Microsoft Certified: Azure AI Apps and Agents Developer Associate (AI-103) (AI-103) AI-103 Questions 376–450 | Page 6/7

376

Multi-Selecthard

You are designing a secure RAG retrieval pipeline in Azure AI Search. Which THREE indexing or search configurations are essential when handling confidential documents? (Choose THREE)

Select 3 answers

A.Encryption at rest using Customer-Managed Keys (CMK)

B.Private Endpoint connections to isolate search traffic within a Virtual Network

C.Disabling all authentication to allow open guest queries

D.Document-level security trimming using OData security filters

E.Storing unencrypted search indexes in public Azure blob containers

AnswersA, B, D

Customer-Managed Keys encrypt Azure AI Search indexes at rest using keys you control in Azure Key Vault, rather than Microsoft-managed keys. This satisfies the confidential-documents constraint by ensuring cryptographic isolation and revocable key control over sensitive indexed content.

Why this answer

Option A is correct because Azure AI Search supports encryption at rest with Customer-Managed Keys (CMK) via Azure Key Vault, which lets you control the key lifecycle and revoke access to confidential index data, a requirement for sensitive workloads. Option B is correct because Private Endpoint connections (via Azure Private Link) give the search service a private IP inside your VNet, so indexing and query traffic never traverses the public internet, isolating confidential retrieval traffic. Option D is correct because document-level security trimming with OData filters (e.g., filtering on a groups or ACL field using search.in or eq) ensures each user only retrieves documents they are authorized to see, preventing leakage of confidential content in RAG responses.

Option C is wrong because disabling authentication would expose the index to anonymous guest queries, directly violating confidentiality. Option E is wrong because storing unencrypted indexes in public blob containers removes both encryption and network isolation, which is unacceptable for confidential documents.

Exam trap

AI-103 often tests whether candidates can identify the three pillars of secure RAG indexing — encryption (CMK), network isolation (Private Endpoint), and authorization (security trimming) — while rejecting obviously insecure options like disabling authentication or public unencrypted storage.

Full explanation →

377

MCQeasy

Your team needs to extract data from a large volume of standardized tax forms using Azure AI Document Intelligence. The forms have a fixed layout where fields are always in the exact same location. Which model type provides the fastest training and highest accuracy for this specific use case?

A.Custom template model

B.Prebuilt receipt model

C.Custom neural model

D.General document model

AnswerA

A custom template model learns fixed field positions from as few as five labelled samples, giving rapid training and high accuracy when layout is consistent. It suits standardised tax forms where fields never move, unlike custom neural or prebuilt models.

Why this answer

A custom template model is designed for documents with a consistent, fixed layout, and it trains quickly because it relies on layout/template matching rather than deep semantic learning. For standardized tax forms where fields are always in the same location, the template model achieves high accuracy with minimal training data and fast training time. This is the classic use case Microsoft documents for the template model type.

Exam trap

AI-103 often tests the trade-off between template and neural models, and candidates incorrectly assume 'neural' is always more accurate, when for fixed-layout forms the template model is faster to train and equally or more accurate.

How to eliminate wrong answers

Option B is wrong because the prebuilt receipt model is trained for sales receipts, not tax forms, and cannot be retrained on your specific form layout. Option C is wrong because the custom neural model is optimized for variable layouts and unstructured documents; it requires more training data and longer training time, and its advantage (handling layout variation) is wasted on fixed-layout forms. Option D is wrong because the general document model extracts only basic entities like key-value pairs and selection marks without customization, and it does not provide the field-level accuracy of a trained template model for your specific form.

Full explanation →

378

MCQeasy

What is the purpose of 'Batch Inference' in Azure AI Foundry?

A.To create a backup of the agent.

B.To test the agent in real-time.

C.To debug individual chat turns.

D.To run multiple prompts against a large dataset for evaluation.

AnswerD

Batch inference submits many prompts against a large dataset in a single managed job rather than one request at a time. This supports systematic evaluation of model outputs across the dataset, matching the stated purpose.

Why this answer

Batch Inference in Azure AI Foundry is designed to process a large dataset of prompts asynchronously against a deployed model, producing outputs for evaluation, scoring, or bulk generation. It is the mechanism used to run evaluation datasets through a model to compute metrics like groundedness, relevance, and coherence at scale. This makes it the correct answer for running multiple prompts against a large dataset for evaluation.

Exam trap

The trap is confusing Batch Inference with interactive testing or debugging features, causing candidates to pick 'test the agent in real-time' when the question emphasizes large datasets and evaluation.

How to eliminate wrong answers

Option A is wrong because Batch Inference does not create backups of agents — agent backup is not a feature of batch inference at all. Option B is wrong because real-time testing of an agent is done through the playground or interactive chat, not batch inference, which is asynchronous and dataset-oriented. Option C is wrong because debugging individual chat turns is a development-time activity performed in the playground or via tracing, not a batch operation over large datasets.

Full explanation →

379

MCQhard

You are monitoring the performance of a deployed Azure OpenAI model. You notice occasional HTTP 429 (Too Many Requests) errors occurring during peak traffic hours. Which monitoring and management action should you take to resolve this issue?

A.Configure Azure Monitor metric alerts for 'Token Generation Rate' and scale up your provisioned throughput deployment capacity.

B.Disable content safety filters to reduce processing latency and throughput overhead.

C.Rotate the primary API key immediately to clear the server-side connection cache.

D.Update the Network Security Group rules to increase the maximum allowed packet size for the API endpoint.

AnswerA

HTTP 429 errors signal rate-limit exhaustion, so raising provisioned throughput capacity directly removes the quota ceiling causing them. Azure Monitor alerts on Token Generation Rate surface the saturation trend early, letting you scale before peak traffic triggers throttling again.

Why this answer

HTTP 429 from Azure OpenAI indicates rate limiting — the deployment's tokens-per-minute (TPM) or requests-per-minute (RPM) quota is being exceeded. The correct response is to monitor token consumption metrics via Azure Monitor and either increase the provisioned throughput (PTU) capacity or raise the standard deployment's TPM quota, which directly addresses the throttling. Configuring metric alerts on token generation rate provides early warning before 429s spike during peak hours.

Exam trap

AI-103 often tests whether candidates recognize 429 as a quota/throttling signal rather than an auth or network issue — the trap is picking key rotation or NSG changes, which address 401/403 or connectivity problems, not rate limits.

How to eliminate wrong answers

Option B is wrong because content safety filters add a small processing step but are not the cause of 429s; disabling them reduces safety and does not increase quota. Option C is wrong because rotating API keys has no effect on rate limiting — 429 is a quota/throttling response, not an authentication failure (which would be 401/403). Option D is wrong because NSG rules govern network packet filtering, not API-level request quotas; packet size has no bearing on token-based throttling.

Full explanation →

380

Multi-Selecthard

Which THREE of the following are components of a standard Prompt Flow DAG (Directed Acyclic Graph)?

Select 3 answers

A.Inputs

B.Outputs

C.Model Training Jobs

D.Tools

E.Data Labeling Projects

AnswersA, B, D

Inputs are a core node type in a Prompt Flow DAG, defining the parameters passed into the flow at runtime. They satisfy the stem's requirement for standard DAG components by acting as entry points that feed data to downstream nodes, alongside other elements such as tools, LLM nodes and outputs.

Why this answer

In a standard Prompt Flow DAG, Inputs (A) are the declared entry points that supply values (for example, a question string or a dataset column) to the flow, making them a required component. Outputs (B) are also required because they define the flow's returned values (such as the model's answer or a score) that downstream consumers or evaluations read. Tools (D) are the executable nodes of the DAG — LLM, Python, Prompt, and built-in tools — and the edges connecting them define the directed acyclic execution order.

Model Training Jobs (C) are not part of a Prompt Flow DAG; training is handled by separate Azure Machine Learning job constructs, not flow nodes. Data Labeling Projects (E) are likewise a distinct Azure Machine Learning feature for annotating data and are not components of a Prompt Flow DAG.

Exam trap

AI-103 often tests whether candidates confuse Prompt Flow DAG components with adjacent Azure ML features like training jobs and labeling projects, which sound pipeline-related but are not part of the flow graph.

Full explanation →

381

Multi-Selectmedium

When planning an Azure AI solution deployment strategy, what are THREE key considerations for model versioning and lifecycle management? (Choose three.)

Select 3 answers

A.Implementing automated regression testing in CI/CD pipelines when upgrading model versions.

B.Reviewing the deprecation and retirement schedule published by Microsoft for specific model versions.

C.Deleting older model versions immediately upon release of a new version to save cloud storage costs.

D.Hardcoding the latest model version string into production code without version specifiers.

E.Establishing a rollback plan in case a newer model version exhibits unexpected behavior.

AnswersA, B, E

Automated regression testing catches accuracy or behavioural degradation introduced by a new model version before promotion, directly satisfying the lifecycle requirement to validate each upgrade against a known baseline rather than trusting vendor release notes.

Why this answer

Option A is correct because automated regression testing in CI/CD pipelines validates that a new model version meets quality, accuracy, and safety baselines before promotion, catching behavioral drift or breaking changes introduced by the upgrade. Option B is correct because Microsoft publishes deprecation and retirement schedules for specific model versions in Azure OpenAI and Azure AI services, and solutions must track these dates to migrate before a version is retired and API calls begin failing. Option E is correct because model upgrades can introduce unexpected behavior, so a documented rollback plan—such as pinning the previous version or using deployment slots to revert traffic—ensures continuity if the new version underperforms.

Option C is not appropriate because deleting older versions immediately removes the ability to roll back and may break pinned deployments still serving production traffic. Option D is not appropriate because hardcoding a 'latest' model string without explicit version specifiers makes behavior non-deterministic and can silently break when Microsoft updates or retires that alias.

Exam trap

AI-103 often tests lifecycle best practices by offering cost-saving or 'always latest' shortcuts that sound efficient but violate change-management and rollback principles.

Full explanation →

382

Multi-Selecthard

You are troubleshooting a low confidence score on specific fields extracted by your Azure AI Document Intelligence custom template model. Which TWO actions can help improve the extraction confidence for those fields? (Choose TWO)

Select 2 answers

A.Increase the confidence threshold parameter in the analyzeDocument request body to 0.99.

B.Convert all training PDFs to grayscale and reduce their resolution to 72 DPI to speed up processing.

C.Ensure all training documents are of high visual quality and free from blur or heavy skew.

D.Add more diverse training documents that represent variations in the document layout where the low-confidence fields appear.

E.Switch the model type from custom template to a prebuilt business card model.

AnswersC, D

Low-quality scans with blur or skew degrade the OCR stage, so extracted text feeding the template model is corrupted, lowering field confidence. Clean, sharp training documents raise OCR accuracy, which directly lifts confidence on the affected fields.

Why this answer

Option C is correct because custom template models learn field positions and visual patterns from the labeled samples, so blurry, skewed, or low-quality training documents degrade the model's ability to reliably locate and read those fields, producing low confidence scores; supplying sharp, straight, high-quality scans directly improves feature extraction and OCR accuracy. Option D is correct because low confidence on specific fields usually indicates the model has not seen enough layout variation for those fields, so adding diverse training documents that capture the different layouts, fonts, and positions where those fields appear lets the model generalize and raise its confidence. Option A is wrong because the confidence threshold in the analyzeDocument request only filters which results are returned or flagged; it does not change the model's actual extraction confidence.

Option B is wrong because converting to grayscale and downsampling to 72 DPI reduces image fidelity and typically lowers OCR and extraction quality rather than improving confidence. Option E is wrong because a prebuilt business card model is a fixed, general-purpose model that cannot be trained on your specific fields or layouts, so it cannot improve confidence for your custom template fields.

Exam trap

AI-103 often tests common misconceptions about improving model performance; candidates might think adjusting API parameters or switching to prebuilt models helps, rather than focusing on training data quality and diversity.

Full explanation →

383

MCQeasy

You are deploying an LLM in Azure AI Foundry and need to monitor for toxicity and jailbreak attempts in real-time. Which feature should you enable?

A.Azure AI Content Safety

B.Azure Policy

C.Microsoft Purview Compliance Manager

D.Azure Monitor Log Analytics

AnswerA

Azure AI Content Safety provides real-time classification of prompts and completions, including toxicity severity scores and jailbreak or prompt-injection detection. Enabling it on the Azure AI Foundry deployment intercepts harmful input and output at inference time, meeting the stem's real-time monitoring requirement.

Why this answer

Azure AI Content Safety is the service designed to detect harmful content, including toxicity and jailbreak attempts, in real time for LLM applications. It provides classifiers for hate, violence, sexual, and self-harm content, plus Prompt Shields for jailbreak and indirect prompt injection detection. Enabling it allows pre- and post-processing moderation of prompts and completions.

Exam trap

AI-103 often tests the distinction between AI-specific safety services and general Azure security/governance tools; candidates may pick Azure Monitor or Purview because they sound like monitoring/compliance solutions, missing that Content Safety is the only service that performs real-time toxicity and jailbreak detection.

How to eliminate wrong answers

Option B is wrong because Azure Policy is a governance service for enforcing organizational rules on Azure resources (e.g., allowed regions, SKUs), not for analyzing content or jailbreak attempts. Option C is wrong because Microsoft Purview Compliance Manager assesses compliance posture against regulations and standards; it does not perform real-time content moderation. Option D is wrong because Azure Monitor Log Analytics collects and queries telemetry and logs; it can store Content Safety logs but does not itself detect toxicity or jailbreaks.

Full explanation →

384

MCQhard

You are deploying an Azure AI Document Intelligence custom extraction model for a regulated workload. Documents are uploaded to Azure Blob Storage in a private container, and no public network access is allowed. Your application runs in an Azure virtual network. You must ensure the Document Intelligence resource can access the documents for training and analysis without exposing them publicly. Which configuration should you implement?

A.Copy the documents into the Document Intelligence resource's own storage and reference them by internal path

B.Generate a long-lived SAS token for the container and embed it in the training request

C.Configure the Document Intelligence resource with a private endpoint and grant it access to the storage account through a managed identity with the Storage Blob Data Reader role

D.Enable public network access on the storage account but restrict it by IP firewall to the Document Intelligence outbound IP range

AnswerC

A private endpoint brings the Document Intelligence resource into the virtual network, and a managed identity with Storage Blob Data Reader lets the service read blobs without public exposure or shared keys. This satisfies the no-public-access requirement while enabling both training and analysis against private containers.

Why this answer

For private, no-public-access scenarios, the supported pattern is a private endpoint for the Document Intelligence resource plus a managed identity granted Storage Blob Data Reader on the storage account. This keeps traffic on the virtual network and avoids shared keys or SAS tokens. IP firewalls and SAS tokens do not satisfy the no-public-access constraint.

Exam trap

The trap here is treating a SAS token as sufficient for private access, when it leaves the storage account publicly reachable and adds credential risk.

Full explanation →

385

MCQhard

You have an Azure AI endpoint exposed to the internet. You want to ensure traffic is inspected for malicious patterns. What is the recommended service?

A.Network Security Group (NSG) rules.

B.Azure Web Application Firewall (WAF).

C.Azure Bastion.

D.Azure AI Content Safety.

AnswerB

Azure Web Application Firewall inspects inbound HTTP/S traffic at the application layer, detecting and blocking injection, cross-site scripting and other malicious patterns. Placing it in front of the internet-exposed Azure AI endpoint satisfies the requirement for traffic inspection.

Why this answer

Azure Web Application Firewall (WAF) inspects HTTP/HTTPS traffic at Layer 7 and applies OWASP rule sets to block SQL injection, cross-site scripting, and other malicious patterns before they reach the AI endpoint. It is the recommended service for inspecting internet-facing web traffic for attacks. NSGs operate at Layers 3/4 and cannot see payload content, so they cannot detect application-layer attacks.

Exam trap

AI-103 often tests the OSI-layer distinction — candidates pick NSG because it 'filters traffic', missing that only a Layer 7 WAF can inspect payloads for malicious patterns.

How to eliminate wrong answers

Option A is wrong because Network Security Group rules filter by IP, port, and protocol at Layers 3/4 — they have no visibility into HTTP payloads and cannot detect injection or XSS patterns. Option C is wrong because Azure Bastion provides secure RDP/SSH access to VMs over TLS; it is a jump-host service, not a traffic inspection or WAF service. Option D is wrong because Azure AI Content Safety moderates AI-generated or user-generated content for harm categories (violence, hate, etc.) — it is not a network-layer web attack firewall.

Full explanation →

386

MCQmedium

When setting up an Azure AI service, you choose the 'Standard' tier. What is the primary benefit of this compared to the 'Free' tier?

A.Lower latency for all requests.

B.Access to experimental AI models.

C.Higher rate limits and production-level SLAs.

D.Automatic data encryption.

AnswerC

The Standard tier removes the Free tier's per-minute and monthly transaction caps and provides financially backed availability SLAs, which production workloads require. Free tier suits evaluation only, lacking both throughput guarantees and contractual uptime commitments.

Why this answer

The Standard tier of Azure AI services provides higher rate limits (e.g., transactions per second) and production-level SLAs (e.g., 99.9% availability) compared to the Free tier, which has strict rate limits and no SLA. The Standard tier is intended for production workloads, while Free is for evaluation and small-scale testing.

Exam trap

AI-103 often tests the misconception that the Standard tier provides lower latency or exclusive features, when the primary differences are rate limits and SLAs.

How to eliminate wrong answers

Option A is wrong because lower latency is not a guaranteed benefit of the Standard tier; latency depends on factors like region, network, and service load, and both tiers may have similar latency under low load. Option B is wrong because access to experimental AI models is not tied to the tier; experimental models may be available in preview regardless of tier. Option D is wrong because automatic data encryption is enabled by default for all tiers, including Free.

Full explanation →

387

MCQhard

You are orchestrating a multi-agent system in Microsoft Foundry where Agent A generates code and Agent B reviews it. Agent B frequently gets stuck in a loop trying to optimize minor syntax formatting. How should you design the agentic workflow to prevent this infinite refinement loop?

A.Disable function calling for Agent B so it can only output text responses.

B.Configure a maximum turn counter and state transition condition in the orchestrator workflow to terminate the loop after a set threshold.

C.Increase the model's presence penalty so Agent B is forced to generate novel review comments.

D.Use a smaller model for Agent B to reduce the quality of its reviews.

AnswerB

A maximum turn counter with an explicit state transition condition caps Agent B's refinement cycles, forcing the orchestrator to exit the review loop once the threshold is reached. This bounds iteration without disabling review, preventing endless syntax optimisation.

Why this answer

Configuring a maximum turn counter and a state transition condition in the orchestrator terminates the loop after a defined threshold, preventing infinite refinement. This is the standard pattern for bounded agentic workflows, ensuring the system exits gracefully when convergence is not reached. It directly addresses the runaway loop without degrading agent capability.

Exam trap

AI-103 often tests the misconception that model parameters (presence penalty, model size) can control agentic loops, when the correct answer is orchestration-level controls like max turn counters and state transition conditions.

How to eliminate wrong answers

Option A is wrong because disabling function calling cripples Agent B's ability to perform reviews that require tools, degrading the workflow rather than fixing the loop. Option C is wrong because presence penalty affects token novelty in generation, not the orchestration logic that causes repeated review cycles — it does not terminate loops. Option D is wrong because using a smaller model reduces review quality and still does not guarantee loop termination; it is a degradation, not a control.

Full explanation →

388

MCQmedium

You are integrating Azure AI Vision Read API into a high-throughput workflow. You receive an HTTP 202 Accepted response. What action should your application take next?

A.Poll the URL specified in the Operation-Location response header until the status is succeeded.

B.Switch to synchronous Read API endpoint.

C.Parse the response body directly for the extracted text results.

D.Immediately retry the POST request with an increased timeout value.

AnswerA

The Read API processes documents asynchronously, so a 202 Accepted means the job is queued rather than finished. The Operation-Location header supplies the polling endpoint; your application must repeatedly query it until the status reports succeeded, satisfying the high-throughput requirement without blocking on a synchronous response.

Why this answer

An HTTP 202 Accepted response from the Azure AI Vision Read API indicates that the request has been accepted for asynchronous processing, and the response includes an Operation-Location header containing a URL to poll for the result. The application must poll that URL (typically with a GET request) until the status field returns 'succeeded', at which point the extracted text is available in the response body. This is the documented asynchronous pattern for the Read API.

Exam trap

AI-103 often tests the assumption that a 202 response contains the final result, causing candidates to parse the body directly instead of polling the Operation-Location header.

How to eliminate wrong answers

Option B is wrong because the Read API does not offer a synchronous endpoint for large documents; switching to a synchronous endpoint is not a valid workaround and would not handle the asynchronous nature of the operation. Option C is wrong because the 202 response body does not contain the extracted text — it only acknowledges the request; the results are available only after polling the Operation-Location URL. Option D is wrong because retrying the POST request immediately would create a duplicate operation and waste resources; the correct action is to poll, not retry.

Full explanation →

389

Multi-Selecthard

You are configuring custom named entity recognition in Language Studio. Which TWO types of entity components can you configure when defining custom entities? (Choose two.)

Select 2 answers

A.Quantum computation component

B.Regex component

C.Neural speech synthesis component

D.Learned component

E.Optical character recognition component

AnswersB, D

A regex component matches entity text against a defined regular expression pattern, satisfying the need to extract structured, predictable formats such as account numbers or dates without labelled examples. Combined with a learned component, it lets you define custom entities using pattern-based rules alongside machine learning within Language Studio.

Why this answer

In Azure AI Language custom NER, when you define an entity you can add a Regex component (option B), which matches entity instances using a regular expression pattern, and a Learned component (option D), which uses labeled training examples so the model learns to recognize the entity from context. These two component types are the ones supported in the entity definition workflow in Language Studio, and they can be combined within a single entity. The other options are unrelated to custom NER entity components: quantum computation (A), neural speech synthesis (C), and optical character recognition (E) are not configurable entity component types in Language Studio.

Exam trap

AI-103 often tests the specific component types available in custom NER, and candidates may confuse them with other AI capabilities like OCR or speech, so remembering 'learned' and 'regex' is key.

Full explanation →

390

MCQmedium

You are evaluating an agentic workflow in Azure AI Foundry that performs multi-step reasoning. The agent occasionally enters an infinite loop, repeatedly calling the same tool with identical arguments. Which architectural pattern or configuration should you apply to prevent this behavior?

A.Increase the vector store chunk overlap size

B.Configure a maximum iteration step limit in the agent orchestration loop

C.Enable Azure API Management caching for the tool endpoint

D.Set the model frequency penalty to 2.0

AnswerB

A maximum iteration step limit caps the orchestration loop, terminating execution once the count is reached. This directly satisfies the stem's constraint of preventing infinite loops from repeated identical tool calls, since the agent cannot exceed the configured ceiling regardless of its reasoning path.

Why this answer

Agentic workflows in Azure AI Foundry (and similar frameworks like Semantic Kernel or AutoGen) require a bounded orchestration loop; configuring a maximum iteration or step limit causes the orchestrator to terminate or escalate when the agent exceeds the threshold, preventing infinite tool-calling loops. This is the standard architectural safeguard for multi-step reasoning agents. It directly addresses the symptom of repeated identical tool calls.

Exam trap

AI-103 often tests whether candidates confuse model-level parameters (temperature, frequency penalty) with orchestration-level controls, tricking them into picking a sampling parameter to fix a control-flow problem.

How to eliminate wrong answers

Option A is wrong because vector store chunk overlap affects retrieval quality and context continuity, not agent loop control — it has no bearing on repeated tool invocations. Option C is wrong because API Management caching would return cached responses for identical requests, which could mask the loop but does not stop the agent from continuing to call the tool and would produce stale results. Option D is wrong because frequency penalty (max 2.0) only discourages token repetition in generated text; it does not prevent the orchestration layer from re-invoking tools with identical arguments.

Full explanation →

391

MCQhard

You are optimizing performance for a real-time video analytics pipeline that sends video frames to Azure AI Vision. Due to network bandwidth constraints, you need to reduce payload sizes while maintaining high OCR accuracy. What is the recommended approach?

A.Encode frames as optimized JPEG images with appropriate resolution scaling, ensuring text remains legible.

B.Convert frames to uncompressed 32-bit BMP format.

C.Extract audio tracks instead of video frames.

D.Send raw YUV byte arrays directly without image containerization.

AnswerA

JPEG compression with resolution scaling directly reduces frame payload size, satisfying the bandwidth constraint. Keeping text legible preserves OCR accuracy, since Azure AI Vision's Read model requires sufficient pixel density; excessive downscaling would degrade recognition despite smaller payloads.

Why this answer

Optimized JPEG encoding with resolution scaling reduces payload size while preserving text legibility, which is the key requirement for OCR accuracy. JPEG's lossy compression is well-suited for photographic frames, and scaling to the minimum resolution that keeps text readable balances bandwidth and accuracy. This is the recommended approach for bandwidth-constrained real-time OCR pipelines.

Exam trap

AI-103 often tests whether candidates understand that OCR accuracy depends on text legibility, so the trap is choosing the smallest possible payload (raw bytes, audio) or the largest (BMP) instead of the balanced JPEG-with-scaling answer.

How to eliminate wrong answers

Option B is wrong because uncompressed 32-bit BMP is far larger than JPEG and would worsen bandwidth constraints, not relieve them. Option C is wrong because extracting audio tracks discards the visual information OCR needs entirely. Option D is wrong because raw YUV byte arrays lack image containerization and metadata, are typically larger than compressed formats, and are not directly accepted by Azure AI Vision's image input requirements.

Full explanation →

392

MCQmedium

You are implementing a multi-turn conversational agent in Azure AI Foundry. As the conversation grows longer, API latency and costs increase significantly. Which strategy should you implement to manage the conversation history effectively?

A.Increase max_tokens to accommodate infinite history growth.

B.Implement a sliding window history buffer combined with periodic summarization of older dialogue turns.

C.Disable the chat history array entirely and treat every turn as a standalone prompt.

D.Truncate the entire conversation history after every single user turn.

AnswerB

A sliding window caps the tokens sent per turn, directly cutting latency and cost, while periodic summarisation preserves older context that would otherwise be discarded. This satisfies the stem's constraint of rising API latency and expense as the conversation lengthens.

Why this answer

For long multi-turn conversations, the standard pattern is a sliding window that keeps only the most recent N turns plus a running summary of older dialogue. This bounds the token count sent to the model, controlling latency and cost, while preserving essential context through summarization. It is the recommended approach in Azure AI Foundry for managing conversation history.

Exam trap

AI-103 often tests whether candidates understand that context windows are finite and that history must be actively managed — the trap is choosing to increase max_tokens or disable history instead of implementing a sliding window with summarization.

How to eliminate wrong answers

Option A is wrong because increasing max_tokens does not solve unbounded history growth — it raises the output ceiling, not the input context, and the model still has a hard context window limit. Option C is wrong because disabling chat history entirely destroys conversational context, making the agent unable to resolve references or maintain coherence. Option D is wrong because truncating after every single turn is effectively the same as having no history and breaks multi-turn reasoning.

Full explanation →

393

MCQeasy

Which Azure AI service should you use if you need to train a model to recognize your company's proprietary product packaging using custom-labeled images with minimal machine learning expertise?

A.Azure AI Vision Prebuilt API

B.Azure AI Custom Vision

C.Azure Machine Learning Automated ML

D.Azure AI Document Intelligence

AnswerB

Custom Vision trains image classification and object detection models from your own labelled images through a guided interface, requiring no ML expertise. This suits recognising proprietary packaging, unlike prebuilt vision services that cannot learn company-specific product categories.

Why this answer

Azure AI Custom Vision is purpose-built for training custom image classification and object detection models using your own labeled images, and it abstracts away the ML pipeline (data splitting, hyperparameter tuning, model selection) behind a simple portal and SDK. It requires no ML expertise — you upload images, tag them, and click Train. This makes it the correct choice for recognizing proprietary product packaging with custom labels.

Exam trap

AI-103 often tests the boundary between 'custom' and 'prebuilt' AI services — candidates pick the Prebuilt API assuming it can be extended with their own labels, when only Custom Vision supports user-supplied training data.

How to eliminate wrong answers

Option A is wrong because the Prebuilt API only offers fixed, general-purpose models (e.g., generic object, brand, or category detection) and cannot be trained on your company's proprietary packaging labels. Option C is wrong because Azure Machine Learning Automated ML, while capable of image tasks, requires configuring workspaces, compute clusters, datasets, and experiment settings — it is not the minimal-expertise path and is overkill for a simple custom vision task. Option D is wrong because Azure AI Document Intelligence is designed for extracting text, key-value pairs, and tables from documents (invoices, forms, receipts), not for recognizing physical product packaging in images.

Full explanation →

394

Multi-Selecthard

You are designing an enterprise RAG solution using Azure AI Search and Azure OpenAI. You need to ensure data security and compliance across different user permission tiers. Which TWO actions should you implement? (Choose two.)

Select 2 answers

A.Ensure all data transfers between Azure AI Search and Azure OpenAI utilize TLS 1.2 or higher.

B.Configure security filters (document-level access control) in Azure AI Search matching user group memberships.

C.Store unencrypted user authentication tokens directly inside the embedding vector metadata.

D.Grant every user application admin privileges to bypass retrieval latency checks.

E.Disable managed identities and store plaintext connection strings in the application source code.

AnswersA, B

TLS 1.2 or higher encrypts data in transit between Azure AI Search and Azure OpenAI, preventing interception or downgrade attacks. This satisfies the stem's compliance requirement by protecting cross-service traffic, though it does not by itself enforce per-user document permissions.

Why this answer

Option A is correct because Azure AI Search and Azure OpenAI communicate over HTTPS endpoints, and enforcing TLS 1.2 or higher protects data in transit against interception or downgrade attacks, which is a baseline requirement for enterprise security and compliance. Option B is correct because Azure AI Search supports document-level access control via security filters (for example, using the user's group IDs in a filterable field or the built-in ACL/rbacScope fields), ensuring each permission tier only retrieves documents its users are authorized to see, which is essential in a multi-tenant RAG solution. Option C is wrong because storing unencrypted authentication tokens in embedding metadata exposes credentials to anyone querying the index and violates token-handling best practices.

Option D is wrong because granting admin privileges to all users removes least-privilege enforcement and does not address retrieval latency in a legitimate way. Option E is wrong because disabling managed identities and embedding plaintext connection strings in source code leaks secrets and discards Azure's keyless, identity-based authentication model.

Exam trap

AI-103 often tests whether candidates conflate transport security (TLS) with authorization (document-level filters), or mistakenly believe that embedding metadata is a safe place for secrets — both are common compliance misconceptions.

Full explanation →

395

MCQmedium

You are building a custom text classification solution using Azure AI Language in Language Studio. You have a dataset where each document can belong to multiple categories simultaneously (e.g., a news article can be tagged as both 'Technology' and 'Finance'). What classification type should you configure?

A.Custom named entity recognition

B.Multi-label classification

C.Single-label classification

D.Extractive summarization

AnswerB

Multi-label classification assigns every applicable category to a document independently, rather than forcing a single mutually exclusive choice. This directly satisfies the stem's constraint that a news article can be tagged as both 'Technology' and 'Finance' simultaneously, which single-label classification cannot represent.

Why this answer

Multi-label classification is designed for scenarios where a single document can be assigned more than one category at the same time. In Azure AI Language, multi-label projects let you apply multiple labels to each document during training and return all applicable labels at inference. This matches the requirement that a news article can be tagged as both 'Technology' and 'Finance'.

Exam trap

AI-103 often tests the distinction between single-label and multi-label classification, and candidates frequently default to single-label because it is the more familiar pattern, missing the keyword 'multiple categories simultaneously' that signals multi-label.

How to eliminate wrong answers

Option A is wrong because custom named entity recognition extracts specific entities (people, places, organizations) from text rather than assigning document-level categories. Option C is wrong because single-label classification forces each document into exactly one mutually exclusive category, which cannot represent simultaneous Technology and Finance tags. Option D is wrong because extractive summarization produces key sentences from a document and performs no classification at all.

Full explanation →

396

MCQmedium

You are testing a Custom Vision model via the REST API. You receive an HTTP 400 Bad Request error with the message 'InvalidProjectId'. What is the most likely cause?

A.The image file format is invalid.

B.The project ID GUID in the request URL is incorrect or does not exist in this Custom Vision account.

C.The subscription has reached its monthly quota limit.

D.The prediction key header was omitted.

AnswerB

The InvalidProjectId message indicates the GUID supplied in the request URL does not match any project in the Custom Vision account, either mistyped or belonging to another resource. Correcting the project ID to an existing GUID resolves the HTTP 400 Bad Request.

Why this answer

The Custom Vision REST API returns HTTP 400 with 'InvalidProjectId' when the project ID GUID supplied in the request URL path (e.g., /customvision/v3.0/training/projects/{projectId}/...) cannot be resolved to a project in the account tied to the submitted training or prediction key. This is a client-side request error, meaning the GUID is malformed, belongs to a different Custom Vision resource/account, or the project was deleted. Because the error is specifically named 'InvalidProjectId', the failure is at the identifier-validation stage, before any image payload is parsed.

Exam trap

AI-103 often tests whether candidates can map specific HTTP status codes and error strings to their root causes — the trap is assuming any 400 means a bad payload (image) rather than a bad identifier in the URL.

How to eliminate wrong answers

Option A is wrong because an invalid image format produces a different error (typically 'BadImageFormat' or 'InvalidImage') and is only evaluated after the project ID is validated. Option C is wrong because exceeding the monthly quota returns HTTP 429 Too Many Requests or a quota-specific error, not a 400 with 'InvalidProjectId'. Option D is wrong because a missing or invalid prediction/training key returns HTTP 401 Unauthorized, not 400 Bad Request.

Full explanation →

397

MCQeasy

Which HTTP status code does the Azure AI Vision Read API return upon successfully initiating an asynchronous OCR analysis job?

A.204 No Content

B.200 OK

C.202 Accepted

D.201 Created

AnswerC

The Read API processes documents asynchronously, so it returns 202 Accepted to confirm the job was queued without blocking. The response includes an Operation-Location header the client polls for results, distinguishing it from synchronous 200 responses.

Why this answer

The Read API is asynchronous: when you POST an image or URL to the Read endpoint, it returns 202 Accepted with an Operation-Location header. You then poll that URL until the job completes and returns 200 OK with results.

Exam trap

AI-103 often tests the distinction between the initial 202 Accepted response and the final 200 OK — candidates may incorrectly choose 200 OK for job initiation.

How to eliminate wrong answers

Option A is wrong because 204 No Content indicates success with no body, which is not how the Read API initiates jobs. Option B is wrong because 200 OK is returned when polling the operation and the job is complete, not upon initiation. Option D is wrong because 201 Created is for resource creation, not asynchronous job initiation.

Full explanation →

398

Multi-Selecthard

You are troubleshooting a multi-agent workflow in Azure AI Foundry where an agent fails to invoke an external API tool correctly. Which THREE debugging steps should you perform? (Choose THREE)

Select 3 answers

A.Test network connectivity and authentication credentials between the agent service and the external API endpoint

B.Restart the client web browser session

C.Inspect the model's generated tool call arguments against the expected JSON Schema definition

D.Review the tool definition schema provided to the model for accurate descriptions and parameter types

E.Increase the Azure DNS TTL value for domain resolution

AnswersA, C, D

A tool invocation crossing a network boundary can fail on connectivity or credential rejection before the model's arguments matter. Verifying reachability and authentication between the agent service and the API endpoint isolates infrastructure faults from schema or prompt faults.

Why this answer

Option A is correct because when an agent fails to invoke an external API tool, you must verify that the agent service can actually reach the endpoint and that its authentication credentials (API keys, OAuth tokens, managed identity, etc.) are valid, since connectivity or auth failures are common root causes of tool invocation errors. Option C is correct because inspecting the model's generated tool call arguments against the expected JSON Schema reveals whether the model is producing malformed, missing, or incorrectly typed parameters that the external API would reject. Option D is correct because the tool definition schema given to the model—including accurate descriptions, parameter names, types, and required fields—directly determines whether the model can generate a valid call; vague or incorrect schemas cause the model to hallucinate or omit arguments.

Option B does not belong because restarting the client browser session only affects the local UI and has no bearing on server-side agent-to-API tool invocation logic. Option E does not belong because increasing Azure DNS TTL affects caching duration for domain resolution and would not fix an agent's failure to correctly construct or authenticate an external API tool call.

Exam trap

AI-103 often tests whether candidates can separate client-side symptoms (browser, DNS TTL) from server-side agent/tool integration issues — the trap is picking a superficially 'networky' answer like DNS TTL instead of the schema and auth checks that actually diagnose tool-call failures.

Full explanation →

399

MCQmedium

Your organization wants to ensure that all Azure AI resources created in any subscription across the tenant adhere to a standardized naming convention (e.g., prefixing with 'ai-'). Which Azure governance tool should you use?

A.Microsoft Defender for Cloud

B.Azure Cost Management budgets

C.Azure Advisor

D.Azure Policy with custom naming convention rules

AnswerD

Azure Policy evaluates resource properties at creation and can deny or audit non-compliant names, applying across every subscription in the tenant through management group scope. Custom naming rules enforce the 'ai-' prefix consistently, which a single subscription-level convention cannot guarantee.

Why this answer

Azure Policy is the governance service designed to enforce organizational standards such as naming conventions across subscriptions and resources. A custom policy definition using a naming pattern rule (e.g., enforcing an 'ai-' prefix) can be assigned at the management group or subscription scope to audit or deny non-compliant resources tenant-wide.

Exam trap

AI-103 often tests the difference between advisory tools (Advisor, Defender for Cloud) and enforcement tools (Azure Policy) — candidates pick a security or cost tool when the requirement is governance enforcement.

How to eliminate wrong answers

Option A is wrong because Microsoft Defender for Cloud provides security posture management and threat protection, not naming-convention enforcement. Option B is wrong because Azure Cost Management budgets track and alert on spending, not resource naming. Option C is wrong because Azure Advisor offers best-practice recommendations but does not enforce or deny resource creation based on naming.

Full explanation →

400

Multi-Selectmedium

When configuring Azure OpenAI Custom Model Evaluation in Azure AI Foundry, which TWO metrics can be computed using LLM-as-a-judge evaluators? (Choose two.)

Select 2 answers

A.CPU utilization percentage

B.DNS propagation delay

C.Disk input/output operations per second (IOPS)

D.Coherence

E.Relevance

AnswersD, E

Coherence suits this scenario because LLM-as-a-judge evaluators assess subjective, generative qualities that deterministic metrics cannot capture. A judge model scores how logically the response flows and connects ideas, satisfying the stem's requirement for judge-computed metrics rather than similarity-based measures such as BLEU or ROUGE.

Why this answer

Options D and E are correct because Azure AI Foundry's LLM-as-a-judge evaluators are designed to assess the quality of generative AI outputs using a language model as the grader, and Coherence (D) and Relevance (E) are two of the standard quality metrics they compute — Coherence measures how logically well-structured and fluent the response is, while Relevance measures how directly the response addresses the user's query or context. These metrics are part of the built-in evaluation metric set (alongside others such as Fluency, Groundedness, and Similarity) that can be run against a model deployment during custom model evaluation. Options A, B, and C are incorrect because CPU utilization percentage, DNS propagation delay, and disk IOPS are infrastructure and networking performance counters, not language-quality judgments, and they cannot be produced by an LLM-as-a-judge evaluator.

Exam trap

AI-103 often tests the confusion between system performance metrics (CPU, DNS, disk) and AI quality metrics (Coherence, Relevance), tricking candidates into selecting infrastructure metrics for model evaluation.

Full explanation →

401

Multi-Selecthard

You are building an autonomous agentic workflow using Microsoft Foundry and Semantic Kernel. You need to ensure the agent operates safely and adheres to human oversight requirements. Which THREE design patterns should you apply? (Choose three.)

Select 3 answers

A.Grant the agent root-level administrative access to all cloud infrastructure services.

B.Implement human-in-the-loop approval gates before the agent executes high-impact tool actions such as database updates or financial transactions.

C.Configure strict idempotency keys for external tool calls to prevent duplicate actions if retried.

D.Implement circuit breakers to halt agent execution if error rates or token consumption spike abnormally.

E.Allow the agent unlimited recursive tool execution without loop detection.

AnswersB, C, D

Human-in-the-loop approval gates directly satisfy the human oversight requirement by pausing execution before high-impact tool actions, such as database updates or financial transactions, so a person authorises the operation rather than the agent acting autonomously.

Why this answer

Option B is correct because human-in-the-loop approval gates are the canonical oversight control for high-impact, irreversible actions such as database updates or financial transactions, ensuring a human authorizes the action before the agent's tool call is dispatched. Option C is correct because idempotency keys make external tool calls safely retryable: if the agent or orchestrator retries after a timeout or transient failure, the downstream service deduplicates the request so a payment or write is not executed twice. Option D is correct because circuit breakers are a resilience pattern that trips when error rates or token consumption exceed thresholds, halting or degrading agent execution to prevent runaway loops, cost blowups, and cascading failures.

Option A is not appropriate because granting root-level administrative access violates least privilege and dramatically expands the blast radius of any agent error or prompt-injection attack. Option E is not appropriate because unlimited recursive tool execution without loop detection invites infinite loops, resource exhaustion, and uncontrolled cost, directly contradicting safe agent design.

Exam trap

AI-103 often tests the misconception that giving an agent broad permissions or unlimited recursion makes it more capable, when in fact safe agentic design requires least privilege, approval gates, idempotency, and circuit breakers.

Full explanation →

402

MCQmedium

Which Azure AI Foundry feature helps identify drift in model performance after deployment?

A.Model Monitoring

B.Data Drift Analysis via Compute

C.Prompt Flow Tracing

D.Model Catalog

AnswerA

Model Monitoring in Azure AI Foundry continuously tracks deployed model performance against baseline metrics, detecting data and prediction drift over time. It satisfies the stem's requirement to identify post-deployment drift by evaluating live inference data and raising alerts when performance degrades, enabling timely retraining or remediation.

Why this answer

Model Monitoring in Azure AI Foundry continuously evaluates deployed models against production data, tracking data drift, prediction drift, and performance metrics like accuracy and latency. It is the built-in feature designed specifically to detect post-deployment degradation. The other options either describe manual or unrelated capabilities that do not provide ongoing drift detection for deployed models.

Exam trap

The trap here is confusing 'drift analysis' as a generic concept with the specific managed Foundry feature — candidates may pick the manually configured compute option because it sounds more technical.

How to eliminate wrong answers

Option B (Data Drift Analysis via Compute) is wrong because it describes a manual, compute-driven analysis pattern rather than the managed Foundry feature that automatically monitors deployed models. Option C (Prompt Flow Tracing) is wrong because tracing captures prompt-level execution telemetry for debugging and evaluation, not statistical drift in model performance. Option D (Model Catalog) is wrong because the catalog is a discovery and deployment repository for models, not a monitoring or drift-detection service.

Full explanation →

403

MCQhard

Your security team discovers that an application using Azure AI Services is intermittently failing with authentication errors due to expired client secrets. You need to refactor the application to use Microsoft Entra ID workﬂows with managed identities. Which authentication library should your application code use to acquire tokens securely without embedding secrets?

A.Azure Storage Data Movement Library

B.System.Net.Http WebRequestHandler

C.Azure SDK for .NET Management Client

D.Microsoft Authentication Library (MSAL)

AnswerD

MSAL acquires tokens from Microsoft Entra ID using managed identities, so no client secret is stored or rotated in code. It handles token caching and refresh against the identity platform, directly removing the expired-secret failures described in the scenario.

Why this answer

Microsoft Authentication Library (MSAL) is the recommended library for acquiring tokens from Microsoft Entra ID, including tokens for managed identities and service principals. MSAL handles token caching, refresh, and secure acquisition without embedding secrets in code. Using MSAL with managed identities lets the application authenticate without storing client secrets, directly addressing the expired-secret failures.

Exam trap

AI-103 often tests authentication library selection, and candidates confuse management-plane SDKs (Azure Management Client) with identity libraries (MSAL), or assume any Azure SDK can acquire tokens.

How to eliminate wrong answers

Option A is wrong because the Azure Storage Data Movement Library is for high-performance transfer of data to and from Azure Storage, not for acquiring Entra ID tokens. Option B is wrong because System.Net.Http WebRequestHandler is a low-level HTTP handler for configuring TLS and proxies, not an authentication library. Option C is wrong because the Azure SDK for .NET Management Client is used to manage Azure resources via ARM, not to acquire user or service tokens for application authentication.

Full explanation →

404

Multi-Selecthard

You are troubleshooting an intermittent connectivity issue between a client application and an Azure AI Speech service that sits behind an Azure Firewall and a Virtual Network. Which THREE diagnostic steps or tools should you utilize? (Choose three.)

Select 3 answers

A.Use Azure Network Watcher IP Flow Verify to check if traffic is being blocked by network security groups.

B.Run the Azure Advisor cost recommendation analyzer to optimize resource pricing tiers.

C.Enable Azure AI service diagnostic settings to send request logs and metrics to a Log Analytics workspace.

D.Inspect Network Security Group (NSG) flow logs to analyze inbound and outbound traffic metadata.

E.Use the Azure Resource Manager template visualizer to inspect regional resource dependencies.

AnswersA, C, D

IP Flow Verify tests a specific five-tuple against the effective NSG rules on a target VM, returning allow or deny plus the blocking rule name. That pinpoints whether a network security group, rather than the firewall, intermittently drops traffic between the client and the Azure AI Speech endpoint.

Why this answer

Option A is correct because Azure Network Watcher IP Flow Verify tests a specific 5-tuple (source/destination IP, port, protocol) against the effective NSG rules on a VM's NIC and reports whether traffic is allowed or denied, which directly isolates NSG-caused intermittent blocking. Option C is correct because enabling Azure AI service diagnostic settings routes request/response logs and metrics to a Log Analytics workspace, letting you correlate service-side failures and latency with client-side symptoms. Option D is correct because NSG flow logs capture inbound and outbound flow metadata (allowed/denied, source/destination, ports, bytes) that reveals intermittent drops and asymmetric traffic patterns over time.

Option B is wrong because Azure Advisor cost recommendations address pricing and utilization, not connectivity diagnostics. Option E is wrong because the ARM template visualizer only shows resource deployment dependencies and cannot trace live network traffic or firewall/NSG behavior.

Exam trap

AI-103 often tests whether candidates pick cost or template tools for a network troubleshooting scenario; the correct answers must be network-observability tools (IP Flow Verify, NSG flow logs, diagnostic settings), not Advisor or ARM visualizer.

Full explanation →

405

MCQmedium

You need to extract data from a highly varied set of tax forms where the layout changes frequently. Which approach is most efficient?

A.Use a Template model

B.Use a Neural custom model

C.Use Read API

D.Use General Document model

AnswerB

Neural custom models learn from labelled examples and tolerate layout variation, so frequently changing tax form structures do not break extraction. Template-based or form-specific approaches fail when fields shift position, whereas the neural model generalises across the highly varied layouts described in the stem.

Why this answer

A Neural custom model in Azure AI Document Intelligence is trained on your own labeled data and can generalize to layout variations, making it ideal for frequently changing tax forms. It learns the structure and key-value pairs from examples rather than relying on fixed templates. This provides the most efficient extraction for highly varied documents.

Exam trap

AI-103 often tests the trade-off between template and neural models, with candidates incorrectly assuming template models are more efficient for varied layouts due to their simplicity.

How to eliminate wrong answers

Option A is wrong because a Template model requires a fixed, consistent layout and fails when the layout changes frequently. Option C is wrong because the Read API only extracts text and does not understand structure or key-value pairs, so it cannot efficiently extract specific fields from tax forms. Option D is wrong because the General Document model is pre-trained for common document types and may not accurately extract specialized tax form fields, especially with layout variations.

Full explanation →

406

MCQhard

A financial services company uses a custom neural model in Azure AI Document Intelligence to extract fields from loan applications. After retraining the model with 30 additional labeled documents, field accuracy for the 'loanAmount' field dropped sharply, although the field was previously accurate. The new documents were labeled by a different team. You need to identify the most likely cause of the regression before retraining again. What should you investigate first?

A.Whether the service endpoint region was changed between trainings

B.Whether the model's confidence threshold was raised in the calling application

C.Whether the labeled documents use inconsistent field labels or value formatting for 'loanAmount' compared with the original training set

D.Whether the training dataset exceeded the maximum page count allowed for custom neural models

AnswerC

Custom neural models learn from label-value associations in the training data. If a second labeling team used a different field name, different currency formatting, or inconsistent bounding regions, the model receives contradictory supervision for the same field. That inconsistency directly degrades per-field accuracy and is the most probable cause of a sudden drop after adding documents.

Why this answer

Custom neural model quality depends heavily on consistent, accurate labels. When multiple teams label the same dataset, differences in field naming, value formatting, or region placement introduce conflicting training signals that disproportionately harm the affected field. Auditing label consistency for 'loanAmount' across both datasets is the correct first step before retraining.

Exam trap

The trap here is blaming infrastructure or configuration changes for a per-field accuracy regression when inconsistent training labels are the more common cause.

Full explanation →

407

Multi-Selecteasy

Which TWO of the following are valid Document Intelligence project types in the Studio?

Select 2 answers

A.Speech-to-text

B.Custom extraction

C.Sentiment analysis

D.Custom classification

E.Face detection

AnswersB, D

Custom extraction is a supported Document Intelligence Studio project type, training models to pull labelled field values from documents using custom neural or template approaches. It satisfies the stem by being one of the recognised project types available for selection in the Studio.

Why this answer

In Azure AI Document Intelligence Studio, the two valid project types are Custom extraction (B) and Custom classification (D). Custom extraction lets you build a model that pulls specific fields (key-value pairs, tables, entities) from documents using labeled samples, which is the core 'custom form'/'custom extraction' capability of Document Intelligence. Custom classification lets you train a model to categorize incoming documents into user-defined classes, a distinct project type also supported in the Studio.

The other options do not belong: Speech-to-text (A) is an Azure AI Speech capability, Sentiment analysis (C) is an Azure AI Language feature, and Face detection (E) is part of Azure AI Vision—none of these are Document Intelligence project types.

Exam trap

AI-103 often tests whether candidates can distinguish Document Intelligence capabilities (extraction, classification) from other Azure AI services (Speech, Language, Vision) — candidates who see 'custom' and assume any AI feature qualifies may incorrectly pick sentiment analysis or face detection.

Full explanation →

408

MCQmedium

Your custom extraction model is failing to identify fields on a new version of a document. You have labeled 50 documents. What is the most likely solution?

A.Change the API version to v2.1

B.Switch to a template model

C.Add more training documents with the new layout

D.Increase the confidence threshold

AnswerC

The model fails because the new layout differs from the 50 labelled samples it trained on. Adding training documents that reflect the new layout supplies the varied examples the model needs to generalise, satisfying the requirement to recognise fields on the updated document version.

Why this answer

When a custom extraction model fails on a new document layout, the most likely fix is to add more training documents that represent the new layout so the model can learn the additional variations. Custom models (especially neural) improve with more diverse, labeled examples covering the layouts they must handle; 50 documents of the old layout are insufficient for the new version. Retraining with representative samples is the standard remedy.

Exam trap

AI-103 often tests the misconception that tuning thresholds or switching model types fixes accuracy problems — the trap is selecting 'increase confidence threshold' or 'switch to template model' when the real issue is insufficient training data for the new layout.

How to eliminate wrong answers

Option A is wrong because changing the API version to v2.1 does not address a model accuracy problem on a new layout — API versions affect features and response schema, not the model's ability to recognize fields it was not trained on. Option B is wrong because switching to a template model would likely perform worse on a new layout — template models are rigid and require exact layout matching, so they fail even more readily on layout variations. Option D is wrong because increasing the confidence threshold only filters out low-confidence predictions; it does not improve the model's ability to detect fields it currently misses, and it would actually reduce recall further.

Full explanation →

409

MCQeasy

You are designing a customer service chatbot using Azure OpenAI. You need to ensure that the model refuses to answer questions outside the scope of your company's products. Where should you define this boundary instruction?

A.In the temperature setting

B.In the user prompt prefix

C.In the max_tokens parameter configuration

D.In the system message

AnswerD

The system message carries persistent instructions that shape model behaviour across every turn, making it the correct place to define scope boundaries and refusal behaviour. User-turn prompts are per-request and cannot reliably enforce a standing restriction.

Why this answer

The system message is correct because it is the designated channel for persistent, high-priority instructions that define the model's persona, scope, and refusal behavior across all turns of a conversation. Placing the boundary instruction there ensures it applies before and throughout the interaction, and it is the standard mechanism for scoping a chatbot to a domain. This is the architecturally correct location for guardrails that must not be overridden by user input.

Exam trap

AI-103 often tests whether candidates confuse generation parameters (temperature, max_tokens) with instruction channels, so the trap is selecting a sampling or length parameter as the place to define behavioral boundaries.

How to eliminate wrong answers

Option A is wrong because temperature controls sampling randomness (0 = deterministic, higher = more creative) and has no semantic content — it cannot express a scope boundary. Option B is wrong because a user prompt prefix is user-controlled and ephemeral; it applies only to that turn, can be overridden or ignored, and is vulnerable to prompt injection. Option C is wrong because max_tokens limits the length of the completion, not its subject matter — it constrains output size, not scope.

Full explanation →

410

MCQhard

You are optimizing a RAG pipeline in Azure AI Foundry. Users complain that answers are missing key numerical details because important tables within PDF documents are being split across chunk boundaries during ingestion. What is the recommended remediation?

A.Use Azure AI Document Intelligence with layout extraction to parse tables and maintain table structure during chunking

B.Increase the embedding model dimensions from 1536 to 3072

C.Set the Azure OpenAI model max_tokens parameter to 4096

D.Apply Azure AI Content Safety image analysis on all PDF pages

AnswerA

Layout extraction preserves table rows and columns as structured cells, so chunking splits on table boundaries rather than mid-table. This directly satisfies the stem's constraint that numerical table details are lost when tables straddle chunk boundaries.

Why this answer

Azure AI Document Intelligence with layout extraction preserves table structure by detecting rows, columns, and cells during parsing, so tables are chunked as coherent units rather than split arbitrarily. This keeps numerical relationships intact and prevents the retrieval step from returning partial tables that omit key figures. It is the recommended remediation for table-splitting problems in RAG ingestion.

Exam trap

AI-103 often tests whether candidates reach for model-side knobs (embedding dimensions, max_tokens) when the actual problem is data-ingestion parsing, so choose the option that fixes table structure at parse time.

How to eliminate wrong answers

Option B is wrong because increasing embedding dimensions from 1536 to 3072 changes vector representation size but does nothing to preserve table structure during chunking — the table is still split before embedding. Option C is wrong because max_tokens controls the model's output length, not how source documents are parsed or chunked. Option D is wrong because Content Safety image analysis screens images for harmful content and has no role in extracting or preserving tabular data from PDFs.

Full explanation →

411

MCQmedium

Your organization is running an Azure AI Search service and needs to restrict management operations (such as index creation and deletion) to administrators while allowing standard client applications to perform search queries only. How should you configure access?

A.Use Admin API keys for management operations and Query API keys for client search queries.

B.Enable IP firewall rules to block index creation.

C.Configure CORS settings in the Azure portal.

D.Assign the Contributor role to client applications.

AnswerA

Azure AI Search issues two key types with distinct privileges: admin keys authorise management operations like index creation and deletion, while query keys permit only search and lookup against existing indexes. This separation enforces least privilege for client applications.

Why this answer

Azure AI Search issues two key types: admin keys (primary and secondary) that grant full control including index creation, deletion, and configuration, and query keys that only permit read operations against indexes. Assigning admin keys to administrators and query keys to client applications enforces least privilege at the data-plane level without needing Azure RBAC.

Exam trap

AI-103 often tests the confusion between control-plane RBAC roles (Contributor) and data-plane key types, tricking candidates into picking a role that grants far more privilege than query-only access.

How to eliminate wrong answers

Option B is wrong because IP firewall rules restrict which networks can reach the service but do not differentiate between management and query operations — an allowed IP can still call index-creation APIs. Option C is wrong because CORS only controls which browser origins may call the search endpoint; it has no bearing on authorization or operation scoping. Option D is wrong because the Contributor role is a control-plane Azure RBAC role that grants management of the resource itself (including key regeneration and service deletion), which is far broader than needed and does not scope clients to query-only data-plane access.

Full explanation →

412

MCQmedium

Your development team is building a global customer support portal. Incoming tickets arrive in dozens of different languages. You need to automatically route tickets to the correct regional team based on the language of the ticket text. Which Azure AI Language feature should you call?

A.Sentiment analysis

B.Named Entity Recognition (NER)

C.Key phrase extraction

D.Language detection

AnswerD

Language detection returns the detected language and ISO code for input text, satisfying the requirement to route tickets by language. Calling it first identifies each ticket's language, letting the portal direct it to the correct regional team without manual triage.

Why this answer

Language detection is the Azure AI Language feature specifically designed to identify the language of a given text and return the detected language name and ISO 639-1 code. For routing tickets based on language, this is the exact capability needed. Sentiment analysis, NER, and key phrase extraction serve different purposes and do not identify language.

Exam trap

AI-103 often tests the confusion between language detection and other text analytics features, so candidates must remember that only language detection returns the language identifier, not sentiment or entities.

How to eliminate wrong answers

Option A is wrong because sentiment analysis determines positive, negative, or neutral tone, not the language of the text. Option B is wrong because Named Entity Recognition extracts entities like people, places, and organizations, but does not detect the overall language. Option C is wrong because key phrase extraction identifies main concepts in text, not the language itself.

Full explanation →

413

Multi-Selecthard

You are designing an architecture that processes incoming multilingual support emails using Azure AI Language. Which THREE processing steps should you include in your pipeline? (Choose three.)

Select 3 answers

A.Execute named entity recognition to extract relevant organizations, locations, and names.

B.Train a new speech-to-text acoustic model for every email received.

C.Call the language detection endpoint to identify the primary language of each incoming email.

D.Convert all email text into raw binary executable files before analysis.

E.Apply sentiment analysis or PII detection on the text documents.

AnswersA, C, E

Named entity recognition extracts organisations, locations, and person names from unstructured email text, populating structured fields for routing and case handling. This satisfies the stem's requirement for a processing step within the multilingual pipeline, operating on translated or native text regardless of the source language.

Why this answer

Option C is correct because Azure AI Language's language detection endpoint must be called first to identify each email's primary language, which is essential for routing multilingual content to the appropriate language-specific models. Option A is correct because named entity recognition (NER) extracts entities such as organizations, locations, and person names from the email text, providing structured, actionable information for support workflows. Option E is correct because sentiment analysis and PII detection are supported Azure AI Language features that help gauge customer tone and identify/redact sensitive data in the email documents.

Option B is not appropriate because speech-to-text acoustic models are unrelated to processing text emails and training a model per email is impractical. Option D is not appropriate because converting text to raw binary executables serves no analytical purpose and would destroy the text needed for language analysis.

Exam trap

AI-103 often tests the inclusion of irrelevant steps (like speech-to-text or binary conversion) to distract candidates, when the correct pipeline focuses on text analysis services like language detection, NER, and sentiment/PII analysis.

Full explanation →

414

MCQhard

You are configuring a webhook to receive notifications when an asynchronous batch text summarization job completes in Azure AI Language. Which Azure service should you use to capture and route the webhook events to your backend function?

A.Azure Event Grid

B.Azure API Management

C.Azure Stream Analytics

D.Azure Logic Apps

AnswerA

Azure Event Grid natively ingests Azure AI Language's asynchronous job-completion webhook events and routes them to your backend function, satisfying the requirement to capture and route notifications. It provides push-based delivery with retry and filtering, unlike polling-based queues, so the summarisation completion event reaches your function promptly.

Why this answer

Azure Event Grid is the event-routing service designed to deliver event notifications from Azure services to subscribers such as webhooks or Azure Functions. Azure AI Language batch jobs publish completion events that Event Grid captures and routes to the configured endpoint. This makes Event Grid the correct service for receiving and forwarding the webhook when the summarization job finishes.

Exam trap

AI-103 often tests event-driven integration patterns, and the trap is choosing Logic Apps because it 'sounds like' the workflow tool that handles events, when Event Grid is the actual routing backbone Azure services publish to.

How to eliminate wrong answers

Option B (Azure API Management) is wrong because APIM is an API gateway for publishing, securing, and throttling APIs, not an event routing or webhook delivery service. Option C (Azure Stream Analytics) is wrong because Stream Analytics performs real-time stream processing on data streams, not event subscription and webhook delivery. Option D (Azure Logic Apps) is wrong because Logic Apps is a workflow orchestration service that can consume events but is not the underlying event routing mechanism Azure services use to publish completion notifications.

Full explanation →

415

MCQhard

You are automating document ingestion for a RAG pipeline. Which feature allows you to automatically split large documents into smaller chunks during indexing?

A.Prompt Flow Vector Store Lookup

B.Model Fine-tuning

C.Indexers with Skillsets

D.Vector Embeddings API

AnswerC

Indexers with skillsets satisfy the chunking constraint: the Text Split skill within a skillset segments large documents into smaller chunks during indexing, and the indexer applies it automatically as content flows from the data source into the search index. This built-in enrichment pipeline removes the need for custom pre-processing code.

Why this answer

In Azure AI Search, indexers with skillsets are the feature that automates document ingestion and chunking: the indexer pulls source documents, and a Text Split skill in the skillset breaks large documents into smaller chunks before they are embedded and written to the index. This is the standard mechanism for building a RAG pipeline where chunk size affects retrieval quality. The skillset can also enrich content (OCR, entity extraction, embeddings) in the same pipeline.

Exam trap

AI-103 often tests the confusion between retrieval-time components (vector store lookup, embeddings API) and ingestion-time components (indexers with skillsets), so candidates pick the embedding API thinking it handles chunking.

How to eliminate wrong answers

Option A is wrong because Prompt Flow Vector Store Lookup is a retrieval-time node that queries an existing vector store; it does not perform ingestion or chunking of documents. Option B is wrong because model fine-tuning adapts model weights to a dataset and has nothing to do with splitting documents during indexing. Option D is wrong because the Vector Embeddings API converts text to vectors but does not itself split documents — chunking must happen before embedding, typically via the indexer/skillset pipeline.

Full explanation →

416

MCQeasy

Which Azure AI Foundry feature allows you to visualize the execution path of an agentic workflow?

A.Azure OpenAI Playground.

B.Azure AI Search dashboard.

C.Model catalog view.

D.Prompt Flow visual designer.

AnswerD

Prompt Flow's visual designer renders the agentic workflow as a directed graph, showing each node, tool call and connection along the execution path. This graphical trace lets developers inspect how the agent progressed between steps.

Why this answer

Prompt Flow's visual designer (DAG view) in Azure AI Foundry renders each node (LLM call, Python tool, retrieval step) and the directed edges between them, letting you trace the exact execution path of an agentic workflow. It also surfaces per-node inputs/outputs and token usage, which is why it is the designated tool for visualizing agent orchestration. The other options are single-purpose surfaces that do not expose flow topology.

Exam trap

AI-103 often tests whether candidates confuse observability surfaces — the trap is picking a playground or search dashboard for a question that is really about visualizing multi-step agent orchestration.

How to eliminate wrong answers

Option A is wrong because the Azure OpenAI Playground is only for interactively testing prompts and model parameters against a deployment; it has no concept of multi-step agent graphs. Option B is wrong because the Azure AI Search dashboard reports index statistics, query latency, and relevance metrics — it does not visualize agent execution. Option C is wrong because the Model Catalog is a discovery/browse experience for foundation models and their metadata, not a runtime tracing tool.

Full explanation →

417

MCQeasy

You need to programmatically delete an iteration of a Custom Vision model that performed poorly. Which API method should you invoke?

A.DELETE /customvision/v3.3/Training/{projectId}/iterations/{iterationId}

B.PUT /customvision/v3.3/Training/{projectId}/iterations/{iterationId}

C.GET /customvision/v3.3/Training/{projectId}/iterations/{iterationId}/purge

D.POST /customvision/v3.3/Training/{projectId}/iterations/{iterationId}/delete

AnswerA

The DELETE iterations endpoint targets a specific iteration by its iterationId within a project, permanently removing that trained model version. This directly satisfies the requirement to programmatically delete one poorly performing iteration, since the method operates at iteration scope rather than project or model scope.

Why this answer

The Custom Vision Training API exposes iteration deletion via the HTTP DELETE method on the path /customvision/v3.3/Training/{projectId}/iterations/{iterationId}. Invoking DELETE on that resource removes the specified iteration from the project. This is the standard RESTful pattern for deleting a subordinate resource identified by its parent project and iteration IDs.

Exam trap

AI-103 often tests REST verb semantics — candidates see a 'delete' action and gravitate toward POST /delete or GET /purge, but the correct answer is the HTTP DELETE verb on the resource path, reflecting standard REST conventions.

How to eliminate wrong answers

Option B is wrong because PUT on the iteration path is used to update an iteration's properties (e.g., rename it), not to delete it. Option C is wrong because GET on a /purge path is not a valid deletion operation — GET is a safe, idempotent read method and cannot delete resources; purge operations, where they exist, use DELETE or POST. Option D is wrong because POST to a /delete sub-resource is not the Custom Vision API's deletion pattern; the API uses the HTTP DELETE verb directly on the iteration resource.

Full explanation →

418

Multi-Selectmedium

When implementing Retrieval-Augmented Generation (RAG) using Azure AI Search and Azure OpenAI, which THREE strategies help optimize chunk retrieval relevance and reduce noise? (Choose three)

Select 3 answers

A.Configure hybrid search combining vector embeddings and keyword full-text search.

B.Store all corporate documents in a single unindexed raw text blob without metadata fields.

C.Set the embedding model dimensions to the lowest possible setting to maximize retrieval speed.

D.Implement document chunking strategies with appropriate overlap to preserve context across boundaries.

E.Enable semantic ranking to re-rank search results using deep learning models.

AnswersA, D, E

Hybrid search runs vector similarity and keyword full-text scoring in parallel, then fuses the rankings. Keyword matching catches exact terms and rare tokens that embeddings blur, while vectors capture paraphrase, so combining them raises relevance and suppresses irrelevant chunks in the RAG pipeline.

Why this answer

Option A is correct because Azure AI Search hybrid search runs vector (kNN/ANN) queries and keyword full-text (BM25) queries in parallel and fuses the results with Reciprocal Rank Fusion (RRF), combining semantic similarity with exact term matching to improve relevance and cut noise. Option D is correct because chunking documents into appropriately sized segments with overlap preserves context that would otherwise be lost at chunk boundaries, so embeddings and retrieved passages remain coherent and relevant. Option E is correct because semantic ranking (semantic reranker) applies deep-learning models to reorder the top results, promoting the most semantically relevant chunks and reducing irrelevant hits.

Option B is wrong because storing documents in a single unindexed blob without metadata prevents fielded search, filtering, and vector indexing, which increases noise rather than reducing it. Option C is wrong because reducing embedding dimensions to the minimum sacrifices semantic fidelity and recall; dimension choice should match the model's trained output (e.g., text-embedding-3-large at 3072 or a supported reduced size), not be minimized purely for speed.

Exam trap

AI-103 often tests whether candidates conflate 'faster retrieval' with 'better retrieval' — the trap is choosing lower embedding dimensions or raw blobs thinking they optimize RAG, when they actually destroy relevance.

Full explanation →

419

MCQeasy

You are using Azure AI Vision 4.0. You want to obtain a concise, single-sentence human-readable description of an image. Which feature parameter should you request?

A.features=caption

B.features=tags

C.features=denseCaptions

D.features=read

AnswerA

The caption feature generates a single human-readable sentence describing the image's content, matching the concise description requirement. Dense captions produce multiple phrases, and tags return keywords rather than sentences, so caption is the precise fit.

Why this answer

The features=caption parameter in Azure AI Vision 4.0 returns a single, concise, human-readable sentence describing the image — exactly what the question asks for. It is the direct match for 'concise, single-sentence description.'

Exam trap

AI-103 often tests caption vs denseCaptions by asking for 'a single sentence' — candidates who pick denseCaptions miss the 'single' qualifier.

How to eliminate wrong answers

Option B is wrong because features=tags returns a list of relevant keywords/objects, not a grammatical sentence. Option C is wrong because features=denseCaptions returns multiple captions for different regions of the image, not one concise sentence. Option D is wrong because features=read performs OCR to extract text from the image, not to describe its visual content.

Full explanation →

420

MCQmedium

Your development team needs to monitor the latency and total call volume of an Azure AI Document Intelligence resource in near real-time. Where can you view these pre-aggregated metrics without writing custom log queries?

A.The Cost Management + Billing cost analysis dashboard

B.The Log Analytics workspace query editor using KQL

C.The Metrics blade of the Azure AI Document Intelligence resource in the Azure portal

D.The Azure Advisor recommendations blade

AnswerC

The Metrics blade exposes platform-level, pre-aggregated telemetry such as latency and call volume natively, with no Log Analytics workspace or KQL query required. This directly satisfies the near real-time monitoring constraint without custom log queries.

Why this answer

The Metrics blade of an Azure AI Document Intelligence resource provides pre-aggregated, near-real-time platform metrics such as latency, call volume, and error rates without requiring any KQL queries. These metrics are emitted automatically by the Azure Monitor platform and are viewable directly in the portal.

Exam trap

AI-103 often tests the distinction between pre-aggregated platform metrics (Metrics blade, no query needed) and diagnostic logs requiring KQL — candidates pick Log Analytics because they assume all monitoring requires queries.

How to eliminate wrong answers

Option A is wrong because Cost Management + Billing only shows financial spend data, not operational telemetry like latency or call volume. Option B is wrong because the Log Analytics KQL query editor requires writing custom queries against diagnostic logs, which the question explicitly excludes. Option D is wrong because Azure Advisor provides best-practice recommendations, not live performance metrics.

Full explanation →

421

Multi-Selecthard

You are hardening an Azure OpenAI service deployment for a financial application that handles sensitive customer data. Which THREE security configurations should you apply? (Choose three.)

Select 3 answers

A.Store API keys in plain text within the client application's source code repository for easy developer access.

B.Authenticate client applications using Microsoft Entra ID (Azure AD) tokens instead of static API keys.

C.Enable anonymous access on the Azure OpenAI endpoint to simplify client connection testing.

D.Disable public network access and configure Azure Private Link with private endpoints.

E.Configure customer-managed keys (CMK) using Azure Key Vault for data encryption.

AnswersB, D, E

Microsoft Entra ID tokens provide short-lived, scoped, identity-based credentials with conditional access and audit trails, unlike static API keys that never expire and are easily leaked. This satisfies the sensitive-data constraint by removing long-lived shared secrets from client applications.

Why this answer

Option B is correct because authenticating with Microsoft Entra ID (Azure AD) tokens enables role-based access control (RBAC), short-lived credentials, and conditional access policies, eliminating the risk of leaked long-lived static API keys. Option D is correct because disabling public network access and using Azure Private Link with private endpoints keeps all traffic to the Azure OpenAI resource on the Microsoft backbone/private IP space, preventing exposure to the public internet and data exfiltration. Option E is correct because customer-managed keys (CMK) stored in Azure Key Vault give the financial organization control over the encryption keys protecting data at rest, satisfying regulatory and key-rotation requirements.

Option A is wrong because storing API keys in plain text in source control exposes credentials to anyone with repository access and is a critical secret-management anti-pattern. Option C is wrong because enabling anonymous access removes authentication entirely, allowing unauthenticated clients to invoke the model and access sensitive data.

Exam trap

AI-103 often tests whether candidates can distinguish genuine hardening controls from convenience features disguised as security — plaintext keys and anonymous access look 'easy' but are the exact anti-patterns the exam wants you to reject.

Full explanation →

422

Multi-Selecteasy

You are configuring an Azure AI Language resource. Which TWO pricing tiers are available for provisioning the resource in the Azure portal? (Choose two.)

Select 2 answers

A.Standard (S)

B.Unlimited (U1)

C.Free (F0)

D.Developer (DEV)

E.Enterprise (E1)

AnswersA, C

The Standard (S) tier is a provisioned pricing option available when creating an Azure AI Language resource in the portal, offering higher throughput and quota than the free tier. It satisfies the stem's requirement for an available provisioning tier.

Why this answer

The Standard (S) tier is correct because Azure AI Language resources are provisioned under the Standard (S) pricing tier, which provides the full production-level throughput and SLA-backed service for text analytics, question answering, and conversational language understanding features. The Free (F0) tier is also correct because Azure AI Language offers a Free (F0) tier for evaluation and low-volume testing, subject to limited transactions, document size, and rate limits per subscription. The Unlimited (U1) option is not a valid Azure AI Language pricing tier; no such tier exists in the Azure portal provisioning flow.

Developer (DEV) is not an Azure AI Language tier; developer-oriented access is handled through keys and endpoints on S or F0, not a separate DEV SKU. Enterprise (E1) is likewise not an Azure AI Language pricing tier; enterprise capabilities are delivered through the Standard tier rather than a distinct E1 SKU.

Exam trap

AI-103 often tests the available pricing tiers, and candidates may incorrectly assume there are developer or enterprise tiers, so remembering only F0 and S is crucial.

Full explanation →

423

MCQmedium

You are training a custom text classification model in Language Studio. After reviewing your trained model's performance metrics, you decide to improve recall for a specific underperforming category. What action should you take?

A.Increase the confidence threshold parameter in your client application code.

B.Delete all training documents for that category.

C.Switch the classification type from multi-label to single-label.

D.Add more varied training examples for that category to the dataset and retrain the model.

AnswerD

Recall improves when the model sees more diverse in-category examples, so adding varied labelled utterances for the underperforming category and retraining directly addresses the stem's recall goal. Precision may dip slightly, so re-evaluate metrics after retraining.

Why this answer

To improve recall for an underperforming category in custom text classification, you should add more varied training examples for that category and retrain the model. Recall measures the model's ability to find all relevant instances, so more diverse examples help the model generalize better. This is a standard approach in machine learning.

Exam trap

AI-103 often tests the misconception that adjusting thresholds or changing classification type improves recall, when the primary fix is more training data.

How to eliminate wrong answers

Option A is wrong because increasing the confidence threshold in the client application would make the model more conservative, likely reducing recall further. Option B is wrong because deleting training documents for that category would remove data and worsen performance. Option C is wrong because switching from multi-label to single-label changes the problem type and may not address the specific category's recall; it could even reduce performance if the data is inherently multi-label.

Full explanation →

424

MCQhard

You are implementing a grounded RAG solution in Azure AI Foundry. Users report that the model occasionally hallucinates answers even when relevant documents are retrieved by Azure AI Search. What is the most effective prompt engineering strategy to mitigate this grounding failure?

A.Remove the system message entirely so the model can freely blend internal training weights with search results.

B.Add strict prompt instructions stating that the model must rely solely on the retrieved context chunks and explicitly output a refusal statement if the answer is not present.

C.Increase the temperature parameter to 1.5 to encourage creative synthesis of the retrieved data.

D.Disable chunking in Azure AI Search so the entire document is passed as a single massive token block.

AnswerB

Explicit grounding instructions constrain generation to the retrieved chunks and mandate a refusal when the answer is absent, directly addressing hallucination despite relevant retrieval. This prompt-level guardrail satisfies the requirement to mitigate grounding failure without changing the search index.

Why this answer

The most effective prompt engineering mitigation for grounding failures is to explicitly instruct the model to answer only from the retrieved context and to refuse when the answer is absent. This constrains the model's generation to the provided evidence and reduces hallucination by giving it a clear escape hatch rather than forcing an answer.

Exam trap

The trap is thinking that more creativity (higher temperature) or more context (no chunking) improves answers, when both actually worsen grounding and increase hallucination in RAG scenarios.

How to eliminate wrong answers

Option A is wrong because removing the system message allows the model to blend internal training knowledge with retrieved context, which increases hallucination rather than reducing it. Option C is wrong because raising temperature to 1.5 increases randomness and creativity, directly worsening hallucination and grounding failures. Option D is wrong because disabling chunking and passing entire documents as massive token blocks dilutes relevant context, increases cost, and can exceed context window limits, making grounding worse rather than better.

Full explanation →

425

MCQmedium

You are building an AI agent in Microsoft Foundry that must answer questions grounded exclusively in a set of internal PDF policy documents. The agent must cite the exact page of the source document in every response, and the documents are updated weekly. You want the least operational overhead while keeping citations accurate. What should you do?

A.Create an Azure AI Search index manually, generate embeddings with a deployment, and wire the index to the agent through a custom function tool.

B.Upload the PDFs to the agent's file store and enable the built-in file search tool, which chunks, embeds, and indexes the documents in a managed vector store.

C.Enable the code interpreter tool and have the agent parse the PDFs at request time to extract text and page numbers.

D.Paste the full text of every PDF into the agent's system instructions so the model can quote the pages directly.

AnswerB

The built-in file search tool in Foundry performs chunking, embedding, and indexing into a managed vector store and returns citation metadata including file and page references. Because the store is managed, weekly updates require only re-uploading or syncing the files, minimizing operational overhead while preserving page-level citations in responses.

Why this answer

The managed file search tool handles chunking, embedding, and vector indexing automatically and returns citation metadata that includes file and page references. Because the index is rebuilt from uploaded files, weekly document refreshes need only file updates rather than custom reindexing pipelines, which satisfies both the accurate-citation requirement and the low-overhead requirement.

Exam trap

The trap here is assuming that a manually built Azure AI Search index is required for page-level citations, when the managed file search tool already returns file and page citation metadata.

Full explanation →

426

MCQeasy

You need to provision an Azure AI resource that allows developers to access multiple AI capabilities (such as Vision, Speech, and Language) using a single endpoint and a single API key. Which resource type should you create?

A.Azure Cognitive Search resource

B.Azure OpenAI resource

C.Azure AI Custom Vision resource

D.Azure AI multi-service account

AnswerD

Azure AI multi-service account bundles Vision, Speech, and Language under one endpoint and one API key, directly satisfying the single-endpoint, single-key constraint. Unlike individual Cognitive Services resources, which require separate endpoints and keys per capability, it consolidates billing and authentication, letting developers call multiple AI services without provisioning each separately.

Why this answer

An Azure AI multi-service account (Cognitive Services multi-service resource) bundles Vision, Speech, Language, Translator, and other services under one endpoint and one API key, simplifying provisioning and billing. It is the correct resource type when developers need multiple AI capabilities through a single credential. Single-service resources like Custom Vision or Azure OpenAI expose only their own capability.

Exam trap

AI-103 often tests whether candidates confuse single-service resources (Custom Vision, Azure OpenAI) with the multi-service account, so the trap is picking a resource that only covers one AI capability when the question demands several.

How to eliminate wrong answers

Option A is wrong because Azure Cognitive Search is a search-index service for full-text and vector search — it does not provide Vision, Speech, or Language AI capabilities. Option B is wrong because Azure OpenAI provides access to OpenAI models (GPT, DALL-E, embeddings) only, not the broader Cognitive Services suite. Option C is wrong because Azure AI Custom Vision is a single-service resource for image classification and object detection, exposing only that capability.

Full explanation →

427

MCQeasy

You are auditing your Azure subscription and find several unused Azure AI Content Safety resources that are incurring base charges. You want to remove the resources permanently. Which operation should you perform?

A.Delete the Azure AI Content Safety resource from the Azure portal.

B.Remove the role assignments from the resource.

C.Stop the Azure AI Content Safety resource state.

D.Disable the keys associated with the resource.

AnswerA

Deleting the resource removes it and its associated base charges permanently. Stopping or scaling down would not eliminate billing for a provisioned Content Safety resource, so deletion is the only operation that satisfies the requirement to remove it permanently.

Why this answer

Deleting the Azure AI Content Safety resource removes it permanently from the subscription, which stops all billing including any base or provisioned charges. Since the requirement is to remove unused resources permanently, deletion is the only operation that eliminates the resource and its associated costs.

Exam trap

AI-103 often tests the difference between disabling access (keys, roles) and deprovisioning a resource — candidates pick 'disable keys' thinking it stops billing, but only deletion removes the billing meter.

How to eliminate wrong answers

Option B is wrong because removing role assignments only revokes access permissions — the resource still exists and continues to incur charges. Option C is wrong because Azure AI Content Safety does not support a 'stop' state like a virtual machine; there is no pause operation that halts billing while retaining the resource. Option D is wrong because disabling keys only prevents authentication to the resource — the resource remains provisioned and billable, and keys can be re-enabled later.

Full explanation →

428

MCQhard

Your enterprise application uses Azure AI Vision. Security policies require that all API keys are rotated every 30 days without causing application downtime. How should you manage the key rotation?

A.Restart the Azure App Service hosting the client application.

B.Switch client applications to Key 2, regenerate Key 1 in the Azure portal, switch clients to Key 1, and then regenerate Key 2.

C.Delete the Azure AI service resource and create a new one every 30 days.

D.Disable authentication on the resource during the rotation window.

AnswerB

Azure AI Vision provides two keys precisely so rotation avoids downtime: clients move to Key 2 while Key 1 is regenerated, then back once Key 2 is replaced. This staggered swap satisfies the 30-day policy without interrupting calls, unlike single-key regeneration, which invalidates live requests immediately.

Why this answer

Azure AI services (including Azure AI Vision) issue two API keys per resource specifically to enable zero-downtime rotation. The correct pattern is to move all clients to Key 2 while Key 1 is still valid, regenerate Key 1, cut clients back over to the new Key 1, then regenerate Key 2 — so at no point is a client using a key that has just been invalidated. This dual-key design is the documented Microsoft approach for credential rotation without service interruption.

Exam trap

AI-103 often tests whether candidates understand that Azure AI resources ship with two keys precisely so rotation can be performed without downtime — the trap is picking an option that either restarts services or disables auth instead of using the dual-key swap.

How to eliminate wrong answers

Option A is wrong because restarting the App Service does nothing to rotate credentials and would cause an outage, not prevent one. Option C is wrong because deleting and recreating the resource destroys the endpoint, custom model data, and configuration, causing a guaranteed outage and requiring re-provisioning. Option D is wrong because disabling authentication exposes the resource to unauthenticated access — a severe security violation that contradicts the stated security policy.

Full explanation →

429

Multi-Selecteasy

You are designing a text analysis solution using Azure AI Language. Which TWO features are available as prebuilt capabilities without requiring custom model training? (Choose two.)

Select 2 answers

A.Custom named entity recognition

B.Language detection

C.Sentiment analysis

D.Custom text classification

E.Custom summarization models

AnswersB, C

Language detection is a prebuilt Azure AI Language capability: the service returns the detected language and confidence score from supplied text without any labelled training data or custom model deployment, satisfying the no-training constraint in the stem.

Why this answer

Option B (Language detection) is correct because Azure AI Language provides it as a prebuilt capability that identifies the language of input text using the ISO 639-1 code, requiring no training data or custom model. Option C (Sentiment analysis) is also correct because it is a prebuilt feature that returns sentiment labels (positive, negative, neutral, mixed) and confidence scores out of the box. In contrast, option A (Custom named entity recognition) and option D (Custom text classification) are custom features that require labeled training data and a trained project before use.

Option E (Custom summarization models) is not a prebuilt capability; summarization is offered as prebuilt abstractive/extractive summarization, and any custom summarization model would require training, so it does not fit the 'no custom training' requirement.

Exam trap

AI-103 often tests the prebuilt-versus-custom distinction, and the trap is selecting custom NER or custom classification because they sound like standard features, when both require labeled data and a training step.

Full explanation →

430

Multi-Selectmedium

You are managing Azure OpenAI deployments and need to monitor token consumption and error rates. Which TWO Azure Monitor features should you use to analyze this telemetry? Each correct answer presents part of the solution.

Select 2 answers

A.Azure Monitor Metrics Explorer to graph real-time metrics like Total Calls and Latency.

B.Azure DevOps commit history dashboard

C.Log Analytics workspace with Kusto Query Language (KQL) queries for deep request log analysis.

D.Microsoft 365 compliance center content search

E.Azure Bastion session monitor

AnswersA, C

Metrics Explorer plots platform metrics emitted by Azure OpenAI, including Total Calls, Latency and token usage, giving near real-time visualisation. It satisfies the monitoring requirement for spotting consumption trends and error-rate spikes without writing queries.

Why this answer

Option A is correct because Azure Monitor Metrics Explorer lets you plot near real-time platform metrics for Azure OpenAI, including Total Calls, Token Usage (prompt/completion tokens), and Latency, which directly addresses monitoring token consumption and error rates at a glance. Option C is correct because Azure OpenAI diagnostic settings can route request/response logs and metrics to a Log Analytics workspace, where Kusto Query Language (KQL) queries enable deep analysis of per-request token counts, status codes, and error patterns over time. Together, Metrics Explorer provides fast visual monitoring while Log Analytics with KQL provides granular, queryable telemetry.

Option B is incorrect because Azure DevOps commit history tracks source code changes, not Azure OpenAI runtime telemetry. Option D is incorrect because Microsoft 365 compliance center content search is for eDiscovery of Microsoft 365 content, not Azure Monitor telemetry. Option E is incorrect because Azure Bastion session monitoring relates to secure RDP/SSH access to VMs, not OpenAI token or error metrics.

Exam trap

AI-103 often tests whether candidates know that platform metrics are automatic and near-real-time while detailed per-request telemetry requires diagnostic settings routed to Log Analytics — mixing these up leads to picking only one of the two correct answers.

Full explanation →

431

Multi-Selecthard

When designing agentic workflows in Azure AI Foundry, what are THREE key characteristics or components of agentic design compared to traditional single-turn prompt completion? (Choose THREE)

Select 3 answers

A.Tool use and external API invocation capabilities

B.Hardcoding every conditional execution path in procedural code

C.Iterative reasoning and self-reflection loops

D.Autonomous planning and task decomposition

E.Executing single stateless prompt completions without memory

AnswersA, C, D

Agents invoke tools and external APIs to act beyond text generation, retrieving live data or triggering operations. This capability distinguishes agentic design from single-turn completion, which only returns a static response with no external side effects.

Why this answer

Option A (Tool use and external API invocation capabilities) is correct because agentic workflows in Azure AI Foundry extend beyond text generation by letting agents call tools, functions, and external APIs (for example via function calling or Azure AI Foundry tool connections) to retrieve data or perform actions. Option C (Iterative reasoning and self-reflection loops) is correct because agents can evaluate intermediate results, critique their own output, and re-plan across multiple reasoning steps rather than producing one final completion. Option D (Autonomous planning and task decomposition) is correct because agentic design centers on breaking a high-level goal into subtasks and deciding the order of execution dynamically, often orchestrated by the model itself.

Option B is incorrect because hardcoding every conditional path is the opposite of agentic design, which relies on model-driven decisions and orchestration rather than fully procedural branching. Option E is incorrect because single stateless prompt completions without memory describe traditional single-turn completion, whereas agentic workflows typically maintain state, context, and memory across turns.

Exam trap

AI-103 often tests whether candidates recognize that agentic design is defined by autonomy, tool use, and iteration, so the trap is selecting a deterministic procedural-coding answer as if it were an agentic characteristic.

Full explanation →

432

Multi-Selectmedium

Which TWO criteria are typically used to evaluate the grounding of an LLM-based application?

Select 2 answers

A.Relevance

B.Latency

C.Token cost

D.Groundedness

E.Model version number

AnswersA, D

Relevance measures how directly retrieved passages address the user's query, filtering out context that is topically adjacent but unhelpful. Grounding evaluation specifically requires this query-to-source alignment, since an answer can be fluent yet unsupported. Relevance therefore satisfies the stem's grounding criterion by verifying the retrieved evidence actually pertains to what was asked.

Why this answer

Option A, Relevance, is correct because grounding evaluation checks whether the model's response actually addresses the user's query using the retrieved context, so relevance measures how well the answer aligns with the question and the supplied source material. Option D, Groundedness, is correct because it directly measures whether the generated statements are supported by, or entailed by, the provided reference documents rather than being hallucinated or drawn from the model's parametric memory. Together, relevance and groundedness are the two standard quality dimensions for assessing retrieval-augmented or otherwise grounded LLM applications.

The unmarked options do not belong: Latency (B) and Token cost (C) are operational efficiency metrics, not grounding quality criteria, and Model version number (E) is a configuration identifier that says nothing about whether an answer is faithful to its sources.

Exam trap

AI-103 often tests whether candidates distinguish quality/grounding metrics from operational metrics like latency, cost, or version numbers, which are commonly mixed into answer choices as distractors.

Full explanation →

433

MCQhard

You are building an Azure AI Document Intelligence solution to process complex financial statements. Due to varying layouts across different financial institutions, you decide to use a custom neural model instead of a custom template model. What is a key architectural advantage of the custom neural model in this scenario?

A.It requires significantly fewer training documents than a template model to achieve high accuracy.

B.It can be trained entirely unsupervised without requiring any human-labeled training documents.

C.It supports cross-page tables and varying visual layouts without requiring strict coordinate alignment.

D.It can extract data from scanned documents without running Optical Character Recognition (OCR) first.

AnswerC

Custom neural models use deep learning to handle variable layouts, cross-page tables and differing visual structures, so financial statements from different institutions need no fixed coordinate templates. This satisfies the varying-layout constraint that a custom template model, reliant on rigid positional alignment, cannot accommodate.

Why this answer

The custom neural model in Azure AI Document Intelligence is designed to handle complex, variable layouts, including cross-page tables and differing visual structures, without requiring strict coordinate alignment. This makes it ideal for financial statements from different institutions where layouts vary significantly. It leverages deep learning to understand document structure and extract fields accurately even when the position of data changes.

Exam trap

AI-103 often tests the assumption that neural models require less training data or are unsupervised, but they actually require labeled data and are supervised; the key advantage is layout flexibility, not data efficiency.

How to eliminate wrong answers

Option A is wrong because custom neural models typically require more training data (at least 5 labeled documents, but often more) compared to template models, which can work with as few as 5. Option B is wrong because custom neural models require human-labeled training documents; they are not unsupervised. Option D is wrong because all Document Intelligence models rely on OCR to extract text from scanned documents before further processing; neural models do not bypass OCR.

Full explanation →

434

Multi-Selectmedium

You are planning a text analytics pipeline that uses asynchronous batch operations. Which THREE text analysis tasks support asynchronous long-running batch operations in Azure AI Language? (Choose three.)

Select 3 answers

A.Simple key phrase extraction

B.Health text analytics

C.Extractive summarization

D.Abstractive summarization

E.Language detection

AnswersB, C, D

Health text analytics runs as an asynchronous long-running batch operation in Azure AI Language, extracting entities and relations from clinical documents. It satisfies the stem's batch constraint because jobs are submitted and polled rather than returning synchronously.

Why this answer

Options B, C, and D are correct because Azure AI Language exposes asynchronous long-running batch operations specifically for Health text analytics (B), Extractive summarization (C), and Abstractive summarization (D); these tasks accept a batch of documents via a submit job request and return results through a polling/operation-location pattern, which is designed for large or long-running workloads. In contrast, A (Simple key phrase extraction) and E (Language detection) are synchronous, single-request operations that return results immediately and do not provide an asynchronous batch job API.

Exam trap

AI-103 often tests the assumption that all text analytics features support async batch jobs, when in fact only health analytics and the two summarization types do — key phrase extraction and language detection remain synchronous.

Full explanation →

435

Multi-Selectmedium

You are designing an information extraction solution using Azure AI Document Intelligence. Which TWO features are supported by Custom Neural models? Each correct selection represents a complete solution element.

Select 2 answers

A.Extraction of key-value pairs and fields defined during training

B.Zero-shot extraction of completely unseen document types without any training data

C.Extraction of tables from documents with variable layouts

D.Real-time voice-to-text transcription of document contents

E.Automatic language translation of extracted text into 100+ languages

AnswersA, C

Custom Neural models are trained on labelled samples to extract the specific key-value pairs and fields you define, generalising across layout variation. This directly satisfies the stem's requirement for trained field extraction, which template models cannot achieve when document structure varies.

Why this answer

Option A is correct because Custom Neural models are trained on labeled samples to extract the specific key-value pairs and fields you define during training, returning structured JSON with confidence scores for those fields. Option C is correct because Custom Neural models support table extraction, including documents with variable layouts, by learning the structure from your labeled training data. Option B is incorrect because zero-shot extraction of unseen document types without training data is not a Custom Neural capability; that is handled by prebuilt models or the newer query-field/zero-shot features, not by training a Custom Neural model.

Option D is incorrect because voice-to-text transcription is a speech service capability, not a Document Intelligence extraction feature. Option E is incorrect because automatic translation into 100+ languages is provided by Azure AI Translator, not by Custom Neural models.

Exam trap

AI-103 often tests the misconception that Custom Neural models can perform zero-shot extraction without training, causing candidates to select that option instead of the actual supervised capabilities.

Full explanation →

436

MCQeasy

Which component in Azure AI Foundry allows you to visualize and debug the execution path of a model call, including input/output mapping?

A.Prompt Flow

B.Compute Instance

C.Model Benchmarks

D.Deployment Endpoint

AnswerA

Prompt Flow in Azure AI Foundry builds a directed graph of nodes, letting you trace each step's inputs and outputs and debug the model call's execution path. This visual tracing directly satisfies the requirement to inspect input/output mapping, which a plain deployment endpoint cannot show.

Why this answer

Prompt Flow in Azure AI Foundry is the orchestration and authoring tool that lets you build flows as DAGs of nodes (LLM, Python, prompt, tool) and visually inspect each node's inputs/outputs, execution time, and token usage. It is purpose-built for debugging and tracing the execution path of a model call, including input/output mapping between nodes.

Exam trap

AI-103 often tests the confusion between Prompt Flow (authoring/debugging/tracing) and deployment endpoints or benchmarks — candidates must know Prompt Flow is the visualization and debugging layer, not the serving or evaluation layer.

How to eliminate wrong answers

Option B is wrong because a Compute Instance is just the VM that hosts the development environment (notebooks, VS Code) — it provides compute, not visualization of execution paths. Option C is wrong because Model Benchmarks evaluate and compare model quality/performance metrics; they do not trace a single call's execution path or input/output mapping. Option D is wrong because a Deployment Endpoint is the runtime URL that serves the model for inference — it exposes the model but does not visualize or debug the flow's internal execution.

Full explanation →

437

MCQhard

Your application analyzes high-resolution engineering blueprints using Azure AI Vision Read API. Users report that small text annotations are occasionally missed. How can you optimize the input image or API configuration to improve recognition of small text?

A.Convert the blueprint to grayscale and reduce color depth to 1-bit monochrome.

B.Switch from Read API to Tagging API.

C.Ensure the image is submitted at full resolution without lossy compression, and verify contrast.

D.Downsample the image resolution to under 1000x1000 pixels before submission.

AnswerC

Submitting the image at full resolution without lossy compression preserves the pixel detail small annotations require, while adequate contrast keeps glyph edges distinct. This satisfies the recognition constraint by preventing downsampling and compression artefacts that erase fine text before the Read API processes it.

Why this answer

The Read API's OCR accuracy for small text depends heavily on the effective pixel density of the characters. Submitting the blueprint at full resolution without lossy compression preserves the fine strokes of small annotations, and verifying sufficient contrast ensures the OCR engine can distinguish text from the background. Lossy compression (JPEG) introduces artifacts that blur small glyphs, directly causing missed detections.

Exam trap

AI-103 often tests the misconception that 'smaller image = faster/cleaner OCR', when in fact the Read API needs sufficient pixel height per character, so downsampling or aggressive compression is the wrong optimization.

How to eliminate wrong answers

Option A is wrong because reducing to 1-bit monochrome discards grayscale anti-aliasing information that the Read API uses to resolve thin or low-contrast strokes, often degrading small-text recognition. Option B is wrong because the Tagging API returns image content tags (e.g., 'outdoor', 'building') and does not perform OCR at all, so it cannot read annotations. Option D is wrong because downsampling to under 1000x1000 pixels reduces the pixel height of small text below the ~12-15 pixel threshold the Read engine needs, making recognition worse, not better.

Full explanation →

438

MCQmedium

You are configuring a Prompt Flow to use a custom Python tool. Which configuration is required to ensure the environment has the necessary libraries?

A.Install them manually on the host machine.

B.Include them in a 'requirements.txt' file in the flow directory.

C.Add them to the System Prompt.

D.Add them to the deployment manifest.

AnswerB

Placing a requirements.txt file in the flow directory lets Prompt Flow resolve and install the declared Python packages into the flow's runtime environment during execution. This directly satisfies the stem's constraint that the custom Python tool has its necessary libraries available, since dependencies are provisioned automatically rather than assumed present.

Why this answer

Prompt Flow resolves Python dependencies for a flow by reading a requirements.txt file located in the flow directory. This file lists the packages and versions needed by custom Python tools, and the Prompt Flow runtime installs them when the flow is executed or built. This ensures reproducibility across local development and cloud execution.

Exam trap

AI-103 often tests whether candidates confuse prompt-level configuration (system prompt) with environment-level configuration (requirements.txt), leading them to pick the system prompt option.

How to eliminate wrong answers

Option A is wrong because manually installing libraries on the host machine does not guarantee the same environment in the Prompt Flow runtime or when the flow is deployed, breaking reproducibility. Option C is wrong because the system prompt is natural-language instruction text sent to the model and has no role in Python package installation. Option D is wrong because the deployment manifest describes how the flow is hosted (endpoint, compute, etc.), not the Python dependency set.

Full explanation →

439

Multi-Selectmedium

You are preparing training images for a Custom Vision object detection project. Which THREE data preparation practices should you follow to ensure high model accuracy? (Choose three.)

Select 3 answers

A.Vary object angles, lighting, and scale across your training images.

B.Use identical background images for all training samples to maintain consistency.

C.Ensure at least 15 tagged images per tag for object detection.

D.Include images with varied backgrounds for the same object class.

E.Keep all training images under 100x100 pixels to speed up training.

AnswersA, C, D

Varying angles, lighting and scale exposes the detector to the appearance variability it will meet at inference, preventing overfitting to a single viewpoint. This directly satisfies the accuracy goal by teaching the model that a class's features persist across transformations rather than being tied to one capture condition.

Why this answer

Option A is correct because varying object angles, lighting, and scale exposes the model to the natural variability it will encounter at inference time, which is essential for a robust Custom Vision object detector. Option C is correct because Microsoft's Custom Vision documentation recommends a minimum of 15 tagged images per tag for object detection, and more images generally yield better accuracy. Option D is correct because including varied backgrounds for the same object class prevents the model from overfitting to a single context and helps it generalize to new environments.

Option B is incorrect because using identical backgrounds for all samples reduces variability and encourages overfitting to that background rather than learning the object itself. Option E is incorrect because forcing all images under 100x100 pixels discards detail and resolution that object detection models need to localize and classify objects accurately.

Exam trap

AI-103 often tests the misconception that 'consistency' in training data (same background, same size) improves accuracy, when in fact diversity of angles, lighting, scale, and backgrounds is what drives generalization.

Full explanation →

440

MCQhard

You are designing an automated deployment pipeline for Azure AI Search using GitHub Actions. The pipeline must authenticate to Azure securely without storing credentials or long-lived secrets in GitHub repository settings. Which authentication mechanism should you implement?

A.Hardcoded administrator API keys stored in GitHub Action variables

B.OpenID Connect (OIDC) with Azure workload identity federation

C.A personal access token (PAT) with full subscription ownership

D.Anonymous access enabled during deployment time

AnswerB

OIDC with workload identity federation lets GitHub Actions exchange a short-lived token for an Azure access token, eliminating stored secrets. This satisfies the stem's constraint of authenticating without credentials or long-lived secrets in repository settings, unlike service principal secrets or PATs.

Why this answer

OpenID Connect (OIDC) with Azure workload identity federation allows GitHub Actions workflows to authenticate to Azure using short-lived tokens issued by GitHub's OIDC provider and exchanged for Azure AD tokens via a federated identity credential. This eliminates the need to store any secret in GitHub, satisfying the 'no long-lived credentials' requirement. It is the Microsoft-recommended pattern for CI/CD pipelines deploying Azure resources such as Azure AI Search.

Exam trap

AI-103 often tests the misconception that storing API keys in GitHub secrets is 'secure enough' — the exam requires recognizing that any stored credential, even encrypted, violates the passwordless/OIDC requirement.

How to eliminate wrong answers

Option A is wrong because hardcoded administrator API keys are long-lived secrets stored in GitHub, directly violating the requirement and exposing the search service to credential leakage. Option C is wrong because a PAT with full subscription ownership is both a long-lived secret and massively over-privileged, violating least privilege and the no-secrets rule. Option D is wrong because anonymous access during deployment would expose the Azure AI Search endpoint to unauthenticated users and is not a valid authentication mechanism for a deployment pipeline.

Full explanation →

441

MCQmedium

You are building an Azure AI Foundry agent that answers employee benefits questions from an HR policy PDF set. Legal requires that every response cite the exact source page, and that unsupported answers are refused. You must maximize groundedness while keeping the existing PDFs unchanged. Which approach should you implement?

A.Extend the agent with the Azure AI Search tool connected to an index built with integrated vectorization and semantic ranking, and instruct the agent to cite retrieved chunks.

B.Upload the PDFs as file attachments to the agent and rely on the model's built-in file search with the default chunking settings.

C.Increase the model temperature and top_p so the agent produces more varied, comprehensive answers with citations.

D.Fine-tune the underlying GPT-4o deployment on the HR policy PDFs and route all benefit questions to the fine-tuned model.

AnswerA

An Azure AI Search index with integrated vectorization and semantic ranker returns the most relevant policy chunks, which the agent can ground on and cite by source page. Because the PDFs stay in Blob Storage and are chunked and embedded during indexing, no source modification is required, and citation instructions keep answers tied to retrieved evidence.

Why this answer

Grounding an Azure AI Foundry agent on an Azure AI Search index with integrated vectorization and semantic ranking supplies the retrieved evidence needed for accurate, page-level citations without altering the source PDFs. Combining that retrieval tool with explicit citation instructions also enables the refusal behavior legal requires when no supporting chunk is found.

Exam trap

The trap here is assuming that attaching files or fine-tuning automatically yields page-accurate citations, when only a retrieval index with controlled chunking and semantic ranking provides verifiable source grounding.

Full explanation →

442

MCQhard

Your application is receiving 429 errors from the Document Intelligence service. What does this indicate?

A.Authentication failed

B.Rate limit exceeded

C.Server-side error

D.The request is malformed

AnswerB

HTTP 429 signals that the Document Intelligence service is throttling requests because the call rate or resource consumption has exceeded the assigned quota. This directly satisfies the stem's constraint: the application is being rate-limited, so retry with exponential backoff or request a quota increase.

Why this answer

HTTP 429 'Too Many Requests' indicates the client has exceeded the rate limit or quota for the Document Intelligence service. Azure Cognitive Services enforce per-second and per-minute transaction limits based on the pricing tier (e.g., S0 has specific TPS limits). The correct response is to implement retry with exponential backoff and/or request a quota increase.

Exam trap

AI-103 often tests the confusion between 429 (rate limit) and 401 (auth) or 400 (bad request) — candidates must map HTTP status codes precisely to their meanings for Azure AI services.

How to eliminate wrong answers

Option A is wrong because authentication failures return HTTP 401 Unauthorized, not 429. Option C is wrong because server-side errors return HTTP 5xx codes (500, 502, 503), not 429. Option D is wrong because malformed requests return HTTP 400 Bad Request, indicating a client-side syntax or validation issue, not a rate limit.

Full explanation →

443

Multi-Selectmedium

When implementing prompt engineering for generative AI applications in Azure AI Foundry, which TWO practices are considered best practices for maintaining prompt maintainability and security? (Choose TWO)

Select 2 answers

A.Writing all prompts as single unending paragraphs without formatting

B.Embedding API keys directly inside prompt strings

C.Storing and versioning prompt templates as assets in Azure AI Foundry

D.Hardcoding user names and passwords into system prompts

E.Separating system instructions from untrusted user inputs using clear delimiters or XML tags

AnswersC, E

Storing and versioning prompt templates as assets in Azure AI Foundry gives prompts source control, audit history and rollback, satisfying the maintainability requirement. Changes are tracked centrally rather than scattered in code, so teams can review, compare and revert prompt revisions safely.

Why this answer

Option C is correct because Azure AI Foundry supports storing and versioning prompt templates as managed assets, which enables change tracking, rollback, reuse across deployments, and consistent governance instead of scattering prompts through application code. Option E is correct because separating system instructions from untrusted user input with clear delimiters or XML tags helps prevent prompt injection and instruction-override attacks, and it makes the prompt structure easier to read and maintain. Options A, B, and D are not best practices: unformatted single-paragraph prompts are hard to review and maintain, embedding API keys in prompt strings exposes secrets that can leak through logs or model output, and hardcoding user names and passwords into system prompts creates a serious credential disclosure risk and violates the principle of keeping secrets out of prompts.

Exam trap

AI-103 often tests the misconception that embedding secrets in prompts or using unformatted prompts is acceptable, when in fact secure and maintainable practices require versioned templates and clear delimiters.

Full explanation →

444

MCQeasy

Your company wants to generate descriptive captions for thousands of product images automatically using Azure AI Vision. Which visual feature parameter should you specify in the analyze request?

A.features=read

B.features=smartCrops

C.features=caption

D.features=objects

AnswerC

Specifying features=caption invokes the Caption visual feature, which returns a single-sentence natural-language description of an image via the Azure AI Vision analyse endpoint. This directly satisfies the stem's requirement to generate descriptive captions for thousands of product images automatically, without custom model training.

Why this answer

The `caption` feature in Azure AI Vision's Analyze Image API generates a human-readable, one-sentence description of the entire image content. It is specifically designed for scenarios like automatically describing product images at scale. The API returns a `captions` array with `text` and `confidence` values, making it the correct choice for descriptive captioning.

Exam trap

AI-103 often tests the confusion between OCR (read) and image captioning (caption), or between object detection (objects) and captioning, so candidates must remember that only `caption` produces descriptive sentences.

How to eliminate wrong answers

Option A is wrong because `features=read` invokes the Read OCR engine, which extracts printed or handwritten text from images rather than generating descriptive captions. Option B is wrong because `features=smartCrops` returns suggested crop regions and aspect ratios for thumbnails, not natural-language descriptions. Option D is wrong because `features=objects` detects and localizes objects with bounding boxes and labels, but does not produce full-sentence captions.

Full explanation →

445

MCQeasy

You are provisioning a new Azure AI Search service. You need to choose a pricing tier that supports semantic search capabilities. Which minimum tier should you select?

A.Basic

B.Free

C.Storage Optimized (L1)

D.Standard (S1)

AnswerD

Semantic search requires the Standard S1 tier or above; Free and Basic tiers lack semantic ranking entirely. Selecting S1 satisfies the stem's constraint that the Azure AI Search pricing tier must support semantic search capabilities.

Why this answer

Semantic search in Azure AI Search requires the Standard tier (S1) or higher. The Basic and Free tiers do not support semantic ranker, which uses deep learning models to re-rank results. Storage Optimized tiers are for large-scale storage but also support semantic search; however, the question asks for the minimum tier, and Standard is the lowest tier that includes semantic search.

Therefore, Standard (S1) is correct.

Exam trap

AI-103 often tests the misconception that Basic or Free tiers support advanced features like semantic search, but candidates must remember that semantic search requires at least the Standard tier.

How to eliminate wrong answers

Option A is wrong because the Basic tier does not support semantic search; it only offers basic full-text search with limited indexing capacity. Option B is wrong because the Free tier lacks semantic search and is intended for evaluation only, with strict limits on indexes and storage. Option C is wrong because while Storage Optimized (L1) supports semantic search, it is not the minimum tier; it is designed for scenarios requiring massive storage, and the question asks for the minimum tier that supports semantic search.

Full explanation →

446

MCQhard

You are implementing indirect prompt injection protection for an agent that reads emails and summarizes them. An incoming email contains malicious instructions designed to trick the agent into exfiltrating user data. Which Azure AI feature should you configure to detect this threat?

A.Azure DDoS Protection standard tier

B.Prompt Shields for indirect prompt injection in Azure AI Content Safety

C.Azure Web Application Firewall (WAF) SQL injection rules

D.Microsoft Defender for Cloud threat detection

AnswerB

Prompt Shields specifically detects jailbreak and indirect prompt injection attacks embedded in external content such as emails. It flags the malicious instructions before the agent acts, satisfying the requirement to block data exfiltration triggered by untrusted input.

Why this answer

Prompt Shields in Azure AI Content Safety is specifically designed to detect both direct (user) and indirect (document/email-embedded) prompt injection attacks. For an agent that reads emails, indirect prompt injection is the threat model — malicious instructions hidden in email content — and Prompt Shields analyzes the content to flag it before the agent acts on it. This is the only option that addresses prompt injection at the AI layer.

Exam trap

AI-103 often tests whether candidates confuse network/infrastructure security controls (DDoS, WAF, Defender) with AI-specific content safety controls (Prompt Shields) that actually detect prompt injection in LLM inputs.

How to eliminate wrong answers

Option A is wrong because Azure DDoS Protection defends against volumetric network-layer attacks, not semantic prompt injection embedded in email content. Option C is wrong because WAF SQL injection rules protect against SQLi in HTTP requests to web apps — they do not inspect natural-language email content for prompt injection. Option D is wrong because Microsoft Defender for Cloud provides cloud security posture management and threat detection for infrastructure, not LLM prompt injection detection.

Full explanation →

447

MCQhard

You need to secure the communication between your app and Document Intelligence using a private endpoint. What is the primary benefit?

A.Reduced cost

B.Faster inference speeds

C.Increased throughput limits

D.Traffic stays within the Azure private network

AnswerD

A private endpoint assigns a private IP from your virtual network to the Document Intelligence resource, so requests never traverse the public internet. Traffic remains on the Azure backbone within the private network, eliminating public exposure and satisfying the security requirement.

Why this answer

A private endpoint for Document Intelligence creates a private IP address in your virtual network via Azure Private Link, so all traffic between your app and the Document Intelligence resource traverses the Microsoft backbone and never the public internet. The primary benefit is network isolation — traffic stays within the Azure private network, reducing exposure to internet-based threats and satisfying compliance requirements for data-in-transit isolation.

Exam trap

AI-103 often tests whether candidates confuse the security benefit of private endpoints (network isolation, no public internet exposure) with performance or cost benefits that private endpoints do not provide.

How to eliminate wrong answers

Option A is wrong because private endpoints do not reduce cost — in fact, Private Link endpoints and the associated data processing can add cost. Option B is wrong because private endpoints do not inherently make inference faster; latency may even increase slightly due to private DNS resolution and routing. Option C is wrong because private endpoints do not raise throughput or quota limits — those are governed by the pricing tier and service quotas, not by network path.

Full explanation →

448

MCQhard

You are managing an Azure AI Custom Vision project and need to automate model training and evaluation updates as new training images are tagged. Which REST API or SDK component should you use to initiate the training iteration process?

A.Azure Machine Learning Pipeline REST endpoint

B.Azure AI Vision Prediction API

C.Azure AI Content Safety moderation batch API

D.Custom Vision Training API TrainProject method

AnswerD

The TrainProject method in the Custom Vision Training API creates a new iteration from the current tagged images, satisfying the requirement to initiate training automatically as tagging progresses. It returns the iteration ID needed for subsequent evaluation and publishing steps.

Why this answer

The Custom Vision Training API provides the TrainProject method, which initiates a training iteration for a specified project. This is the correct API to automate model training when new tagged images are added. It allows you to programmatically trigger training and then evaluate the resulting iteration.

Exam trap

AI-103 often tests the confusion between training and prediction APIs, or between Custom Vision and Azure Machine Learning services, leading candidates to pick Azure ML pipelines for Custom Vision tasks.

How to eliminate wrong answers

Option A is wrong because Azure Machine Learning Pipeline REST endpoint is used for orchestrating ML workflows in Azure ML, not for Custom Vision training. Option B is wrong because the Azure AI Vision Prediction API is for making predictions with a trained model, not for training. Option C is wrong because Azure AI Content Safety moderation batch API is for content moderation, unrelated to Custom Vision training.

Full explanation →

449

Multi-Selecthard

You are auditing a generative AI application for responsible AI compliance. Which THREE practices should you implement to mitigate risks? (Choose 3)

Select 3 answers

A.Store all user PII in plain text logs

B.Increase the temperature of the model to 1.0

C.Configure system messages to explicitly define boundaries

D.Enable Azure AI Content Safety filtering

E.Implement manual review cycles for high-impact outputs

AnswersC, D, E

System messages set behavioural guardrails by instructing the model to refuse out-of-scope requests, which directly mitigates harmful or off-topic outputs. This satisfies the responsible AI requirement to constrain generated content at inference time, before responses reach users, without retraining or additional infrastructure.

Why this answer

Option C is correct because configuring system messages to explicitly define boundaries constrains the model's behavior, scope, and tone, helping prevent prompt injection, off-topic responses, and harmful or non-compliant outputs. Option D is correct because enabling Azure AI Content Safety filtering applies classifiers for hate, violence, sexual, and self-harm content (plus prompt shields and groundedness detection), directly reducing the risk of harmful or unsafe generated content. Option E is correct because implementing manual review cycles for high-impact outputs adds human oversight, which is a core responsible AI requirement for consequential decisions where automated errors could cause real harm.

Option A is not appropriate because storing user PII in plain text logs violates privacy and data-protection principles and increases breach exposure rather than mitigating risk. Option B is not appropriate because raising the temperature to 1.0 increases randomness and creativity, which raises the likelihood of hallucinated, inconsistent, or unsafe outputs instead of reducing risk.

Exam trap

AI-103 often tests whether candidates confuse model-tuning parameters (temperature, max_tokens) with governance controls, so the trap is picking a sampling knob as if it were a responsible AI safeguard.

Full explanation →

450

MCQmedium

Your development team needs to deploy an Azure OpenAI model deployment in an enterprise environment. They require provisioned throughput units (PTUs) to guarantee predictable latency and throughput for a mission-critical chatbot. Where can you purchase and manage PTU allocations?

A.Azure Portal or Azure OpenAI Studio during model deployment creation by selecting the Provisioned tier.

B.Azure DevOps pipeline YAML configuration files.

C.Microsoft 365 Admin Center under billing subscriptions.

D.Visual Studio Enterprise subscription benefit portal.

AnswerA

Provisioned throughput units are purchased and managed directly through the Azure Portal or Azure OpenAI Studio when creating the deployment, selecting the Provisioned tier. This satisfies the requirement for guaranteed, predictable latency and throughput by reserving dedicated model capacity rather than consuming shared pay-as-you-go quota.

Why this answer

Provisioned Throughput Units (PTUs) for Azure OpenAI are purchased and managed through the Azure Portal or Azure OpenAI Studio when creating or updating a model deployment by selecting the Provisioned deployment type. This is where you specify the PTU quantity, associate it with the Azure OpenAI resource, and manage capacity reservations for predictable latency and throughput.

Exam trap

AI-103 often tests whether candidates know that PTU purchase and management is an Azure resource-level operation performed in the Portal or Azure OpenAI Studio, not a licensing or DevOps pipeline activity.

How to eliminate wrong answers

Option B is wrong because Azure DevOps pipelines are CI/CD orchestration tools; they can deploy infrastructure via IaC but are not a marketplace or management surface for purchasing PTU capacity. Option C is wrong because the Microsoft 365 Admin Center manages Microsoft 365 licenses and subscriptions, not Azure OpenAI provisioned throughput, which is an Azure resource-level commitment. Option D is wrong because Visual Studio Enterprise subscription benefits cover developer tooling and Azure credits, not PTU allocation or management for Azure OpenAI deployments.

Full explanation →

Microsoft Certified: Azure AI Apps and Agents Developer Associate (AI-103) (AI-103) (AI-103) — Questions 376–450