prompts for malicious content and restricted topics?

A.Azure Key Vault access policies

B.Azure AI Content Safety integrated into the model gateway

C.Azure Cognitive Search semantic ranker

D.Azure API Management rate limiting policy

AnswerB

Azure AI Content Safety integrated into the model gateway inspects prompts in transit, applying harm categories and restricted-topic filters before the model receives them. This satisfies the stem's requirement to evaluate user inputs safely, blocking malicious content at the gateway layer rather than relying on post-generation moderation or agent-level prompt instructions.

Why this answer

Azure AI Content Safety integrated into the model gateway inspects incoming prompts for harmful content, jailbreak attempts, and restricted topics before they reach the model, providing a centralized safety layer for multi-agent orchestration. This is the correct component for pre-model prompt evaluation and filtering.

Exam trap

AI-103 often tests whether candidates confuse infrastructure security controls (Key Vault, API Management) with semantic content safety, when only Azure AI Content Safety evaluates prompts for malicious or restricted content.

How to eliminate wrong answers

Option A is wrong because Azure Key Vault access policies manage secrets and cryptographic keys, not prompt content inspection or topic filtering. Option C is wrong because the Azure Cognitive Search semantic ranker reorders search results for relevance; it does not evaluate prompts for malicious content. Option D is wrong because API Management rate limiting controls request throughput and quotas, not the semantic safety of prompt content.

Full explanation →

113

MCQmedium

You are using Custom Vision to train an object detection model. You want to evaluate model performance before publishing an iteration. Which metric represents the overall accuracy of the model across all classes at various confidence thresholds?

A.Perplexity

B.Word Error Rate (WER)

C.BLEU Score

D.Mean Average Precision (mAP)

AnswerD

Mean Average Precision (mAP) averages precision across all classes and confidence thresholds, giving one overall accuracy figure for the object detection iteration. This directly satisfies the stem's requirement to evaluate performance across all classes at various confidence thresholds before publishing, unlike single-threshold metrics such as precision or recall alone.

Why this answer

Mean Average Precision (mAP) is the standard metric for object detection models because it aggregates precision across all classes and Intersection over Union (IoU) thresholds. It computes the average precision (AP) for each class—the area under the precision-recall curve—and then averages those AP values. This gives a single score that reflects how well the model localizes and classifies objects across different confidence thresholds, which is exactly what Custom Vision reports for object detection iterations.

Exam trap

AI-103 often tests whether candidates confuse metrics from different AI domains—such as using NLP metrics (perplexity, WER, BLEU) for computer vision tasks—so the trap is selecting a familiar but irrelevant metric instead of the correct object detection metric, mAP.

How to eliminate wrong answers

Option A is wrong because perplexity is a language modeling metric that measures how well a probability distribution predicts a sample, not object detection. Option B is wrong because Word Error Rate (WER) is used in speech recognition to count word-level transcription errors, not visual object detection. Option C is wrong because BLEU Score is a text generation metric for comparing machine-translated text to reference translations, unrelated to object detection.

Full explanation →

114

MCQhard

You are deploying a custom text classification model using Azure AI Language. The model undergoes frequent retraining as new business data arrives. You need to establish a model versioning strategy that allows zero-downtime client updates while ensuring seamless rollback capability if regression occurs. What is the recommended architectural approach?

A.Delete the original resource before deploying the updated model version.

B.Overwrite the existing model artifact in place with the new training run using the exact same name.

C.Create a completely new Azure AI resource in a different subscription for every single model iteration.

D.Deploy each model version with a unique deployment name and update the application client configuration or routing rules to shift traffic progressively.

AnswerD

Unique deployment names let each model version run concurrently behind a stable endpoint, so clients switch by updating routing rather than redeploying. Progressive traffic shifting gives zero-downtime cutover, and reverting the routing rule restores the prior version instantly if regression appears.

Why this answer

Azure AI Language custom models are consumed through deployments, and each deployment has a unique name that maps to a specific model version. By deploying each new version under a unique deployment name and updating client configuration or routing rules, you can shift traffic progressively and roll back instantly by pointing clients back to the previous deployment. This achieves zero-downtime updates and seamless rollback.

Exam trap

AI-103 often tests whether candidates think retraining overwrites a model in place, missing that Azure AI Language uses named deployments to enable versioning, progressive rollout, and rollback.

How to eliminate wrong answers

Option A is wrong because deleting the original resource destroys the existing model and causes downtime with no rollback path. Option B is wrong because overwriting the model artifact in place removes the previous version, eliminating rollback capability and risking regression. Option C is wrong because creating a new Azure AI resource per iteration is operationally heavy, costly, and unnecessary when deployments within the same resource support versioning.

Full explanation →

115

Multi-Selectmedium

When configuring role-based access control (RBAC) in Azure AI Foundry and Azure OpenAI, which TWO built-in Azure roles are commonly assigned to developers and services? (Choose TWO)

Select 2 answers

A.Storage Blob Data Owner

B.Cognitive Services OpenAI User

C.Virtual Machine Contributor

D.Network Contributor

E.Azure AI Developer

AnswersB, E

Cognitive Services OpenAI User grants data-plane access for calling Azure OpenAI models, inference and completions, without permitting key or account management. It satisfies the stem's requirement for a built-in role assigned to developers and services needing model invocation, complementing the control-plane Cognitive Services OpenAI Contributor role.

Why this answer

Option B, Cognitive Services OpenAI User, is correct because it is the built-in role that grants data-plane access to Azure OpenAI resources, allowing developers and services to call inference endpoints such as chat completions and embeddings without granting management permissions. Option E, Azure AI Developer, is correct because it is the built-in role designed for building and managing Azure AI Foundry projects and hubs, giving developers the permissions needed to create projects, connections, and deploy models while scoping access appropriately. Option A, Storage Blob Data Owner, is not appropriate here since it governs blob storage data access, not Azure OpenAI or AI Foundry RBAC.

Option C, Virtual Machine Contributor, manages VMs and is unrelated to AI service access. Option D, Network Contributor, manages networking resources and does not grant the AI/OpenAI data-plane permissions required for this scenario.

Exam trap

AI-103 often tests the distinction between control-plane roles (Azure AI Developer, Contributor) and data-plane roles (Cognitive Services OpenAI User), so candidates mistakenly pick generic roles like Storage Blob Data Owner or Network Contributor that have nothing to do with AI model access.

Full explanation →

116

MCQmedium

You are building an automated support ticketing system. You need to extract names, email addresses, and phone numbers from user messages while leaving non-sensitive text untouched. Which feature of Azure AI Language should you use?

A.Key phrase extraction

B.Personally Identifiable Information (PII) detection

C.Named Entity Recognition (NER)

D.Custom text classification

AnswerB

PII detection identifies and redacts entities such as names, email addresses and phone numbers within text, returning offsets for each. This extracts exactly the sensitive fields requested while leaving surrounding non-sensitive content unchanged, matching the ticketing requirement.

Why this answer

PII detection in Azure AI Language is purpose-built to identify and categorize sensitive personal data such as names, email addresses, phone numbers, SSNs, and credit card numbers, and it returns entity categories plus redaction offsets so you can mask or remove them. Because the requirement is specifically to isolate sensitive identifiers while leaving other text untouched, PII detection is the exact fit. It is essentially NER with a curated, privacy-focused entity taxonomy and built-in masking support.

Exam trap

AI-103 often tests the overlap between NER and PII detection, tricking candidates into choosing generic NER when the scenario explicitly requires privacy-specific categories like phone numbers or SSNs that only PII detection returns.

How to eliminate wrong answers

Option A is wrong because key phrase extraction surfaces salient topics and noun phrases for summarization or search, not sensitive data categories, and it cannot distinguish a person's name from any other proper noun. Option C is wrong because generic NER returns broad entity types (Person, Location, Organization, DateTime, Quantity) but does not include the privacy-specific categories like PhoneNumber, Email, or USSocialSecurityNumber that PII detection provides, nor does it offer redaction. Option D is wrong because custom text classification assigns whole documents or sentences to user-defined labels; it does not extract entity spans from within text.

Full explanation →

117

MCQhard

You are implementing a text classification pipeline using Azure AI Language. You want to deploy your trained custom model to a production endpoint. What is required before you can assign traffic to the model deployment in Language Studio?

A.You must successfully complete model training and create a deployment in Language Studio.

B.You must export the model container and deploy it to an Azure Kubernetes Service (AKS) cluster.

C.You must submit a support ticket to Microsoft to enable custom model endpoints.

D.You must convert the model files to ONNX format.

AnswerA

A deployment must exist before traffic can be assigned, and it can only be created after training completes successfully. Language Studio therefore requires both the trained model and its deployment as prerequisites for routing production traffic.

Why this answer

In Azure AI Language's custom text classification, the workflow requires two distinct steps before traffic can be routed: first, you must train the model (which produces a trained model artifact tied to your project), and second, you must create a deployment that hosts that trained model at an endpoint. Only after a deployment exists can you assign traffic to it in Language Studio. The [CORRECT] option captures both prerequisites — successful training and deployment creation — which are the actual gating requirements.

Exam trap

AI-103 often tests the distinction between training a model and deploying it — candidates frequently assume that a successfully trained model is immediately queryable, but Azure AI Language requires an explicit deployment step before any endpoint traffic can be assigned.

How to eliminate wrong answers

Option B is wrong because Azure AI Language custom models are fully managed PaaS — you do not export containers or run them on AKS; that pattern belongs to Azure Machine Learning or Azure AI Vision Docker containers, not Language Studio custom classification. Option C is wrong because custom model endpoints are a standard, self-service feature in Azure AI Language; no support ticket or feature-flag enablement is required. Option D is wrong because ONNX conversion is not part of the Azure AI Language custom text classification deployment path — the service handles model packaging and serving internally, and ONNX is relevant to other runtimes (e.g., Windows ML, ONNX Runtime) not this deployment flow.

Full explanation →

118

MCQhard

You are tuning the prompt for a generative AI application in Azure AI Foundry. The model frequently hallucinates when asked complex multi-part questions. You decide to use chain-of-thought (CoT) prompting. How should you structure the prompt to implement CoT effectively?

A.Set the model seed parameter to a fixed integer value for reproducibility

B.Reduce the maximum completion token limit to 150 tokens

C.Instruct the model to think step-by-step and show its intermediate reasoning before providing the final answer

D.Instruct the model to output its response exclusively in JSON format with strict schema validation

AnswerC

Instructing the model to reason step-by-step and expose intermediate reasoning decomposes complex multi-part questions into sequential sub-steps, reducing hallucination by grounding each inference in prior output. This directly satisfies the stem's constraint: complex multi-part questions that trigger hallucination. CoT works by allocating generation tokens to intermediate computation rather than jumping straight to a final answer.

Why this answer

Chain-of-thought prompting works by instructing the model to reason step-by-step and show intermediate reasoning before the final answer. This decomposition reduces hallucination on multi-part questions because the model is forced to work through each part logically rather than jumping to a conclusion. The prompt should explicitly ask for step-by-step thinking.

Exam trap

The trap is confusing output formatting (JSON) or sampling parameters (seed, temperature) with reasoning techniques — CoT is specifically about eliciting intermediate reasoning steps.

How to eliminate wrong answers

Option A is wrong because setting a seed only affects reproducibility, not reasoning quality or hallucination. Option B is wrong because reducing max tokens truncates reasoning, which harms CoT and can increase errors. Option D is wrong because forcing JSON output constrains format but does not improve reasoning; it may even suppress the step-by-step thinking CoT requires.

Full explanation →

119

MCQmedium

Your application uploads images to Azure AI Vision via URLs. One of your URLs is failing with an invalid image error, even though the image opens correctly in web browsers. What is the most likely reason?

A.Azure AI Vision does not support URL-based inputs; you must upload binary bytes.

B.The image file size exceeds 500 MB.

C.Image URLs are restricted to PNG format only.

D.The URL requires HTTP Basic Authentication or redirects that the vision service cannot resolve.

AnswerD

Azure AI Vision fetches the URL server-side without browser credentials or redirect handling, so a URL behind HTTP Basic Authentication or a redirect chain fails as an invalid image. This explains the failure despite the image opening correctly in a browser, satisfying the stem's scenario.

Why this answer

Azure AI Vision's URL-based image input requires the service to fetch the image directly from the provided URL. If the URL is behind HTTP Basic Authentication, requires a token, or returns a redirect (3xx) that the service cannot follow, the fetch fails and the service returns an 'invalid image' error even though a browser (which handles auth prompts and redirects) can open it fine. The fix is to make the image publicly accessible or upload the binary bytes instead.

Exam trap

AI-103 often tests the difference between what a browser can do (handle auth prompts, cookies, redirects) and what a server-side service fetch can do, so candidates wrongly assume 'it opens in my browser' means the URL is valid for the API.

How to eliminate wrong answers

Option A is wrong because Azure AI Vision explicitly supports URL-based image inputs via the 'url' property in the ImageAnalysis request body — binary upload is an alternative, not a requirement. Option B is wrong because the documented size limit for images is 4 MB (and 20 MB for some operations), not 500 MB, and an oversized image would typically return a size-related error rather than a generic invalid image error. Option C is wrong because Azure AI Vision supports JPEG, PNG, GIF, BMP, and other common formats — PNG is not the only allowed format.

Full explanation →

120

MCQhard

You are implementing a multi-region disaster recovery strategy for an Azure OpenAI deployment. If the primary region goes down, how should you configure traffic redirection to the secondary region using Azure Front Door?

A.Use geographic routing based on the client IP location.

B.Use weighted routing with a 50/50 split between regions.

C.Use priority routing with the primary region set to priority 1 and the secondary region set to priority 2.

D.Use performance-based routing to the lowest latency node.

AnswerC

Priority routing directs all traffic to the priority 1 primary region and only fails over to priority 2 when the primary becomes unhealthy, satisfying the disaster recovery requirement for automatic redirection during a regional outage.

Why this answer

Azure Front Door priority routing assigns a priority number to each origin, and traffic is sent to the lowest-numbered (highest-priority) healthy origin first. Setting the primary region to priority 1 and the secondary to priority 2 ensures all traffic goes to the primary during normal operation and automatically fails over to the secondary only when the primary becomes unhealthy. This is the standard active-passive DR pattern.

Exam trap

AI-103 often tests whether candidates confuse weighted or performance-based routing with failover, when only priority routing provides deterministic active-passive redirection.

How to eliminate wrong answers

Option A is wrong because geographic routing directs users based on their IP location, which is used for latency or compliance routing, not for active-passive failover — it would send some users to the secondary even when the primary is healthy. Option B is wrong because weighted routing with a 50/50 split is active-active load distribution, not DR failover; it does not prioritize the primary region. Option D is wrong because performance-based routing selects the lowest-latency origin, which may route to the secondary even when the primary is healthy, and it does not provide deterministic failover semantics.

Full explanation →

121

MCQmedium

Your application calls Azure AI Vision API and receives an HTTP 429 Too Many Requests response. What does this error signify, and how should your application handle it?

A.The service is undergoing maintenance. Fail over to a secondary region.

B.The image file format is unsupported. Convert to PNG.

C.Your API key is invalid. Refresh your subscription key immediately.

D.You have exceeded your pricing tier rate limits. Implement exponential backoff and retry the request.

AnswerD

HTTP 429 signals the subscription has surpassed its pricing tier's request rate limit, so the service is throttling calls. Exponential backoff with retry spacing requests progressively further apart, letting the rate window reset and avoiding repeated rejections while satisfying the need to handle the throttling gracefully.

Why this answer

HTTP 429 Too Many Requests indicates that the application has exceeded the rate limits for its Azure AI Vision pricing tier (e.g., transactions per second or per minute). The correct handling is to implement exponential backoff with jitter and retry the request after a delay, respecting the 'Retry-After' header if provided. This prevents overwhelming the service and allows the request to succeed once the rate limit resets.

Exam trap

AI-103 often tests the distinction between 429 (rate limit) and 401/403 (auth) or 400 (bad request); candidates who misattribute 429 to an invalid key or maintenance will choose the wrong remediation.

How to eliminate wrong answers

Option A is wrong because 429 does not indicate maintenance; maintenance would typically return 503 Service Unavailable. Option B is wrong because an unsupported image format returns 400 Bad Request, not 429. Option C is wrong because an invalid API key returns 401 Unauthorized or 403 Forbidden, not 429.

Full explanation →

122

MCQhard

You are implementing opinion mining as part of sentiment analysis in Azure AI Language. An incoming review states, 'The room was spacious, but the staff was rude.' What does opinion mining return for this sentence?

A.A binary classification indicating whether the review is spam or genuine.

B.A single neutral sentiment score for the entire document.

C.A translated version of the review in English.

D.Aspect targets ('room', 'staff') along with their corresponding evaluations ('spacious', 'rude') and individual sentiments.

AnswerD

Opinion mining extracts aspect targets and their evaluations, returning 'room' with 'spacious' (positive) and 'staff' with 'rude' (negative), each with individual sentiment. This aspect-level breakdown satisfies the stem's requirement, unlike sentence-level sentiment that would return one mixed score.

Why this answer

Opinion mining in Azure AI Language returns aspect-based sentiment analysis, identifying aspect targets (e.g., 'room', 'staff') and their corresponding evaluations (e.g., 'spacious', 'rude') with individual sentiment polarities. For the given review, it would return 'room' as positive and 'staff' as negative, providing granular insights beyond document-level sentiment.

Exam trap

AI-103 often tests the difference between document-level sentiment and aspect-based sentiment, and candidates may incorrectly choose a single sentiment score when the question asks for opinion mining output.

How to eliminate wrong answers

Option A is wrong because spam classification is not a feature of opinion mining; it is a separate task. Option B is wrong because opinion mining does not return a single neutral sentiment score; it returns aspect-level sentiments, and the overall sentiment might be mixed. Option C is wrong because translation is a different Azure AI service (Translator), not part of sentiment analysis.

Full explanation →

123

MCQhard

You are processing large PDF documents. You only need the text in the table of contents. What is the most cost-effective approach?

A.Use a custom template model

B.Split the PDF into individual pages before uploading

C.Use the 'pages' query parameter

D.Process the entire PDF

AnswerC

Using the 'pages' query parameter restricts processing to the table of contents pages, so you are billed only for those pages rather than the entire PDF. This directly satisfies the cost-effectiveness constraint, since Azure AI Document Intelligence charges per page analysed.

Why this answer

The 'pages' query parameter in Azure AI Document Intelligence allows you to specify which pages to process, such as a single page or a range. If you only need the table of contents, you can identify the page number(s) containing it and pass the 'pages' parameter to limit processing to those pages. This reduces the number of pages processed, thereby lowering cost, since Document Intelligence pricing is based on the number of pages analyzed.

Exam trap

AI-103 often tests the misconception that you must process the entire document or split it manually, when in fact the 'pages' parameter provides a cost-effective way to target specific pages.

How to eliminate wrong answers

Option A is wrong because a custom template model is used for extracting structured data from forms with a consistent layout, not for selectively processing pages; it would still process the entire document unless combined with the 'pages' parameter. Option B is wrong because splitting the PDF into individual pages before uploading is a manual, inefficient approach that still requires processing each page separately and may not be cost-effective if you only need one page. Option D is wrong because processing the entire PDF incurs cost for all pages, which is not cost-effective when only the table of contents is needed.

Full explanation →

124

MCQmedium

You are evaluating a Custom Vision classification model. You want to programmatically test the model against a validation dataset without using the portal UI. Which endpoint should your script call?

A.POST /computervision/imageanalysis:analyze

B.POST /customvision/v3.3/Training/{projectId}/images

C.GET /vision/v3.2/models

D.POST /customvision/v3.3/Prediction/{projectId}/classify/iterations/{publishedName}/image

AnswerD

The Prediction endpoint accepts an image directly and returns classification results for the published iteration, so a script can test against a validation dataset without portal interaction. The publishedName path segment targets the deployed iteration, satisfying the requirement to bypass the UI.

Why this answer

To programmatically test a published Custom Vision classification model, you call the prediction endpoint POST /customvision/v3.3/Prediction/{projectId}/classify/iterations/{publishedName}/image. This endpoint accepts an image and returns the predicted tags with confidence scores for the specified published iteration. It is the correct REST path for classification predictions outside the portal.

Exam trap

AI-103 often tests the distinction between Training and Prediction endpoints and between classify and detect paths, and candidates pick the training endpoint or the wrong project-type path.

How to eliminate wrong answers

Option A is wrong because /computervision/imageanalysis:analyze belongs to Azure AI Vision's Image Analysis service, not Custom Vision, and it performs general analysis rather than invoking your trained model. Option B is wrong because /customvision/v3.3/Training/{projectId}/images is a training endpoint used to upload and tag images, not to run predictions. Option C is wrong because GET /vision/v3.2/models is a legacy Computer Vision endpoint that lists models and has nothing to do with Custom Vision predictions.

Full explanation →

125

MCQeasy

You want to monitor the cost and latency of your agentic application in production. Which service should you integrate with your Azure AI Foundry project?

A.Azure OpenAI Studio.

B.Azure AI Search.

C.Azure Application Insights.

D.Azure AI Content Safety.

AnswerC

Application Insights captures request traces, dependency calls and latency telemetry from the agentic application, and its token and usage metrics feed cost monitoring, satisfying the requirement to track both cost and latency in production. Azure Monitor alone lacks the application-level distributed tracing needed here.

Why this answer

Azure Application Insights is the Application Performance Monitoring (APM) service that Azure AI Foundry integrates with to capture traces, latency, token usage, and cost telemetry for agentic applications. It provides the dashboards, query language (KQL), and alerting needed to monitor production behavior.

Exam trap

AI-103 often tests whether candidates confuse development surfaces (OpenAI Studio) with observability services — the trap is picking a playground or search service when the question asks about production cost and latency monitoring.

How to eliminate wrong answers

Option A is wrong because Azure OpenAI Studio is a development and experimentation surface for prompts and deployments, not a production monitoring or cost-tracking service. Option B is wrong because Azure AI Search is the retrieval/index service; it reports search metrics but not end-to-end agent latency or cost. Option D is wrong because Azure AI Content Safety filters harmful content; it does not collect performance or billing telemetry.

Full explanation →

126

MCQhard

You are designing a multi-agent solution in Azure AI Foundry where agents collaborate asynchronously. One agent produces an intermediate JSON payload that must be strictly validated before being consumed by the next agent. What feature should you implement?

A.Model temperature set to 1.0

B.Azure API Management XML transformation policy

C.Structured Outputs with strict JSON Schema enforcement

D.Azure AI Content Safety hate speech detection

AnswerC

Structured Outputs constrain generation to a strict JSON Schema, guaranteeing the payload conforms to declared types and required fields before the next agent consumes it. This satisfies the stem's strict validation constraint, unlike prompt-based formatting requests, which cannot guarantee schema compliance across asynchronous handoffs.

Why this answer

Structured Outputs with strict JSON Schema enforcement forces the model to emit JSON that conforms to a defined schema, which is exactly what is needed to validate an intermediate payload before the next agent consumes it. It reduces parsing failures and ensures required fields, types, and enums are respected.

Exam trap

AI-103 often tests the difference between prompt-level requests for JSON and true schema-enforced Structured Outputs, tempting candidates to pick content safety or gateway policies instead.

How to eliminate wrong answers

Option A is wrong because setting temperature to 1.0 increases randomness and makes output less predictable, which is the opposite of strict validation. Option B is wrong because Azure API Management XML transformation policies handle XML, not JSON schema validation for model outputs, and they operate at the gateway layer rather than constraining generation. Option D is wrong because Azure AI Content Safety hate speech detection filters harmful content; it does not validate JSON structure or enforce a schema.

Full explanation →

127

MCQhard

You are training a Custom Vision object detection model. You have 1,000 images, but only 200 of them contain the rare object you want to detect. The other 800 images are background images without the object. What impact will this class imbalance have on training, and how should you address it?

A.Reduce training iteration time to 1 minute.

B.Convert the project to multilabel classification.

C.Ensure that you have tagged at least 15 bounding boxes per tag; having background images is beneficial for reducing false positives.

D.Delete all 800 background images immediately, as they will cause training to fail.

AnswerC

Custom Vision requires roughly 15 bounding boxes per tag, so 200 positive images suffice, and the 800 background images suppress false positives by teaching the model what absence looks like. This satisfies the stem's class-imbalance constraint without discarding the valuable negative examples.

Why this answer

Custom Vision object detection requires a minimum of 15 tagged bounding boxes per tag for a tag to be trainable, and background (negative) images are explicitly beneficial because they teach the model what the object is not, reducing false positives. With only 200 positive images, keeping the 800 backgrounds provides valuable negative examples. The correct guidance is to ensure sufficient tagged instances while retaining background images.

Exam trap

The trap here is assuming background images are harmful and should be deleted, when in object detection they are actually beneficial negative examples that reduce false positives.

How to eliminate wrong answers

Option A is wrong because reducing training time to 1 minute has no bearing on class imbalance and would likely under-train the model. Option B is wrong because multilabel classification is a different problem type and does not address object detection imbalance. Option D is wrong because deleting background images removes useful negative examples and can increase false positives; it also won't cause training to fail, so the premise is false.

Full explanation →

128

MCQmedium

You are processing medical intake forms using Azure AI Language. You need to identify and categorize medical terms, medications, and dosages. Which feature should you use?

A.Health text analytics

B.Key phrase extraction

C.Standard Named Entity Recognition (NER)

D.Custom text classification

AnswerA

Health text analytics applies a domain-specific model that extracts and categorises medical entities such as medications, dosages, and clinical terms from unstructured text. This directly satisfies the stem's requirement to identify and categorise medical terminology within intake forms, which generic entity extraction cannot reliably achieve.

Why this answer

Health text analytics in Azure AI Language is specifically designed for the medical domain and can identify medical entities such as medications, dosages, diagnoses, and symptoms. It goes beyond generic NER by recognizing healthcare-specific terminology and relationships. This makes it the correct choice for processing medical intake forms where medications and dosages must be categorized.

Exam trap

AI-103 often tests whether candidates assume standard NER covers medical terminology, when in fact healthcare-specific extraction requires the dedicated health text analytics feature.

How to eliminate wrong answers

Option B is wrong because key phrase extraction returns salient phrases but does not categorize them as medical terms, medications, or dosages. Option C is wrong because standard NER recognizes general entity types like Person, Location, and Organization, not medical-specific categories such as medication or dosage. Option D is wrong because custom text classification assigns whole-document or sentence-level labels from your own taxonomy, not entity extraction of medical terms.

Full explanation →

129

MCQeasy

You are planning to deploy an Azure AI solution that requires a dedicated endpoint for a custom model. You need to choose a deployment option that provides a static endpoint and supports autoscaling. What should you use?

A.Azure Functions with an HTTP trigger.

B.Azure Machine Learning managed online endpoint.

C.Azure Kubernetes Service (AKS) with a custom inference server.

D.Azure Container Instances with a public IP address.

AnswerB

Azure Machine Learning managed online endpoints provide a static scoring URI, support autoscaling, and handle infrastructure management. They are designed for deploying models with dedicated endpoints and can integrate with Azure AI services. This meets the requirement for a static endpoint and autoscaling.

Why this answer

Azure Machine Learning managed online endpoints offer a static endpoint and autoscaling, making them ideal for deploying custom models. They abstract infrastructure management and integrate with Azure AI services. Other options lack either static endpoints or autoscaling capabilities.

Exam trap

The trap here is assuming that Azure Functions or AKS provide the same managed endpoint experience as Azure Machine Learning managed online endpoints.

Full explanation →

130

MCQeasy

Your team needs to extract the most important talking points from research papers using the Azure AI Language service. You want the service to return a list of salient terms and phrases from the input document. Which API task should you execute?

A.Key Phrase Extraction

B.PII Detection

C.Sentiment Analysis

D.Language Detection

AnswerA

Key Phrase Extraction returns the salient terms and phrases present in the input document, which is exactly the talking-points output required. It analyses the supplied research paper text directly, so no training or custom model is needed to surface the main concepts.

Why this answer

Key Phrase Extraction is the Azure AI Language service feature specifically designed to identify and return the main concepts or talking points from unstructured text. It analyzes the input document and outputs a list of salient terms and phrases that represent the key ideas, which directly matches the requirement to extract important talking points from research papers. The other options serve different purposes: PII Detection identifies personal information, Sentiment Analysis determines emotional tone, and Language Detection identifies the language of the text.

Exam trap

AI-103 often tests the distinction between different Azure AI Language service tasks, and candidates may confuse Key Phrase Extraction with Sentiment Analysis or PII Detection if they do not carefully read the requirement for 'salient terms and phrases'.

How to eliminate wrong answers

Option B is wrong because PII Detection is used to identify and redact personally identifiable information such as names, addresses, and phone numbers, not to extract key phrases. Option C is wrong because Sentiment Analysis determines the emotional tone (positive, negative, neutral) of the text, not the salient terms. Option D is wrong because Language Detection identifies the language in which the text is written, not the key phrases within it.

Full explanation →

131

MCQeasy

You are building an image search engine. You want to extract visual feature tags from user photos using Azure AI Vision 4.0. Which visual feature should you request?

A.features=read

B.features=caption

C.features=smartCrops

D.features=tags

AnswerD

Requesting `features=tags` returns content tags describing objects, scenes and actions detected in each photo, which is precisely the visual feature extraction the search engine needs for indexing. Azure AI Vision 4.0 exposes this through the Image Analysis API, satisfying the stem's requirement to derive searchable visual feature tags from user photos.

Why this answer

The features=tags parameter in Azure AI Vision 4.0 returns a list of content tags describing the objects, scenes, and concepts detected in an image, which is exactly what an image search engine needs for indexing and retrieval. Tags are the visual feature purpose-built for extracting searchable descriptors from photos.

Exam trap

AI-103 often tests the confusion between the read feature (OCR text extraction) and the tags feature (visual concept extraction), so candidates who associate 'extract' with 'read' pick the wrong option.

How to eliminate wrong answers

Option A is wrong because features=read runs Optical Character Recognition (OCR) to extract printed or handwritten text, not visual feature tags. Option B is wrong because features=caption generates a single human-readable sentence describing the image, which is too coarse for tag-based search indexing. Option C is wrong because features=smartCrops returns suggested crop regions that preserve salient content, which is a composition feature, not a tagging feature.

Full explanation →

132

MCQmedium

You are developing a RAG solution using Azure AI Search. Users are searching using domain-specific acronyms that are not well-represented in standard embedding models, resulting in poor retrieval scores. What feature should you enable in Azure AI Search to improve keyword matching for exact acronyms?

A.Increasing the embedding dimension size to 3072

B.Enabling Azure AI Content Safety filters

C.Setting the semantic ranker threshold to zero

D.Hybrid search combining vector search and full-text BM25 search

AnswerD

Hybrid search pairs vector similarity with BM25 full-text matching, so exact acronym tokens that embeddings under-represent are still matched lexically. This directly satisfies the stem's requirement for improved keyword matching on domain-specific acronyms, since BM25 scores literal term overlap while vectors handle semantic similarity.

Why this answer

Hybrid search in Azure AI Search runs vector (k-NN) search and full-text BM25 keyword search in parallel, then fuses the results using Reciprocal Rank Fusion (RRF). BM25 excels at exact term matching, so domain-specific acronyms that embedding models underrepresent will still be matched lexically, dramatically improving recall for those queries. This is the standard remedy when semantic/vector retrieval alone misses rare tokens.

Exam trap

AI-103 often tests the misconception that bigger embeddings or semantic ranking alone solve all retrieval problems — the trap is overlooking that lexical BM25 matching is still essential for exact terms, acronyms, and rare tokens that embeddings handle poorly.

How to eliminate wrong answers

Option A is wrong because increasing embedding dimensions (e.g., to 3072 with text-embedding-3-large) improves general semantic capacity but does not fix out-of-vocabulary acronyms — the tokenizer still fragments them and the model has no learned representation for rare domain terms. Option B is wrong because Azure AI Content Safety filters detect and block harmful content (hate, violence, self-harm, sexual) and have nothing to do with retrieval scoring or keyword matching. Option C is wrong because setting the semantic ranker threshold to zero does not improve matching — it merely changes the cutoff for which re-ranked results are returned, and if the acronym never surfaced in L1 retrieval, semantic ranking has nothing to re-rank.

Full explanation →

133

MCQmedium

You are building an agent in Azure AI Foundry that needs to retrieve information from a company SharePoint site. Which component should you implement to ground the agent?

A.Blob Storage integration

B.Azure AI Search with a SharePoint connector

C.Direct REST API call to SharePoint

D.Microsoft Graph Data Connect directly to the model

AnswerB

Azure AI Search with a SharePoint connector indexes SharePoint content and exposes it as a queryable knowledge source, grounding agent responses in actual company documents. This satisfies the retrieval requirement, since the connector handles authentication and indexing without custom ingestion code.

Why this answer

Azure AI Search with a SharePoint connector (via the SharePoint indexer or the newer SharePoint ingestion pipeline) is the supported, first-party way to ground an Azure AI Foundry agent on SharePoint content. The connector crawls SharePoint sites, extracts documents, chunks and indexes them, and exposes them through a retriever that the agent can call as a knowledge tool. This gives the agent permission-aware, up-to-date retrieval without custom code.

Exam trap

AI-103 often tests whether candidates understand that grounding requires an indexed retrieval layer (Azure AI Search) rather than a raw API call or bulk-export tool like Graph Data Connect.

How to eliminate wrong answers

Option A is wrong because Blob Storage integration only indexes files that have been copied into Azure Blob Storage — it has no native awareness of SharePoint sites, permissions, or document updates, so it cannot ground the agent on live SharePoint content. Option C is wrong because a direct REST API call to SharePoint bypasses the indexing, chunking, embedding, and ranking pipeline that makes retrieval accurate and scalable; it also forces you to handle auth, pagination, and relevance yourself. Option D is wrong because Microsoft Graph Data Connect is a bulk-extraction pipeline for analytics (landing data in Azure Data Lake/Storage), not a real-time retrieval mechanism, and it cannot be wired directly into a model as a grounding source.

Full explanation →

134

Multi-Selectmedium

You are evaluating generative AI model outputs in Azure AI Foundry. Which TWO automated evaluation metrics assess the quality and correctness of RAG responses against source data? (Choose TWO)

Select 2 answers

A.Groundedness (measuring adherence to source context)

B.Virtual network gateway throughput

C.Storage account transaction count per hour

D.CPU utilization percentage of the Azure App Service plan

E.Relevance (measuring how well the response addresses the user query)

AnswersA, E

Groundedness scores whether the response's claims are actually supported by the retrieved source context. For RAG correctness this is the decisive metric, since it detects hallucinated content that contradicts or falls outside the supplied documents.

Why this answer

Option A, Groundedness, is correct because it is a built-in Azure AI Foundry evaluation metric that measures how well a generative AI response is supported by and adheres to the retrieved source context, which is exactly what is needed to verify RAG responses against source data and detect ungrounded or hallucinated content. Option E, Relevance, is correct because it is another built-in Foundry evaluation metric that assesses how well the model's response addresses the user's query, complementing groundedness by judging answer quality and correctness relative to the question. The remaining options are infrastructure and platform telemetry rather than AI quality metrics: B (virtual network gateway throughput) measures network bandwidth, C (storage account transaction count per hour) measures storage operations, and D (CPU utilization percentage of the Azure App Service plan) measures compute load, none of which evaluate the semantic quality or correctness of RAG responses.

Exam trap

AI-103 often tests whether candidates can distinguish AI quality metrics (groundedness, relevance) from generic infrastructure metrics (CPU, storage transactions, network throughput) that are irrelevant to model evaluation.

Full explanation →

135

MCQeasy

You are planning the disaster recovery strategy for an Azure AI Translator resource. In the event of a regional outage in East US, you want to ensure minimal downtime and continuity of service. Which deployment model should you select?

A.Deploy a primary resource in East US and a secondary resource in the paired region (West US), then implement client-side failover logic.

B.Configure an Azure Backup vault to snapshot the AI model weights daily.

C.Store all translation requests in Azure Blob Storage and replay them when the region recovers.

D.Rely on Azure's native synchronous cross-region database replication for cognitive services.

AnswerA

Azure AI Translator is a regional service, so pairing East US with its paired region West US and adding client-side failover logic achieves continuity during a regional outage. The paired-region topology satisfies the minimal-downtime constraint without a global endpoint.

Why this answer

Azure AI Translator does not provide automatic cross-region failover or built-in geo-replication for the translation endpoint. The recommended disaster recovery pattern is to deploy the resource in two regions (ideally a paired region such as East US / West US) and implement client-side failover logic that switches the endpoint and key when the primary region is unavailable. This achieves minimal downtime without relying on features the service does not offer.

Exam trap

AI-103 often tests the assumption that Azure provides automatic geo-failover for all services — the trap is that Cognitive Services like Translator require customer-implemented failover, unlike services with built-in geo-replication such as Azure SQL or Cosmos DB.

How to eliminate wrong answers

Option B is wrong because Azure Backup vaults back up data and VMs, not AI model weights — Cognitive Services models are Microsoft-managed and not exposed as backup-able artifacts. Option C is wrong because queuing requests in Blob Storage and replaying them later is a batch-recovery pattern, not a continuity strategy; it does not provide real-time translation during the outage. Option D is wrong because Azure Cognitive Services do not expose a synchronous cross-region database replication feature — there is no customer-visible database to replicate, and the service's internal replication is not a customer-configurable DR mechanism.

Full explanation →

136

MCQeasy

You are designing a generative AI application in Azure AI Foundry and want to categorize user queries into predefined intent classes before routing them to specific agents. Which prompt engineering technique should you apply?

A.Zero-shot random generation

B.Enabling JSON output mode without schema definition

C.Few-shot prompting with labeled classification examples

D.Increasing model temperature to 1.5

AnswerC

Few-shot prompting supplies labelled examples of each intent class within the prompt, letting the model infer the classification boundary without fine-tuning. This directly satisfies the requirement to categorise queries into predefined intents before routing, since the examples anchor output to your specific class labels.

Why this answer

Few-shot prompting provides the model with labeled examples of each intent class, allowing it to learn the classification pattern in-context without fine-tuning. This is the standard technique for intent categorization when you have predefined classes and want consistent, schema-aligned outputs. The labeled examples anchor the model's output format and decision boundaries.

Exam trap

AI-103 often tests the confusion between few-shot prompting (providing labeled examples in the prompt) and fine-tuning or structured output modes — candidates must recognize that in-context labeled examples define few-shot classification.

How to eliminate wrong answers

Option A is wrong because 'zero-shot random generation' is not a real technique — zero-shot prompting uses no examples, and 'random generation' implies uncontrolled output, which is the opposite of classification. Option B is wrong because enabling JSON output mode without a schema definition does not tell the model which intent classes exist or how to map queries to them; structured output alone does not solve classification. Option D is wrong because raising temperature to 1.5 increases randomness and creativity, which degrades classification consistency and increases hallucinated or off-schema labels.

Full explanation →

137

Multi-Selecthard

Which THREE features are critical for managing the lifecycle of an AI model in a production environment?

Select 3 answers

A.Model registry for version control.

B.Hardcoding all environment settings.

C.CI/CD integration for automated deployments.

D.Manual file copying via FTP.

E.Real-time monitoring of inference metrics.

AnswersA, C, E

A model registry stores versioned artefacts with lineage and metadata, enabling rollback, audit and promotion between environments. It satisfies the lifecycle requirement by tracking which model version is deployed where, supporting reproducibility and governance across training, staging and production stages.

Why this answer

Option A (Model registry for version control) is correct because a model registry tracks model artifacts, versions, lineage, and stage transitions (e.g., staging to production), which is essential for reproducibility, rollback, and governance across the model lifecycle. Option C (CI/CD integration for automated deployments) is correct because CI/CD pipelines automate testing, packaging, and deployment of models and their dependencies, enabling consistent, repeatable releases and reducing manual error during lifecycle transitions. Option E (Real-time monitoring of inference metrics) is correct because production models require continuous observation of latency, throughput, drift, and accuracy degradation so teams can trigger retraining or rollback before business impact occurs.

Option B (Hardcoding all environment settings) is not appropriate because it prevents portability across environments and breaks configuration management best practices. Option D (Manual file copying via FTP) is not appropriate because it lacks versioning, auditability, and automation, making deployments error-prone and non-reproducible.

Exam trap

The trap is selecting options that seem convenient but are anti-patterns in production, such as hardcoding settings or manual file transfers, instead of recognizing the need for automation and monitoring.

Full explanation →

138

Multi-Selecthard

You are optimizing the cost of your Azure AI infrastructure. You notice several resources are underutilized or incorrectly provisioned. Which THREE actions should you take to reduce unnecessary spending? Each correct answer presents part of the solution.

Select 3 answers

A.Downgrade development and test resources from high-tier production SKUs to free or lower standard tiers where appropriate.

B.Provision dedicated Azure OpenAI PTUs for idle development sandboxes.

C.Identify and delete abandoned or unused Azure AI resource accounts.

D.Implement Azure Cost Management budgets and anomaly detection alerts.

E.Scale out search service partitions and replicas to maximum capacity permanently.

AnswersA, C, D

Downgrading non-production workloads to free or standard tiers directly addresses the stem's underutilised and over-provisioned resources, since development and test environments rarely need production-grade throughput or SLA guarantees. Matching SKU tier to actual workload demand eliminates the premium capacity charges that high-tier SKUs impose regardless of consumption.

Why this answer

Option A is correct because non-production dev/test workloads rarely need high-tier production SKUs, so moving them to free or lower standard tiers (e.g., Azure AI Search Basic/Standard, Cognitive Services standard pricing tiers) directly cuts recurring compute and feature charges without affecting production SLAs. Option C is correct because abandoned or unused Azure AI resource accounts (e.g., orphaned Cognitive Services, Azure OpenAI, or Azure AI Search resources) still incur charges for provisioned capacity and stored data, so identifying and deleting them eliminates pure waste. Option D is correct because Azure Cost Management budgets and cost anomaly alerts provide proactive visibility and governance, letting you detect and remediate unexpected spend before it accumulates.

Option B is not appropriate because Provisioned Throughput Units (PTUs) reserve dedicated Azure OpenAI capacity at a fixed hourly rate and are cost-effective only for steady, high-volume production workloads, not idle development sandboxes. Option E is not appropriate because permanently scaling Azure AI Search partitions and replicas to maximum capacity increases cost and is only justified by sustained indexing/query load or high-availability requirements, not cost reduction.

Exam trap

AI-103 often tests whether candidates confuse 'more capacity' with 'better performance' — options that scale up resources or add PTUs sound like optimization but actually increase cost, so read for the direction of the spend impact.

Full explanation →

139

MCQeasy

When engineering prompts for Azure OpenAI models, you want to guide the model's output format and persona effectively. Which practice represents the most reliable prompt engineering technique?

A.Embedding all context and instructions exclusively inside user messages without utilizing system messages.

B.Providing clear role definition, explicit formatting instructions, and few-shot examples within the system message.

C.Using overly complex, ambiguous phrasing to encourage the model to explore diverse latent spaces.

D.Relying entirely on zero-shot inference without system messages to maximize model creativity.

AnswerB

The system message carries the highest-priority instructions, so embedding role definition, explicit output formatting rules, and few-shot examples there reliably steers persona and structure. Few-shot examples demonstrate the exact pattern, which is more dependable than instructions alone.

Why this answer

Option B is correct because the system message is the authoritative channel for defining the model's role, persona, output format constraints, and few-shot examples. Azure OpenAI models weight system messages heavily for behavioral guidance, and combining role definition, explicit formatting rules, and few-shot examples yields the most deterministic, reliable outputs. This is the recommended pattern in Microsoft's prompt engineering guidance.

Exam trap

AI-103 often tests the misconception that 'more creative' or 'zero-shot' prompting is better, when the exam expects candidates to recognize that structured system messages with role, format, and few-shot examples are the most reliable technique.

How to eliminate wrong answers

Option A is wrong because embedding all instructions in user messages removes the persistent behavioral framing that system messages provide, making the model more susceptible to prompt injection and inconsistent formatting across turns. Option C is wrong because ambiguous, complex phrasing increases variance and degrades output reliability — clarity and specificity are the goals, not ambiguity. Option D is wrong because zero-shot inference without system messages sacrifices controllability and persona consistency, maximizing creativity at the expense of the format and behavior the scenario requires.

Full explanation →

140

MCQhard

You are managing an enterprise Azure AI Search deployment. During an audit, you notice that index data is encrypted at rest using Microsoft-managed keys. Management mandates that you switch to customer-managed keys (CMK) stored in Azure Key Vault immediately without losing index data or rebuilding the service. How should you implement this?

A.Delete the search service and provision a brand new one with CMK enabled from inception.

B.Modify the DNS records to point to a new CMK-encrypted endpoint.

C.Configure encryption with customer-managed keys in the Azure portal on the existing search service instance and assign necessary Key Vault access permissions.

D.Run an Azure Automation runbook that exports indices to blob storage, deletes the service, and restores it.

AnswerC

Configuring customer-managed keys directly on the existing search service satisfies the mandate to avoid data loss or a rebuild, since CMK encryption is applied at the service level rather than requiring index recreation. Granting the search service's managed identity Key Vault access permissions enables it to wrap and unwrap the key, switching encryption from Microsoft-managed to customer-managed in place.

Why this answer

Azure AI Search supports enabling customer-managed key (CMK) encryption on an existing service without data loss or service recreation. You configure CMK in the Azure portal (or via ARM/Bicep) under the Encryption blade, then grant the search service's managed identity the necessary Key Vault permissions (get, wrapKey, unwrapKey). The service re-encrypts the index encryption key with your CMK, and no reindexing or rebuild is required.

Exam trap

AI-103 often tests whether candidates assume CMK requires rebuilding the service or migrating data, when in fact Azure AI Search supports in-place CMK enablement on existing services via the portal with Key Vault permissions.

How to eliminate wrong answers

Option A is wrong because deleting and recreating the service causes downtime, loses all indexes and data, and is completely unnecessary — CMK can be enabled in place. Option B is wrong because DNS records have nothing to do with encryption key management; changing DNS does not switch the encryption key source. Option D is wrong because exporting/deleting/restoring is a destructive, high-risk migration approach that is not required and would cause data loss and extended downtime.

Full explanation →

141

MCQmedium

You are configuring an Azure AI Search index for use with Azure OpenAI On Your Data. The index contains technical documentation with complex nested tables. Which search configuration will yield the best retrieval performance for semantic queries?

A.Configure a pure OData filter index without vector embeddings.

B.Enable semantic search alongside vector and keyword search (hybrid search).

C.Use simple query syntax with exact phrase matching only.

D.Disable all vector fields and rely exclusively on traditional Lucene scoring.

AnswerB

Hybrid search fuses BM25 keyword matching with vector similarity, while semantic ranking reranks the fused results using a language model. Together they handle exact terminology in nested tables and paraphrased semantic queries, outperforming either keyword or vector search alone.

Why this answer

Hybrid search combines keyword (BM25/Lucene) scoring with vector similarity, and semantic search adds a Microsoft-hosted reranker (based on deep learning models) on top of the merged results. For complex nested tables in technical documentation, keyword matching alone misses paraphrased queries, while pure vector search can lose exact identifiers; the semantic reranker then reorders the top ~50 results for relevance. This layered approach yields the best retrieval quality for RAG grounding in Azure OpenAI On Your Data.

Exam trap

AI-103 often tests the misconception that vector search alone is sufficient for RAG, when in fact Microsoft recommends hybrid (keyword + vector) plus semantic reranking for best retrieval quality.

How to eliminate wrong answers

Option A is wrong because OData filters only narrow results by structured field predicates and provide no relevance ranking or semantic understanding, so they cannot retrieve conceptually similar content. Option C is wrong because simple query syntax with exact phrase matching only performs literal term matching and fails on synonyms, paraphrases, or natural-language semantic queries. Option D is wrong because disabling vector fields removes the dense embedding channel entirely, leaving only traditional Lucene scoring, which cannot capture semantic similarity for complex technical content.

Full explanation →

142

MCQeasy

When evaluating an AI agent, what does the 'Coherence' metric measure?

A.The cost of the API call.

B.The logical flow and clarity of the generated text.

C.The number of sources cited.

D.The speed of the response.

AnswerB

Coherence measures how logically the generated text flows, ensuring sentences and ideas connect clearly rather than contradicting or jumping abruptly. This satisfies the evaluation requirement by assessing structural consistency across the response, distinct from groundedness (factual accuracy against sources) or fluency (grammatical correctness at sentence level).

Why this answer

The Coherence metric in AI agent evaluation measures the logical flow and clarity of the generated text. It assesses whether the response is well-structured, easy to follow, and logically consistent. This is one of the key quality metrics for generative AI outputs.

Exam trap

AI-103 often tests the confusion between coherence and other metrics like fluency or relevance, with candidates incorrectly associating coherence with source citation or response speed.

How to eliminate wrong answers

Option A is wrong because the cost of the API call is a financial metric, not a measure of text quality. Option C is wrong because the number of sources cited relates to groundedness or attribution, not coherence. Option D is wrong because the speed of the response is a performance metric (latency), not a quality metric.

Full explanation →

143

Multi-Selectmedium

You are designing a RAG solution using Azure AI Search and Azure OpenAI. Which TWO factors significantly influence the effectiveness of chunking documents during the ingestion phase? (Choose TWO)

Select 2 answers

A.Chunk size measured in tokens or characters

B.Azure OpenAI model temperature setting

C.Chunk overlap percentage between consecutive segments

D.Azure API Management timeout duration

E.Virtual network subnet IP address range

AnswersA, C

Chunk size, measured in tokens or characters, determines how much context each indexed segment carries. It directly governs retrieval granularity: oversized chunks dilute embedding relevance, undersized chunks fragment meaning, so this axis materially shapes RAG answer quality during ingestion.

Why this answer

Option A (chunk size measured in tokens or characters) is correct because the size of each chunk directly determines how much context is embedded and retrieved; chunks that are too large dilute semantic relevance and may exceed the embedding model's token limit, while chunks that are too small lack sufficient context, so tuning size in tokens or characters is a primary lever for RAG quality. Option C (chunk overlap percentage between consecutive segments) is correct because overlap preserves continuity across chunk boundaries, preventing sentences or ideas split between segments from losing meaning, and the overlap ratio materially affects retrieval recall and answer coherence. Option B is incorrect because the Azure OpenAI model temperature only controls randomness during text generation at inference time and has no effect on how documents are chunked during ingestion.

Option D is incorrect because Azure API Management timeout duration governs request/response time limits for API calls and does not influence chunking strategy. Option E is incorrect because virtual network subnet IP address range is a networking concern for connectivity and has no bearing on document segmentation.

Exam trap

AI-103 often tests the misconception that generation-side parameters like temperature or infrastructure settings like API timeouts affect ingestion-time chunking, when only chunk size and overlap are chunking-specific factors.

Full explanation →

144

Multi-Selecteasy

You need to ensure data privacy and compliance when processing user feedback containing sensitive information. Which TWO types of sensitive data can Azure AI Language PII detection identify and redact? (Choose two.)

Select 2 answers

A.Weather forecast coordinates

B.Source code syntax errors

C.Phone numbers

D.Email addresses

E.CPU utilization metrics

AnswersC, D

PII detection in Azure AI Language recognises phone numbers as a distinct entity category, returning them for redaction in processed text. This satisfies the stem's compliance requirement by letting user feedback be scrubbed of that sensitive identifier before storage or analysis.

Why this answer

Azure AI Language's PII detection feature is designed to identify and redact personal data entities, and phone numbers (option C) are a supported PII entity category, so they can be detected and masked during text processing. Email addresses (option D) are likewise a core supported PII entity in Azure AI Language, allowing them to be recognized and redacted to meet privacy and compliance requirements. Weather forecast coordinates (option A) are geographic data points, not personal identifiers recognized as PII entities by the service.

Source code syntax errors (option B) relate to code analysis, not PII detection, and CPU utilization metrics (option E) are operational telemetry rather than sensitive personal information, so neither is identified or redacted by PII detection.

Exam trap

AI-103 often tests whether candidates confuse general telemetry or infrastructure data (CPU metrics, coordinates) with true PII entity categories, causing them to pick non-personal data types as valid redaction targets.

Full explanation →

145

MCQeasy

You are developing a .NET application that extracts bounding boxes of text from scanned documents using Azure AI Vision 4.0. Which asynchronous API operation should you call first to initiate the OCR process?

A.POST /computervision/imageanalysis:analyze?api-version=2024-02-01

B.POST /vision/v3.2/ocr

C.POST /vision/v2.1/recognizeText

D.POST /customvision/v3.3/Prediction

AnswerA

The POST imageanalysis:analyze endpoint initiates the OCR request, submitting the document for processing under the specified API version. It returns the operation location for polling, making it the required first call before retrieving extracted bounding-box results asynchronously.

Why this answer

The Azure AI Vision 4.0 Image Analysis analyze endpoint is POST /computervision/imageanalysis:analyze with api-version=2024-02-01, and it supports the 'read' feature for OCR with bounding boxes. This is the correct modern endpoint for initiating asynchronous OCR in the 4.0 API surface.

Exam trap

The trap is endpoint version confusion — candidates pick the familiar /vision/v3.2/ocr path from older tutorials instead of the 4.0 /computervision/imageanalysis:analyze endpoint with the required api-version parameter.

How to eliminate wrong answers

Option B is wrong because /vision/v3.2/ocr is the legacy Computer Vision 3.2 OCR endpoint, not the 4.0 Image Analysis API. Option C is wrong because /vision/v2.1/recognizeText is an even older Read/recognizeText endpoint from the v2.1 API generation. Option D is wrong because /customvision/v3.3/Prediction belongs to Azure Custom Vision, which performs image classification/object detection on custom-trained models, not general OCR.

Full explanation →

146

Multi-Selecthard

When configuring vector embeddings and search indexes in Azure AI Search for RAG, which THREE factors directly impact vector search performance and recall accuracy? (Choose THREE)

Select 3 answers

A.Embedding model dimension size and consistency with source generation

B.Azure App Service CPU architecture (x64 vs ARM)

C.HNSW (Hierarchical Navigable Small World) algorithm parameters such as m and efSearch

D.Choice of vector similarity metric (Cosine, Dot Product, or Euclidean)

E.Configuring Azure Front Door DNS CNAME records

AnswersA, C, D

Embedding dimension determines vector size and index memory, while the query embedding must be generated by the same model as the documents. Mismatched dimensions or models break comparability, directly degrading recall accuracy and search latency.

Why this answer

Option A is correct because the embedding model's dimension size determines the vector field's dimensionality in the Azure AI Search index, and it must match the dimensions produced by the model used to generate the query embeddings; a mismatch causes search failures or degraded recall, and higher dimensions generally increase accuracy at the cost of memory and latency. Option C is correct because HNSW parameters directly govern the approximate nearest neighbor graph: 'm' controls the number of bidirectional links per node (affecting index size and recall), while 'efSearch' (efConstruction/efSearch) tunes the breadth of the search during query time, trading latency for recall. Option D is correct because the similarity metric (Cosine, Dot Product, or Euclidean) must align with how the embedding model was trained; using the wrong metric distorts distance comparisons and reduces recall accuracy.

Option B is incorrect because Azure App Service CPU architecture is unrelated to Azure AI Search vector indexing or querying, which runs on the search service's own infrastructure. Option E is incorrect because Azure Front Door DNS CNAME records only affect traffic routing and hostname resolution, not vector search performance or recall.

Exam trap

AI-103 often tests the misconception that infrastructure choices (CPU architecture, CDN/DNS) affect vector search quality — candidates must focus on embedding model properties, HNSW tuning, and similarity metric selection as the true performance and recall levers.

Full explanation →

147

MCQeasy

You are setting up cost alerts for your Azure subscription, which contains several Azure AI services resources. You want to receive an email notification when your month-to-date spending reaches 80% of your $5,000 budget. Which feature should you configure?

A.Azure Cost Management budgets and alerts

B.Microsoft Defender for Cloud security alerts

C.Azure Monitor activity log alerts

D.Azure Advisor cost recommendations

AnswerA

Azure Cost Management budgets track actual or forecast spend against a defined amount and trigger alert thresholds, such as 80% of a $5,000 monthly budget, notifying specified recipients by email. This directly satisfies the requirement to be alerted before the subscription budget is fully consumed.

Why this answer

Azure Cost Management budgets and alerts allow you to set a budget and configure alerts based on actual or forecasted spending thresholds, such as 80% of a $5,000 budget. This is the correct feature for cost monitoring and notifications.

Exam trap

AI-103 often tests the confusion between cost management and security or monitoring features, leading candidates to choose Azure Monitor or Defender for cost alerts.

How to eliminate wrong answers

Option B is wrong because Microsoft Defender for Cloud security alerts are for security threats, not cost management. Option C is wrong because Azure Monitor activity log alerts track control-plane operations, not spending. Option D is wrong because Azure Advisor cost recommendations provide suggestions to optimize costs but do not send alerts based on budget thresholds.

Full explanation →

148

MCQeasy

You are deploying a generative AI chat application on Azure OpenAI. Product owners want the model's responses to follow a strict JSON schema so the front end can render structured cards. You need the highest possible reliability that outputs conform to the schema. What should you configure?

A.Use structured outputs with a JSON schema definition on the model deployment.

B.Add few-shot examples of well-formed JSON to every request payload.

C.Increase the max_tokens value and set top_p to 0.1 to narrow sampling.

D.Set the temperature parameter to 0 and include the schema in the system message.

AnswerA

Structured outputs constrain decoding so that generated tokens conform to the supplied JSON schema, giving the highest reliability that responses parse and validate. This removes the need for prompt-only enforcement and lets the front end depend on a stable contract, which is exactly the requirement described.

Why this answer

Structured outputs enforce the JSON schema during decoding, so the model cannot emit tokens that violate the schema. This provides a deterministic contract that the front end can rely on, unlike temperature tuning, sampling adjustments, or prompt examples, all of which only increase the probability of conformance without guaranteeing it.

Exam trap

The trap here is treating temperature 0 as a guarantee of schema conformance, when deterministic sampling still permits structurally invalid output.

Full explanation →

149

MCQhard

You are managing costs for a multi-tenant Azure AI project. You need to attribute usage costs specifically to the 'Research' department. What is the most effective approach?

A.Implement Azure Policy to enforce mandatory tags on all new resources.

B.Disable diagnostic logging for the research resources to reduce billing.

C.Create a separate resource group for each department.

D.Change the pricing tier of the resource to 'Free'.

AnswerA

Azure Policy enforces mandatory tag requirements at resource creation, ensuring every resource carries a department tag such as Research. Cost Management can then filter and attribute spend by that tag, giving the required per-department cost visibility.

Why this answer

Azure Policy can enforce mandatory tags (such as Department=Research) on resource creation, ensuring consistent cost attribution across the multi-tenant project. Once tags are enforced, Azure Cost Management can group and filter costs by the Department tag, giving accurate per-department chargeback. This is the most effective and scalable approach because it prevents untagged resources from being created in the first place.

Exam trap

AI-103 often tests whether candidates confuse resource groups (a scoping/organizational construct) with tags (the actual cost-attribution mechanism), leading them to pick 'separate resource groups' when the question asks about attributing costs to a department.

How to eliminate wrong answers

Option B is wrong because disabling diagnostic logging does not reduce compute or inference billing — it only removes telemetry, harming observability while leaving costs essentially unchanged. Option C is wrong because separate resource groups alone do not provide cost attribution unless combined with tags or subscription-level separation; resource groups are a scoping mechanism, not a billing attribution mechanism, and multi-tenant resources may span groups. Option D is wrong because changing the pricing tier to Free is not a cost-attribution strategy, and Free tiers have severe quota limits that would break a production multi-tenant workload.

Full explanation →

150

MCQeasy

You are configuring a connection between Azure AI Foundry and an Azure OpenAI resource. Which credential type is recommended for secure, keyless authentication between services in Azure?

A.Account primary access key stored in application configuration

B.Hardcoded personal access tokens (PAT)

C.Anonymous public access with IP whitelisting

D.Managed Identity with Azure RBAC

AnswerD

Managed Identity issues a Microsoft Entra ID token to the Azure AI Foundry resource, and Azure RBAC grants it the required role on the Azure OpenAI resource, eliminating stored keys. This satisfies the stem's keyless authentication constraint, unlike API keys or connection strings, which must be rotated and can leak.

Why this answer

Managed Identity with Azure RBAC is the recommended keyless authentication method because Azure AI Foundry and Azure OpenAI are both Azure services that support Microsoft Entra ID token-based authentication. A system-assigned or user-assigned managed identity eliminates secrets entirely, and RBAC role assignments (such as Cognitive Services OpenAI User) scope exactly what the identity can do. This removes the risk of key leakage, rotation overhead, and credential sprawl in configuration files.

Exam trap

AI-103 often tests the misconception that API keys are acceptable for service-to-service auth in Azure, when the exam consistently favors managed identity plus RBAC as the keyless, least-privilege answer.

How to eliminate wrong answers

Option A is wrong because storing the account primary access key in application configuration embeds a long-lived shared secret that must be rotated manually and can be exfiltrated from config files or source control. Option B is wrong because hardcoded personal access tokens are still static secrets, are tied to a user identity rather than a workload, and violate the keyless authentication principle. Option C is wrong because anonymous public access with IP whitelisting provides no authentication at all — IP allowlists can be spoofed or bypassed via compromised hosts inside the allowed range and expose the endpoint to unauthenticated abuse.

Full explanation →

Microsoft Certified: Azure AI Apps and Agents Developer Associate (AI-103) (AI-103) (AI-103) — Questions 76–150