Microsoft Certified: Azure AI Apps and Agents Developer Associate (AI-103) (AI-103) AI-103 Questions 301–375 | Page 5/7

301

MCQhard

You are configuring diagnostic settings for an Azure AI Search service to audit administrative actions and query logs. Which destination should you choose if you want to stream these logs into a SIEM tool like Microsoft Sentinel for real-time security analysis?

A.Send to an Azure DevOps pipeline

B.Export to a local CSV file via Azure CLI

C.Send to an Event Hub

D.Archive to a storage account

AnswerC

Event Hubs provides the streaming ingestion endpoint that Microsoft Sentinel consumes natively, so diagnostic logs flow into the SIEM in near real time. Log Analytics and Storage would require extra connectors or polling, failing the real-time security analysis constraint.

Why this answer

Azure AI Search diagnostic settings support three export destinations: Log Analytics workspace, Azure Storage account, and Azure Event Hubs. Event Hubs is the correct choice for streaming telemetry in near real-time to external SIEM platforms such as Microsoft Sentinel, Splunk, or QRadar, because Sentinel natively ingests from Event Hubs via its data connectors. This enables continuous, low-latency forwarding of administrative and query logs for correlation and alerting.

Exam trap

AI-103 often tests the confusion between storage accounts (for archival/retention) and Event Hubs (for real-time streaming to SIEM), so candidates who see 'audit logs' and reflexively pick storage miss the 'real-time' keyword.

How to eliminate wrong answers

Option A is wrong because Azure DevOps pipelines are CI/CD orchestration tools, not log ingestion endpoints, and cannot receive diagnostic settings streams. Option B is wrong because exporting to a local CSV via Azure CLI is a manual, pull-based operation with no real-time streaming capability and no native SIEM integration. Option D is wrong because archiving to a storage account is designed for long-term retention and compliance, not for real-time streaming; storage is a batch/query destination, not a push-based event pipeline.

Full explanation →

302

MCQhard

You are building a custom named entity recognition model in Language Studio. You want to extract product serial numbers that follow a specific pattern (two uppercase letters followed by four digits, a hyphen, and three digits). Which entity type definition method should you use?

A.List entity type

B.Learned entity type

C.Regex entity type

D.Prebuilt entity type

AnswerC

Regex entity types match text against a defined regular expression pattern, so the serial number format of two uppercase letters, four digits, a hyphen and three digits can be captured precisely. List and prebuilt entity types cannot enforce this positional character structure.

Why this answer

A regex entity type is designed to extract entities that match a specific regular expression pattern, such as two uppercase letters followed by four digits, a hyphen, and three digits (e.g., '[A-Z]{2}\d{4}-\d{3}'). In Language Studio, you define the regex pattern and optionally provide examples, making it ideal for structured identifiers like serial numbers.

Exam trap

AI-103 often tests the misconception that learned entities can handle strict patterns without training data, or that list entities can match patterns; the key is that regex is for pattern-based extraction.

How to eliminate wrong answers

Option A (List entity type) is wrong because a list entity matches a fixed set of synonyms or values (e.g., a list of product names), not arbitrary patterns. Option B (Learned entity type) is wrong because learned entities require labeled training data and are suited for context-dependent extraction, not strict pattern matching. Option D (Prebuilt entity type) is wrong because prebuilt entities cover common domains like datetime, email, or URL, not custom serial number formats.

Full explanation →

303

MCQmedium

Your application processes receipt images using Azure AI Vision Read API. You need to extract text specifically from tables within the receipts. Which service is better suited for advanced table structure and key-value pair extraction compared to base Azure AI Vision Read?

A.Azure AI Face Service

B.Azure AI Document Intelligence

C.Custom Vision Service

D.Azure AI Content Safety

AnswerB

Document Intelligence provides dedicated layout and prebuilt receipt models that output structured tables and key-value pairs. Base Read returns plain text lines and words, so it cannot satisfy the requirement for advanced table structure extraction from receipts.

Why this answer

Azure AI Document Intelligence (formerly Form Recognizer) is purpose-built for structured document extraction, including tables, key-value pairs, and layout analysis. Its prebuilt receipt model and general document/layout models return structured JSON with table cells, row/column spans, and semantic key-value associations that the base Read API does not provide. The Read API only returns raw text lines and words with bounding boxes, without table structure or field semantics.

Exam trap

AI-103 often tests the distinction between OCR-only services (Read API) and structured document understanding services (Document Intelligence), causing candidates to assume any Azure AI Vision offering can extract tables.

How to eliminate wrong answers

Option A is wrong because Azure AI Face Service performs facial detection, recognition, and attribute analysis (age, emotion, landmarks) and has no OCR or document-parsing capability. Option C is wrong because Custom Vision Service is an image classification and object detection service for training custom models on labeled images, not a text or table extraction engine. Option D is wrong because Azure AI Content Safety detects harmful content categories (hate, violence, self-harm, sexual) in text and images and does not perform OCR or document structure extraction.

Full explanation →

304

MCQmedium

You are analyzing product reviews and need to identify specific product features mentioned by users along with whether users expressed positive or negative opinions about those specific features. Which advanced feature of sentiment analysis should you enable?

A.Extractive summarization

B.Opinion mining

C.Key phrase extraction

D.Named Entity Recognition

AnswerB

Opinion mining (aspect-based sentiment analysis) extracts specific product features mentioned in reviews and assigns a separate positive or negative polarity to each. This directly satisfies the need to link opinions to the particular features users discuss.

Why this answer

Opinion mining, also called aspect-based sentiment analysis, is the advanced sentiment analysis feature that identifies specific aspects (product features) mentioned in text and determines the sentiment expressed toward each one. It returns aspect-sentiment pairs, which is exactly what is needed to know whether users felt positively or negatively about particular product features. This makes opinion mining the correct choice.

Exam trap

AI-103 often tests the difference between sentiment analysis variants, and the trap is choosing key phrase extraction because it identifies features, when only opinion mining links each feature to its sentiment polarity.

How to eliminate wrong answers

Option A (Extractive summarization) is wrong because it produces a condensed summary of the text, not aspect-level sentiment pairs. Option C (Key phrase extraction) is wrong because it extracts important phrases but does not associate sentiment polarity with each phrase. Option D (Named Entity Recognition) is wrong because NER identifies entities like people, places, and organizations, not the sentiment expressed toward product features.

Full explanation →

305

MCQeasy

You need to determine the primary dominant colors and whether an image is predominantly black-and-white using Azure AI Vision. Which visual feature parameter provides this information?

A.features=denseCaptions

B.features=color

C.features=people

D.features=tags

AnswerB

The color visual feature returns dominant foreground and background colours plus an isBwImg boolean indicating whether the image is predominantly black-and-white. Requesting features=color therefore supplies exactly the dominant-colour and monochrome detection the stem requires from Azure AI Vision.

Why this answer

The features=color parameter in Azure AI Vision returns color-related analysis, including the dominant foreground and background colors and whether the image is black-and-white (isBWImg). This directly provides the dominant colors and monochrome determination requested. Other features return captions, people detection, or tags, not color metadata.

Exam trap

AI-103 often tests feature parameter selection, and candidates confuse tags or captions with color analysis, forgetting that color is a distinct visual feature with its own parameter.

How to eliminate wrong answers

Option A is wrong because features=denseCaptions generates captions for multiple regions of an image, not color information. Option C is wrong because features=people detects human figures and does not report dominant colors or black-and-white status. Option D is wrong because features=tags returns content tags describing objects and scenes, not color analysis.

Full explanation →

306

MCQmedium

Your company requires strict data privacy. Where should you host your Document Intelligence resources to ensure data remains in a specific region?

A.Regional Azure resource

B.Any available region

C.Global service endpoint

D.Public multi-region endpoint

AnswerA

A regional Azure resource pins Document Intelligence processing and storage to a single chosen geography, satisfying the strict data-residency constraint. Unlike multi-region or global endpoints, which may replicate data across boundaries, regional deployment guarantees that documents and extracted results never leave the specified region, meeting the company's privacy requirement.

Why this answer

To guarantee that Document Intelligence data (documents, extracted results, and processing) stays within a specific geography, you must deploy a regional Azure resource in the desired region. Regional resources pin both compute and data residency to that region, satisfying strict data-privacy and sovereignty requirements. Global or multi-region endpoints may replicate or route data outside the chosen geography, violating the requirement.

Exam trap

AI-103 often tests the misconception that 'any Azure region' or a 'global endpoint' satisfies data residency, when only a region-specific resource guarantees data stays in the required geography.

How to eliminate wrong answers

Option B (Any available region) is wrong because it provides no data-residency guarantee — Azure may place or process data in any region, breaching the privacy requirement. Option C (Global service endpoint) is wrong because global endpoints can route requests and store data across multiple regions, which is incompatible with strict regional data residency. Option D (Public multi-region endpoint) is wrong because multi-region endpoints inherently distribute data across regions and expose a public endpoint, both of which violate the requirement to keep data in a specific region.

Full explanation →

307

MCQhard

You have an Azure AI application that is hitting rate limits (429 errors). You need to determine which specific client application is causing the issue. How do you identify this?

A.Check the 'Activity Log'.

B.Increase the resource quota in the 'Quotas' blade.

C.Enable diagnostic logging to Log Analytics and run a KQL query.

D.Restart the AI service.

AnswerC

Diagnostic logging streams per-request records into Log Analytics, where KQL can filter and aggregate 429 responses by caller identity, such as the client application ID. This satisfies the stem's requirement to pinpoint which specific client application triggers the rate limiting, rather than merely confirming that throttling occurs.

Why this answer

Diagnostic logging to Log Analytics captures per-request telemetry including the caller identity, subscription, resource, operation name, and HTTP status code, so a KQL query filtering on ResponseCode == 429 and summarizing by CallerIpAddress or identity reveals which client is exhausting the quota. The Activity Log only records control-plane operations (resource create/update/delete), not data-plane inference calls, so it cannot show which app is making the requests.

Exam trap

AI-103 often tests whether candidates confuse the Activity Log (control plane) with diagnostic logs (data plane), so the trap is picking Activity Log because it sounds like the place to 'see what happened' when only diagnostic logs contain per-request caller data.

How to eliminate wrong answers

Option A is wrong because the Activity Log records ARM control-plane events such as deployments and role assignments, not individual inference requests, so it has no visibility into which client triggered 429s. Option B is wrong because raising the quota masks the symptom rather than identifying the offending client, and it may not even be permitted if the subscription is already at the regional ceiling. Option D is wrong because restarting the AI service does not diagnose anything; it briefly clears state and may even worsen throttling if clients retry aggressively.

Full explanation →

308

MCQhard

You are building an agent in Azure AI Foundry using a prompt flow. You need to ensure that the agent only accesses specific internal documents and avoids hallucinating outside the provided context. Which grounding technique should you implement?

A.Implement grounding with data using Azure AI Search

B.Use a system message to forbid outside information

C.Enable prompt caching

D.Increase the temperature parameter to 1.0

AnswerA

Azure AI Search indexes the specific internal documents and returns retrieved chunks as grounding context, constraining the model's answer to that content. This satisfies the stem's constraint of restricting access to specific documents and preventing hallucination beyond the provided context, unlike relying on the model's parametric knowledge alone.

Why this answer

Grounding with data using Azure AI Search connects the agent to a specific indexed document set, so responses are generated only from retrieved content. This directly prevents hallucination by constraining the model's context to the internal documents. System messages alone cannot guarantee the model will ignore its pretrained knowledge.

Exam trap

AI-103 often tests the misconception that a system message or prompt instruction can reliably prevent hallucination, when in fact only grounding with a trusted data source enforces factual boundaries.

How to eliminate wrong answers

Option B is wrong because a system message is a soft instruction that the model may ignore, especially with ambiguous prompts; it does not enforce data boundaries. Option C is wrong because prompt caching improves latency and cost, not grounding or factual accuracy. Option D is wrong because increasing temperature to 1.0 makes output more random and creative, worsening hallucination.

Full explanation →

309

MCQeasy

You are using Azure AI Document Intelligence to extract information from identity documents, such as passports and driver licenses. Which prebuilt model should you select?

A.prebuilt-tax.us.w2

B.prebuilt-idDocument

C.prebuilt-W2

D.prebuilt-businessCard

AnswerB

The prebuilt-idDocument model is trained specifically for passports, driver licences and identity cards, extracting fields such as name, date of birth and document number. General or custom models lack this document-type-specific field schema, so they return weaker results for identity documents.

Why this answer

The prebuilt-idDocument model is specifically designed to extract key information from identity documents such as passports, driver licenses, and U.S. state IDs. It returns structured fields like name, date of birth, document number, and expiration date. This is the only prebuilt model among the options that targets identity documents.

Exam trap

AI-103 often tests the ability to distinguish between prebuilt models for different document types; candidates may confuse the W-2 model with identity documents or select an invalid model name like prebuilt-W2.

How to eliminate wrong answers

Option A is wrong because prebuilt-tax.us.w2 is for extracting data from U.S. W-2 tax forms, not identity documents. Option C is wrong because prebuilt-W2 is not a valid Azure AI Document Intelligence model name; the correct model for W-2 forms is prebuilt-tax.us.w2.

Option D is wrong because prebuilt-businessCard is for extracting information from business cards, not passports or driver licenses.

Full explanation →

310

MCQeasy

A retail analytics team wants to count how many people enter a store each hour. They plan to analyze frames from an existing ceiling-mounted camera feed using Azure AI Vision Image Analysis 4.0. They need the count of detected people and the bounding box for each person, without storing the images long term. Which approach should they use?

A.Call the Image Analysis 4.0 analyze endpoint with the features parameter set to people, and read the peopleResult object from the response

B.Call the Image Analysis 4.0 analyze endpoint with the features parameter set to tags, and count tags named person

C.Call the Image Analysis 4.0 analyze endpoint with the features parameter set to smartCrops, and use the crop regions as person locations

D.Call the Image Analysis 4.0 analyze endpoint with the features parameter set to objects, and filter the results for the person tag

AnswerA

The people feature returns a peopleResult containing the count and bounding boxes for detected persons. This directly provides the hourly count and per-person coordinates the team needs, and because analysis is stateless, no images need to be retained after the API response is processed.

Why this answer

The people feature is purpose-built for detecting humans and returns both a count and bounding boxes in peopleResult. That maps exactly to the hourly footfall metric and the per-person coordinates. The other features describe objects, tags, or suggested crops and do not provide dependable person-level detection or counting.

Exam trap

The trap here is confusing general object detection with the dedicated people feature that returns a count and person bounding boxes.

Full explanation →

311

MCQeasy

A developer on your team is configuring an Azure AI Foundry agent that must answer questions about a product catalog stored in Azure Cosmos DB for NoSQL. The agent must retrieve the current price and stock for a given product SKU at query time. Which tool should the agent use?

A.The code interpreter tool to run a Python script that connects to Cosmos DB and computes the price.

B.A function tool that queries Cosmos DB for the SKU and returns the current price and stock as JSON.

C.The file search tool with a CSV export of the catalog uploaded to the agent's vector store.

D.The Azure AI Search tool pointed at an index populated nightly from Cosmos DB.

AnswerB

A function tool lets the agent execute deterministic code that reads the authoritative Cosmos DB record at request time, returning current price and stock. Structured lookups by key are exactly the case function calling is designed for, and the returned JSON can be inserted into the model's context for a grounded answer.

Why this answer

A function tool is the correct choice because it allows the agent to execute authenticated code that queries Cosmos DB for the exact SKU and returns live price and stock values. Function calling is intended for deterministic, structured lookups where a vector index or static file would return stale data.

Exam trap

The trap here is reaching for a retrieval tool by habit when the requirement is a live, exact-key lookup that only a function tool can satisfy.

Full explanation →

312

Multi-Selecthard

You are configuring network isolation and security for an Azure AI Search service that connects to an Azure SQL Database backend. Which THREE features can you utilize to secure this data flow? (Choose three.)

Select 3 answers

A.IP firewall rules restricting Azure SQL Database access to the Azure AI Search service outbound IPs.

B.Azure Private Endpoint for Azure AI Search and Azure SQL Database

C.Managed identities and secure connection strings with Entra authentication where supported.

D.Disabling all transport layer security (TLS) requirements to speed up query execution.

E.Storing plain-text database admin credentials in the search index configuration file.

AnswersA, B, C

IP firewall rules on Azure SQL Database whitelist the Azure AI Search service's outbound IP addresses, permitting indexer connections while blocking other internet sources. This satisfies the restriction requirement at the database server level, though it operates over the public endpoint rather than a private one.

Why this answer

Option A is correct because Azure SQL Database supports server-level and database-level IP firewall rules, and you can allow the specific outbound IP addresses of the Azure AI Search service (its public egress IPs) so only that traffic reaches the SQL server. Option B is correct because Azure Private Endpoint creates a private IP address in your VNet for both Azure AI Search and Azure SQL Database, so the data flow stays on the Microsoft backbone/private network instead of traversing the public internet, and it can be combined with disabling public access. Option C is correct because Azure AI Search indexers can use a managed identity (system-assigned or user-assigned) to authenticate to Azure SQL Database with Microsoft Entra ID, and secure connection strings avoid embedding credentials in code or configuration.

Option D is wrong because disabling TLS would remove encryption in transit and weaken security, not legitimately secure the data flow, and TLS does not meaningfully slow query execution in this context. Option E is wrong because storing plain-text admin credentials in a search index configuration file exposes secrets and is a security anti-pattern, not a supported or recommended way to secure the connection.

Exam trap

AI-103 often tests the misconception that disabling TLS or embedding credentials improves performance or simplicity — both are security failures, not valid isolation features.

Full explanation →

313

MCQhard

You are managing an Azure AI Search service with multiple replicas and partitions. During peak query hours, latency increases significantly, and CPU utilization on the search service hits 95%. You need to scale the service to improve query throughput without altering the index data schema. Which scaling action should you take?

A.Increase the number of search partitions.

B.Disable semantic ranking to reduce compute overhead.

C.Increase the number of search replicas in the Azure portal or via CLI.

D.Migrate the service to the Free pricing tier.

AnswerC

Adding replicas distributes query load across more instances, directly relieving the 95% CPU saturation during peak hours and raising query throughput. Replicas serve queries only; partitions store and index data, so adding replicas leaves the index schema untouched, satisfying the stem's constraint.

Why this answer

Adding replicas increases the number of copies of the index that can serve queries, directly improving query throughput and distributing CPU load across more nodes. Since the issue is high query CPU utilization and latency during peak hours, scaling replicas is the correct horizontal scale-out for read/query performance.

Exam trap

AI-103 often tests the replica-vs-partition distinction, so candidates pick partitions for a query-throughput problem when partitions actually address storage and indexing scale, not query concurrency.

How to eliminate wrong answers

Option A is wrong because partitions scale storage and write/indexing throughput, not query concurrency; adding partitions can even increase query fan-out cost. Option B is wrong because disabling semantic ranking reduces functionality rather than scaling capacity, and it does not address the underlying CPU saturation. Option D is wrong because the Free tier has severe limits (no scaling, low storage) and would worsen the problem.

Full explanation →

314

MCQhard

You are building an agentic workflow in Azure AI Foundry. You need the agent to perform a multi-step task: search a database, summarize results, and then send an email. Which component is best suited to manage this orchestration?

A.Azure AI Agent Service.

B.Azure OpenAI fine-tuning service.

C.Azure AI Search indexer.

D.Azure AI Content Safety API.

AnswerA

Azure AI Agent Service provides built-in orchestration for multi-step tool use, chaining database search, summarisation and email sending through its thread and tool-calling runtime. It satisfies the requirement to manage sequential task orchestration without custom glue code.

Why this answer

Azure AI Agent Service is purpose-built to orchestrate multi-step agentic workflows, including tool/function calling, planning, and chaining actions like database queries, summarization, and email dispatch. It manages conversation state, tool invocation, and step sequencing natively, which is exactly what this scenario requires.

Exam trap

AI-103 often tests whether candidates confuse model customization (fine-tuning) or data retrieval (AI Search) with orchestration, so they pick a component that handles only one sub-task instead of the agent runtime that manages the whole multi-step flow.

How to eliminate wrong answers

Option B is wrong because fine-tuning only customizes a model's behavior on training data; it does not orchestrate tools or multi-step actions. Option C is wrong because an Azure AI Search indexer ingests and enriches data into a search index; it cannot plan or execute workflows. Option D is wrong because Content Safety only classifies and filters harmful content, providing no orchestration capability.

Full explanation →

315

MCQhard

You are processing customer feedback using sentiment analysis in the Azure AI Language service. You need to extract finer-grained information about specific product attributes mentioned in the reviews (e.g., finding out that the screen is bright but the battery life is poor). Which feature should you enable in your API request?

A.Opinion mining

B.PII detection

C.Language detection

D.Custom text classification

AnswerA

Opinion mining extends sentiment analysis by linking each detected sentiment to the specific target it describes, so 'screen' is scored positive while 'battery life' is scored negative. Plain sentiment analysis returns only document or sentence-level polarity.

Why this answer

Opinion mining in Azure AI Language sentiment analysis extracts fine-grained opinions about specific aspects or attributes mentioned in text, such as 'screen' and 'battery life'. It provides sentiment for each aspect, allowing you to determine that the screen is bright (positive) but battery life is poor (negative). This is exactly what the scenario requires.

Exam trap

AI-103 often tests the difference between general sentiment analysis and opinion mining; candidates may overlook that opinion mining must be explicitly enabled in the API request to get aspect-level details.

How to eliminate wrong answers

Option B is wrong because PII detection identifies and redacts personal information, not opinions about product attributes. Option C is wrong because language detection identifies the language of the text, not aspect-based sentiment. Option D is wrong because custom text classification categorizes text into custom labels, but does not extract aspect-level sentiment.

Full explanation →

316

Multi-Selecthard

You are optimizing an enterprise RAG application in Azure AI Foundry for high accuracy and minimal latency. Which THREE advanced retrieval techniques should you consider implementing? (Choose THREE)

Select 3 answers

A.Semantic ranking to re-rank top search results using deep learning models

B.Using random sentence selection without indexing

C.Hybrid search combining full-text BM25 and vector embeddings

D.Disabling vector search entirely to rely solely on random string matching

E.Query expansion and re-writing to enrich user prompts with synonyms and related terms

AnswersA, C, E

Semantic ranking applies deep-learning re-ranking to the top results returned by the initial retrieval pass, lifting relevance precision beyond keyword or vector similarity alone. This directly serves the stem's high-accuracy constraint while capping latency, since re-ranking runs only over a small candidate set rather than the full index.

Why this answer

Option A (Semantic ranking) is correct because Azure AI Search's semantic ranker applies deep-learning language models to re-rank the top BM25/vector results, promoting the most contextually relevant passages and directly improving RAG answer accuracy. Option C (Hybrid search combining full-text BM25 and vector embeddings) is correct because fusing keyword and dense-vector retrieval via Reciprocal Rank Fusion captures both exact term matches and semantic similarity, which raises recall and precision over either method alone. Option E (Query expansion and re-writing) is correct because enriching the user prompt with synonyms and related terms broadens the candidate set and aligns the query with the indexed content, improving retrieval quality for ambiguous or short queries.

Option B does not belong because random sentence selection without indexing provides no relevance signal or deterministic retrieval, guaranteeing poor accuracy. Option D does not belong because disabling vector search and relying on random string matching discards semantic retrieval entirely, which destroys both accuracy and the ability to handle natural-language queries.

Exam trap

AI-103 often tests whether candidates confuse retrieval-quality techniques (hybrid, semantic, query expansion) with unrelated or destructive options like disabling vector search or random selection, which sound plausible but break the RAG pipeline.

Full explanation →

317

MCQmedium

You need to ensure that your agent is compliant with regional data residency requirements. Where should you check this?

A.In the Prompt Flow runtime.

B.In the Content Safety policy.

C.In the model configuration settings.

D.In the Azure AI Foundry resource's region settings.

AnswerD

Regional data residency is enforced by the Azure AI Foundry resource's own region, since data is stored and processed there. Checking that resource's region settings confirms whether the deployed agent keeps data within the required geography, directly satisfying the compliance constraint in the stem.

Why this answer

Data residency for Azure AI Foundry is governed by the region of the Azure AI Foundry resource (and its dependent resources such as the model deployment and storage). To verify compliance with regional data residency requirements, you check the resource's region settings, because that determines where data is stored and processed. Content safety policies, prompt flow runtime, and model configuration do not control geographic placement.

Exam trap

AI-103 often tests the confusion between configuration settings that affect behavior (content safety, model parameters, prompt flow) and the resource-level region setting that actually governs data residency.

How to eliminate wrong answers

Option A is wrong because the Prompt Flow runtime is an execution environment for flows and does not determine where data is stored or processed for residency purposes. Option B is wrong because Content Safety policies define filtering thresholds and categories, not geographic data placement. Option C is wrong because model configuration settings control parameters like temperature, max tokens, and model version, not the region in which the resource and its data reside.

Full explanation →

318

MCQhard

You are building an automated asset management system that processes millions of images annually. To optimize costs and latency, you want to minimize the number of API calls while extracting tags, bounding box objects, and OCR text simultaneously. How should you structure your API request?

A.Submit three separate asynchronous requests for tags, objects, and read.

B.Include multiple comma-separated values in the features parameter, such as features=tags,objects,read.

C.Use the legacy Computer Vision v3.2 batch analysis endpoint.

D.Train a Custom Vision model to combine all three tasks.

AnswerB

The features parameter accepts multiple comma-separated capabilities in one analyse call, returning tags, objects and read results together. This collapses three separate requests into one, directly satisfying the constraint to minimise API calls across millions of images.

Why this answer

The Image Analysis 4.0 analyze endpoint accepts a comma-separated features parameter, so a single POST with features=tags,objects,read returns all three result sets in one response. This minimizes billable transactions and network round-trips versus issuing separate calls. The response JSON contains separate sections (tags, objects, readResult) that the client parses independently.

Exam trap

AI-103 often tests the assumption that each AI capability requires its own API call, when the modern Image Analysis endpoint explicitly supports multiple comma-separated features in a single request.

How to eliminate wrong answers

Option A is wrong because three separate asynchronous requests triple the transaction count and latency, defeating the stated goal of minimizing API calls, and the async pattern is unnecessary for a single-image analysis. Option C is wrong because the legacy Computer Vision v3.2 batch endpoint is deprecated and does not offer the unified multi-feature parameter in the same way as the current Image Analysis 4.0 API. Option D is wrong because Custom Vision is a classification/object-detection training service and does not perform OCR or return general-purpose tags; it cannot combine the three tasks.

Full explanation →

319

MCQhard

You are deploying an Azure AI Speech container on an Azure Kubernetes Service (AKS) cluster. Your security policy dictates that pod-to-pod communication within the cluster must be encrypted in transit using mTLS, and egress traffic to Azure must flow through a strict outbound firewall. Which Kubernetes feature should you implement?

A.Azure Container Registry vulnerability scanning

B.Kubernetes Network Policies combined with a service mesh for mTLS

C.Azure Bastion deployed inside the AKS worker node subnet

D.AKS cluster auto-scaler configuration

AnswerB

Network policies govern IP-level pod isolation, while a service mesh enforces mTLS encryption for pod-to-pod communication.

Why this answer

Kubernetes Network Policies enforce pod-level traffic rules (including egress restrictions to a firewall), while a service mesh (e.g., Istio, Linkerd, or Open Service Mesh) provides automatic mTLS for pod-to-pod communication. Together they satisfy both requirements: encrypted in-transit pod traffic and strict egress control. This is the standard AKS pattern for zero-trust pod networking.

Exam trap

AI-103 often tests whether candidates conflate cluster-level security features (image scanning, auto-scaler, Bastion) with pod-level networking controls (Network Policies, service mesh), tricking them into picking a tangential AKS feature.

How to eliminate wrong answers

Option A is wrong because Azure Container Registry vulnerability scanning inspects image layers for CVEs; it has nothing to do with runtime pod-to-pod encryption or egress firewalling. Option C is wrong because Azure Bastion provides secure RDP/SSH access to VMs from the Azure portal — it does not encrypt pod-to-pod traffic or control egress. Option D is wrong because the AKS cluster auto-scaler adjusts node pool size based on demand; it does not address mTLS or egress firewall requirements.

Full explanation →

320

MCQeasy

You need to evaluate customer reviews to determine whether the overall tone is positive, negative, or neutral. Which Azure AI Language feature should you use?

A.Language detection

B.Named Entity Recognition

C.Key phrase extraction

D.Sentiment analysis

AnswerD

Sentiment analysis returns positive, negative, neutral or mixed labels with confidence scores for each document, directly satisfying the requirement to determine overall review tone. It is a prebuilt Azure AI Language feature requiring no training.

Why this answer

Sentiment analysis is the Azure AI Language feature that evaluates text and returns a sentiment label (positive, negative, neutral, or mixed) along with confidence scores. It is specifically designed to determine the overall tone of customer reviews. The other features do not provide sentiment scoring.

Exam trap

AI-103 often tests the distinction between sentiment analysis and other text analytics features, so candidates must remember that only sentiment analysis returns positive/negative/neutral labels.

How to eliminate wrong answers

Option A is wrong because language detection identifies the language of the text, not its tone. Option B is wrong because Named Entity Recognition extracts entities such as people, places, and dates, but does not assess sentiment. Option C is wrong because key phrase extraction identifies main topics, not positive or negative tone.

Full explanation →

321

MCQeasy

When sharing a Prompt Flow with a teammate, which action allows them to run the flow without having to recreate the connections?

A.Delete the connection and recreate it

B.Grant permissions to the shared Connection resource

C.Hardcode credentials into the flow

D.Export the flow as a text file

AnswerB

Prompt Flow connections hold credentials and endpoint configuration, so a teammate needs permission on the shared Connection resource itself. Granting that access lets them run the flow using the existing connection rather than recreating it with their own credentials.

Why this answer

When sharing a Prompt Flow with a teammate, the flow references a Connection resource by name. To let the teammate run the flow without recreating the connection, you grant them permissions (e.g., Cognitive Services User or equivalent RBAC role) on the shared Connection resource in Azure AI Foundry. This way the flow resolves the same connection at runtime for both users, and credentials remain centralized and secure.

Exam trap

The trap is thinking that sharing the flow file itself is sufficient — candidates may pick 'export the flow as a text file' because it sounds like sharing, but the flow only works if the recipient also has RBAC permissions on the referenced Connection resource.

How to eliminate wrong answers

Option A is wrong because deleting and recreating the connection would break the flow's reference and force reconfiguration — it is the opposite of what you want when sharing. Option C is wrong because hardcoding credentials into the flow embeds secrets in the flow artifact, which is insecure and defeats the purpose of using connections; it also means the teammate uses your credentials directly, which is an audit and rotation nightmare. Option D is wrong because exporting the flow as a text file does not transfer connection permissions — the teammate would still lack access to the referenced connection and would have to recreate it, which is exactly what the question says to avoid.

Full explanation →

322

MCQeasy

You are provisioning an Azure AI services account and want to view your current API key. Where can you retrieve the API key in the Azure portal?

A.Resource graph explorer

B.Azure Advisor cost optimization pane

C.Azure Active Directory overview page

D.Keys and Endpoint blade of the resource

AnswerD

The Keys and Endpoint blade exposes both Key1 and Key2 for the Azure AI services resource, letting you copy a key directly without regenerating it. This satisfies the stem's requirement to view the current API key in the Azure portal, since the blade displays existing key values rather than only offering rotation.

Why this answer

In the Azure portal, the API key and endpoint for an Azure AI services (Cognitive Services) resource are found on the 'Keys and Endpoint' blade of that resource. This blade displays Key1 and Key2, the endpoint URL, and the region, and allows regenerating the keys. It is the standard location for retrieving credentials used by SDKs and REST calls.

Exam trap

AI-103 often tests portal navigation knowledge — candidates may overthink and pick a governance or monitoring blade, but the API key is always on the resource's own Keys and Endpoint blade, not in directory or advisor views.

How to eliminate wrong answers

Option A is wrong because Resource Graph Explorer is a tool for querying and exploring Azure resources via Kusto queries — it does not expose service keys. Option B is wrong because Azure Advisor's cost optimization pane provides recommendations to reduce spend, not credentials. Option C is wrong because the Azure Active Directory (Entra ID) overview page manages identities, tenants, and directory settings, not AI service API keys.

Full explanation →

323

Multi-Selecthard

When configuring a custom Named Entity Recognition (NER) project in Language Studio using the Azure AI Language service, which THREE steps or configuration options are required during the model lifecycle?

Select 3 answers

A.Define your entity schema and types

B.Fine-tune the underlying base transformer weights via direct GPU gradient backpropagation code

C.Manually convert all input text files into pre-tokenized JSON vectors

D.Train the model using your labeled dataset

E.Deploy the trained model to a deployment name for client application consumption

AnswersA, D, E

Defining the entity schema and types is mandatory because the NER model learns to classify tokens against exactly the categories you declare; without labelled entity types, training data cannot be mapped to outputs. This satisfies the stem's requirement for a required model-lifecycle configuration step in Language Studio.

Why this answer

Option A is correct because a custom NER project requires you to define the entity schema — the entity types (labels) and their descriptions — in Language Studio before you can label data, as this schema drives the labeling experience and the model's output classes. Option D is correct because after labeling your documents you must start a training job in Language Studio (or via the REST API/az CLI) that builds the model from your labeled dataset; the service handles the underlying training without you writing training code. Option E is correct because a trained custom NER model is not queryable until it is assigned to a deployment (a named deployment in the Language resource), which the client application then calls via the analyze-text endpoint.

Option B is not required because Azure AI Language abstracts the base model and performs fine-tuning internally; you do not write GPU gradient backpropagation code. Option C is not required because Language Studio accepts plain-text/JSON documents in its supported formats and performs tokenization internally, so manual pre-tokenized JSON vectors are unnecessary.

Exam trap

AI-103 often tests the misconception that you need to manually handle tokenization or fine-tune weights, when the service automates these steps.

Full explanation →

324

MCQmedium

You are using Prompt Flow to iterate on an agent. What is the benefit of using 'Environment Variables' in your flow?

A.To manage configuration settings like API keys and endpoints across environments.

B.To reduce the model response time.

C.To store large data files.

D.To increase the prompt length.

AnswerA

Environment variables hold configuration such as API keys and endpoints outside the flow definition, so the same flow can be promoted between development, test and production without editing nodes. This satisfies the stem's need to manage settings across environments.

Why this answer

Environment variables in Prompt Flow let you externalize configuration such as API keys, endpoints, and deployment names so the same flow can run against different environments (dev, test, prod) without code changes. This is the standard twelve-factor app pattern and is the primary benefit of using environment variables in a flow.

Exam trap

The trap is conflating environment variables with performance or data-storage features — candidates may pick 'reduce response time' or 'store large data files' because they associate configuration with optimization, but environment variables are strictly for externalized configuration and secrets management.

How to eliminate wrong answers

Option B is wrong because environment variables do not affect model inference latency — response time is determined by model size, prompt length, and service load, not by how configuration is stored. Option C is wrong because environment variables are meant for small configuration values, not large data files; storing large files in environment variables would hit size limits and is an anti-pattern. Option D is wrong because environment variables have no effect on prompt length — prompt length is determined by the text you send to the model, not by configuration storage.

Full explanation →

325

MCQeasy

You are building an application that analyzes user feedback. You need to detect instances where users share their phone numbers, email addresses, and credit card numbers so that you can automatically redact them. Which feature should you use?

A.Sentiment analysis

B.Language detection

C.Named Entity Recognition (NER)

D.PII detection

AnswerD

PII detection identifies and returns entities such as phone numbers, email addresses and credit card numbers, with optional redaction masking in the response. This directly satisfies the requirement to detect and automatically redact those specific data types from user feedback.

Why this answer

PII detection in Azure AI Language is specifically designed to identify and redact personally identifiable information such as phone numbers, email addresses, and credit card numbers. It returns the detected entities along with their categories and confidence scores, enabling automatic redaction. This is the exact feature needed for the scenario.

Exam trap

AI-103 often tests the confusion between NER and PII detection; while NER can identify some PII, PII detection is specifically optimized for sensitive data and includes redaction capabilities.

How to eliminate wrong answers

Option A is wrong because sentiment analysis determines the emotional tone of text, not the presence of PII. Option B is wrong because language detection identifies the language of the text, not sensitive information. Option C is wrong because Named Entity Recognition (NER) identifies general entities like people, places, and organizations, but does not specifically target PII categories like phone numbers or credit card numbers with the same precision and redaction focus as PII detection.

Full explanation →

326

Multi-Selectmedium

You are planning the security architecture for an enterprise Azure AI multi-service account. Management requires that access to the resource be strictly controlled and audited. Which TWO security features or configurations should you implement? Each correct answer presents part of the solution.

Select 2 answers

A.Disable local authentication (API key access) and require Microsoft Entra ID authentication.

B.Store all admin API keys permanently in an unencrypted Azure Blob Storage container.

C.Configure an Azure Private Endpoint to restrict data access to the virtual network.

D.Grant the Contributor role to all end-user client applications.

E.Enable anonymous public access across all subnets.

AnswersA, C

Disabling local authentication removes API-key access, forcing every request to authenticate through Microsoft Entra ID. This satisfies the stem's requirement for strictly controlled, auditable access, since Entra ID issues tokens tied to identities that can be logged and governed.

Why this answer

Option A is correct because disabling local authentication on the Azure AI multi-service account removes API-key-based access, forcing all callers to authenticate through Microsoft Entra ID (formerly Azure AD), which enables identity-based access control, conditional access, and full audit logging of who accessed the resource. Option C is correct because configuring an Azure Private Endpoint assigns a private IP address from your virtual network to the Azure AI service, so data-plane traffic stays on the Microsoft backbone and the resource can be restricted from public network access, satisfying strict network-level control and auditability. Option B is wrong because storing admin API keys permanently in an unencrypted Blob container exposes credentials and directly contradicts the requirement for strictly controlled, audited access.

Option D is wrong because granting the Contributor role to all end-user client applications violates least privilege and gives broad management-plane rights rather than scoped data access. Option E is wrong because enabling anonymous public access across all subnets removes authentication and network restrictions entirely, which is the opposite of the required security posture.

Exam trap

The trap is selecting options that sound convenient (storing keys, broad roles) instead of least-privilege and network isolation — candidates must recognize that 'strictly controlled and audited' implies Entra ID auth and Private Endpoint.

Full explanation →

327

MCQhard

You are designing an Azure AI Language solution that redacts PII from scanned insurance claim narratives. Because the PII is stored with strict data-residency requirements, you must call the service from a specific Azure region and confirm which categories are detected. Which response element should you inspect to determine the exact PII category and confidence for each detected span?

A.The 'entities' array, where each entry includes 'category', 'confidenceScore', 'offset', and 'length'.

B.The 'statistics' object, where each entry includes 'documentsCount' and 'transactionsCount'.

C.The 'warnings' array, where each entry includes a 'code' and 'message' for each document.

D.The 'documents' array, where each entry includes a 'redactedText' field only.

AnswerA

PII detection returns an 'entities' array whose entries carry the detected 'category' such as PhoneNumber or CreditCardNumber, a 'confidenceScore', and the 'offset' and 'length' of the span. Inspecting these fields lets you confirm both the exact category and its confidence for each detected span in the claim narrative.

Why this answer

To confirm both the category and confidence of each detected PII span, you must examine the 'entities' array, whose entries pair a 'category' with a 'confidenceScore' and the span's 'offset' and 'length'. The redacted text alone hides that metadata, and the warnings and statistics objects describe document health and counts rather than individual PII detections.

Exam trap

The trap here is treating the redacted narrative as proof of correct detection, when the category and confidence only appear in the entities array.

Full explanation →

328

MCQmedium

Your organization is deploying an Azure OpenAI resource. Your security policy requires that all network traffic destined for the OpenAI service must traverse a secure hub-and-spoke virtual network topology and pass through an Azure Firewall instance. How should you route outbound traffic from your application subnet to Azure OpenAI?

A.Configure Azure Traffic Manager dns routing.

B.Attach an Azure Bastion subnet to the firewall.

C.Rely on automatic default system routing without route tables.

D.Configure User Defined Routes (UDR) on the application subnet with a next hop pointing to the Azure Firewall.

AnswerD

User Defined Routes override Azure's default system routes, forcing the application subnet's outbound traffic to the firewall's private IP as next hop. This satisfies the policy that all traffic to Azure OpenAI traverses the hub-and-spoke topology and the Azure Firewall instance.

Why this answer

User Defined Routes (UDR) let you override Azure's default system routes by associating a custom route table with the application subnet. By setting the next hop type to 'Virtual appliance' and pointing to the Azure Firewall's private IP, all egress traffic (including to the Azure OpenAI endpoint) is forced through the firewall for inspection and policy enforcement. This is the standard hub-and-spoke pattern for centralized network security.

Exam trap

AI-103 often tests whether candidates confuse DNS-based traffic steering (Traffic Manager) or access services (Bastion) with actual packet-path control, which only UDRs and a virtual appliance next hop provide.

How to eliminate wrong answers

Option A is wrong because Azure Traffic Manager is a DNS-based global load balancer and cannot force traffic through a firewall or alter the data path. Option B is wrong because Azure Bastion provides RDP/SSH access to VMs over TLS and has nothing to do with routing application subnet traffic. Option C is wrong because default system routes send internet-bound traffic directly out via Azure's backbone, bypassing the firewall entirely, which violates the security policy.

Full explanation →

329

MCQhard

You are experiencing high latency in your RAG agent. You confirm the model response time is fine, but the retrieval step is slow. What should you investigate?

A.The system prompt length.

B.The content safety threshold.

C.The model temperature.

D.The index retrieval configuration and semantic ranker parameters.

AnswerD

Since generation latency is acceptable, the bottleneck lies in retrieval. Index configuration, chunking, vector search settings and semantic ranker parameters directly govern how long candidate documents take to fetch and rerank, so these are the components to profile.

Why this answer

Since the LLM response time is confirmed fine, the bottleneck is in the retrieval pipeline. Index configuration (chunk size, vector dimensions, HNSW/IVF parameters) and semantic ranker settings (which add an L2 reranking pass) directly determine retrieval latency. Tuning these — e.g., reducing semantic ranker usage or adjusting efSearch/kNN — is the correct investigation path.

Exam trap

The trap here is assuming latency is always a model/prompt issue; AI-103 tests whether you can isolate the retrieval stage and correctly attribute slowness to index and semantic ranker configuration rather than generation parameters.

How to eliminate wrong answers

Option A is wrong because system prompt length affects token processing and generation time on the model side, which the question states is already fine. Option B is wrong because content safety thresholds affect whether content is blocked, not retrieval latency. Option C is wrong because temperature only influences sampling randomness during generation, not retrieval speed.

Full explanation →

330

MCQmedium

You are using Azure AI Vision Read API in your application. You want to ensure that the extracted text blocks return word-level bounding polygons rather than just line-level bounding boxes. What parameter or response structure should you examine?

A.Add ?wordLevel=true to the query string.

B.Examine the 'words' array within each returned 'lines' object in the JSON response payload.

C.Call the legacy OCR endpoint instead.

D.Set the model-version parameter to v2.0.

AnswerB

The Read API JSON nests a 'words' array inside each 'lines' object, and every word entry carries a bounding polygon with four corner coordinates. Inspecting that array satisfies the stem's requirement for word-level polygons rather than the line-level bounding boxes returned alongside.

Why this answer

The Azure AI Vision Read API returns OCR results as a hierarchy of pages → lines → words, and each word object contains a bounding polygon (boundingPolygon) with four (x,y) points. To get word-level polygons, you examine the 'words' array nested inside each 'lines' object in the JSON response — no query parameter is needed.

Exam trap

AI-103 often tests whether candidates know that word-level polygons are part of the default Read response schema, not something enabled by a query parameter or model version.

How to eliminate wrong answers

Option A is wrong because there is no ?wordLevel=true query parameter in the Read API — word-level data is returned by default in the analyze/read response structure. Option C is wrong because the legacy OCR endpoint (vision/v3.x/ocr) is deprecated and returns a different schema; it does not provide the modern Read API word-level polygon structure. Option D is wrong because model-version=v2.0 refers to an older Read model and does not change the response schema to include word-level polygons — the schema already includes them.

Full explanation →

331

Multi-Selecthard

You are troubleshooting a Custom Vision object detection model that exhibits poor performance in production. Which THREE actions can you take using the Custom Vision portal or APIs to diagnose and improve the model? (Choose three.)

Select 3 answers

A.Delete all project iteration history to reset model weights to random initialization.

B.Examine prediction suggestions to find low-confidence or incorrectly predicted images and add them to training data.

C.Review precision and recall metrics per tag in the Performance tab.

D.Adjust the probability threshold on the performance page to evaluate precision/recall trade-offs.

E.Manually edit model neural network weights via the REST API.

AnswersB, C, D

Examining prediction suggestions surfaces production images where confidence scores fall below threshold or labels diverge from ground truth, exposing distribution gaps between training and live data. Adding these misclassified samples to the training set directly addresses the poor production performance constraint by retraining the model on representative edge cases it previously never encountered.

Why this answer

Option B is correct because the Custom Vision portal's Predictions view surfaces images the deployed model scored with low confidence or misclassified, and feeding those hard examples back into the training set as labeled data is the standard active-learning loop for improving an object detection model. Option C is correct because the Performance tab reports per-tag precision, recall, and mean average precision (mAP), letting you identify which object classes the model detects poorly and target them with more or better-labeled images. Option D is correct because the Performance page includes a probability threshold slider that recomputes precision and recall at different confidence cutoffs, so you can pick an operating point that balances false positives and false negatives for your production scenario.

Option A is wrong because Custom Vision does not expose iteration history deletion as a diagnostic or improvement action, and resetting to random weights would destroy the trained model rather than improve it. Option E is wrong because Custom Vision is a managed service that does not expose neural network weights for manual editing through the REST API; you can only influence the model through training data, tags, domains, and training parameters.

Exam trap

AI-103 often tests the misconception that you can directly manipulate model internals (weights, architecture) in Custom Vision, when in reality improvement comes only from data curation, threshold tuning, and retraining.

Full explanation →

332

MCQeasy

Which command line tool is primarily used for managing Azure resources, including AI services, from a terminal?

A.Azure PowerShell.

B.Azure CLI.

C.Docker.

D.Kubectl.

AnswerB

Azure CLI provides cross-platform terminal commands that directly manage Azure resources, including AI services, satisfying the stem's requirement for a command line tool. Unlike portal-based or SDK approaches, it exposes resource provisioning and configuration through scriptable `az` commands, enabling automation of AI service deployment and management from any shell.

Why this answer

Azure CLI is the cross-platform command-line tool designed to create, manage, and automate Azure resources, including Azure AI services, from a terminal. It provides commands such as 'az cognitiveservices account create' and 'az ml' for AI workloads, making it the primary terminal-based management tool.

Exam trap

The trap here is conflating Azure PowerShell with Azure CLI; both manage Azure, but the question asks for the primary terminal tool, and Azure CLI is the cross-platform, shell-agnostic answer.

How to eliminate wrong answers

Option A is wrong because Azure PowerShell is a PowerShell module for managing Azure, not the primary cross-platform CLI tool referenced by the question's 'command line tool' phrasing; it is also shell-specific. Option C is wrong because Docker is a containerization platform, not an Azure resource management tool. Option D is wrong because Kubectl manages Kubernetes clusters, not general Azure resources or AI services.

Full explanation →

333

Multi-Selectmedium

Which TWO of the following are valid ways to authenticate to the Document Intelligence service?

Select 2 answers

A.FTP Credentials

B.SSH Key

C.SQL Server Authentication

D.API Key

E.Microsoft Entra ID (RBAC)

AnswersD, E

API keys authenticate requests to Document Intelligence by passing a shared secret in the Ocp-Apim-Subscription-Key header, satisfying the requirement for a valid service authentication method. Unlike Microsoft Entra ID tokens, keys grant full resource access without role scoping, but they remain a supported credential for the service.

Why this answer

Option D (API Key) is correct because Azure AI Document Intelligence (formerly Form Recognizer) issues subscription keys that clients pass in the Ocp-Apim-Subscription-Key header (or as a query parameter) to authenticate REST/SDK requests. Option E (Microsoft Entra ID with RBAC) is correct because the service supports Entra ID token-based authentication, where a bearer token is obtained for the https://cognitiveservices.azure.com/ scope and access is governed by Azure RBAC roles such as Cognitive Services User. The remaining options do not apply: FTP credentials (A) are for file transfer servers, not Azure AI services; an SSH key (B) is for secure shell login to hosts; and SQL Server Authentication (C) is a database login mechanism unrelated to Document Intelligence's HTTP-based authentication.

Exam trap

AI-103 often tests whether candidates confuse generic IT authentication mechanisms (FTP, SSH, SQL logins) with Azure AI service authentication, which is limited to API keys and Microsoft Entra ID.

Full explanation →

334

MCQhard

An agent is designed to summarize meeting transcripts but keeps including internal meeting notes in the final summary. What is the most effective way to address this?

A.Update the system prompt to clarify which sections of the transcript to include.

B.Lower the temperature.

C.Switch to a different model.

D.Add a 'stop_sequence' to the API call.

AnswerA

System prompts define the agent's scope and boundaries, so explicitly restricting the summary to transcript sections excludes internal notes. This is a prompt-engineering fix applied before retrieval or post-processing, directly addressing the unwanted content at generation time.

Why this answer

The system prompt is the primary control surface for defining an agent's scope, tone, and what content it should include or exclude. When an agent incorrectly includes internal meeting notes, the root cause is usually ambiguous instructions about which transcript sections are in-scope, so clarifying the inclusion/exclusion rules in the system prompt directly fixes the behavior. This is a prompt engineering fix, not a model parameter fix, because the issue is about instruction-following scope rather than randomness or creativity.

Exam trap

AI-103 often tests the misconception that model parameters like temperature or stop sequences can fix content-scope problems, when in fact scope and inclusion rules belong in the system prompt.

How to eliminate wrong answers

Option B is wrong because lowering temperature only reduces randomness in token sampling and does not change what content the model considers in-scope; the agent would still include internal notes, just more deterministically. Option C is wrong because switching models does not address the underlying ambiguity in the instructions and may introduce new behavioral differences without solving the scope problem. Option D is wrong because a stop_sequence only halts generation when a specific string is emitted; it cannot selectively filter which transcript sections are summarized and would truncate output rather than correct inclusion logic.

Full explanation →

335

MCQhard

You are implementing a RAG pipeline where documents are chunked before embedding generation. You notice that semantic sentences are frequently split across chunk boundaries, resulting in degraded retrieval quality. Which chunking strategy should you implement to resolve this?

A.Sliding window chunking with token overlap

B.Fixed-character chunking with zero overlap

C.Paragraph-only splitting with maximum size limits of 10,000 tokens

D.Random token slicing

AnswerA

Sliding window chunking with token overlap carries trailing tokens from one chunk into the next, so sentences spanning a boundary still appear intact in at least one chunk. This directly resolves the semantic splitting that degraded retrieval quality.

Why this answer

Sliding window chunking with token overlap is correct because it preserves semantic continuity by carrying a portion of the previous chunk's tokens into the next chunk. This overlap ensures that sentences or ideas spanning a boundary appear intact in at least one chunk, so the embedding captures the full semantic unit. It directly addresses the root cause of degraded retrieval: fragmented context at chunk edges.

Exam trap

AI-103 often tests the misconception that smaller, non-overlapping chunks improve precision, when in fact zero overlap causes boundary fragmentation that silently degrades recall.

How to eliminate wrong answers

Option B is wrong because fixed-character chunking with zero overlap guarantees that any sentence crossing a boundary is permanently split with no recovery, worsening the fragmentation problem. Option C is wrong because paragraph-only splitting with a 10,000-token maximum produces enormous chunks that exceed typical embedding model context windows (e.g., 8,191 tokens for text-embedding-ada-002) and dilute semantic focus, and paragraphs alone do not prevent mid-sentence splits when a paragraph exceeds the limit. Option D is wrong because random token slicing destroys all linguistic structure, producing incoherent chunks that yield meaningless embeddings and unusable retrieval.

Full explanation →

336

MCQhard

You are implementing a disaster recovery strategy for a critical image processing application that relies on Azure AI Vision. If the primary Azure region experiences an outage, how should you configure failover?

A.Use geo-replication on the Azure Blob Storage account to automatically fail over the AI service endpoints.

B.Configure Azure Site Recovery to replicate the Azure AI Vision service backend VMs.

C.Rely on automatic synchronous replication of Custom Vision models across all global Azure datacenters without application changes.

D.Deploy a secondary Azure AI Vision resource in a paired Azure region and configure your application to fail over to the secondary endpoint if errors occur.

AnswerD

Azure AI Vision is regional, so an outage in the primary region makes that endpoint unavailable. Deploying a duplicate resource in a paired region and failing over to its endpoint on errors provides the cross-region redundancy the disaster recovery requirement demands.

Why this answer

Azure AI Vision is a regional PaaS service, so disaster recovery requires deploying a secondary resource in a paired region and having the application fail over to that endpoint when the primary returns errors. There is no automatic cross-region replication of the service itself, so the application must implement the failover logic. This matches Azure's shared responsibility model for PaaS DR.

Exam trap

AI-103 often tests the misconception that geo-replication of storage or Azure Site Recovery automatically protects PaaS AI services, when in reality each regional AI resource must be independently deployed and the application must handle failover.

How to eliminate wrong answers

Option A is wrong because geo-replicating a Blob Storage account only protects stored data — it does not replicate or fail over the Azure AI Vision compute endpoint, which is a separate regional resource. Option B is wrong because Azure AI Vision is a PaaS service, not a set of customer-managed VMs, so Azure Site Recovery (which replicates IaaS VMs) cannot be applied to its backend. Option C is wrong because Custom Vision models are not synchronously replicated to all global datacenters; they are regional resources requiring explicit export/import or redeployment.

Full explanation →

337

MCQhard

You are implementing Retrieval-Augmented Generation (RAG) using Azure AI Search as the retriever and Azure OpenAI as the generator. Users report that the generated responses frequently miss critical context located in the middle of long retrieved documents. Which grounding optimization technique should you apply?

A.Increase the temperature parameter of the Azure OpenAI model to 1.2 to encourage creative synthesis of the middle content.

B.Switch the Azure AI Search retriever from hybrid search to pure keyword BM25 search.

C.Implement semantic ranking and reorder retrieved documents so the highest-scored chunks appear at the very beginning and end of the prompt context.

D.Reduce the chunk size to 64 tokens to force the retriever to return hundreds of tiny fragments.

AnswerC

Positional bias causes models to attend most strongly to the start and end of context, neglecting middle chunks. Semantic ranking selects the most relevant chunks, and deliberately placing the highest-scored ones at the prompt's beginning and end counteracts the lost-in-the-middle effect.

Why this answer

LLMs exhibit a 'lost in the middle' bias where content at the beginning and end of the context window is attended to more strongly than content in the middle. Applying semantic ranking to surface the most relevant chunks and then reordering them so the highest-scored chunks sit at the start and end of the prompt mitigates this positional bias and improves grounding.

Exam trap

AI-103 often tests the misconception that increasing temperature or switching search modes fixes grounding issues, when the real fix targets the LLM's positional attention bias through semantic ranking and chunk reordering.

How to eliminate wrong answers

Option A is wrong because raising temperature to 1.2 increases randomness and hallucination rather than improving retrieval of middle content, and temperature does not affect attention to context position. Option B is wrong because switching from hybrid search to pure BM25 keyword search discards vector similarity, typically reducing recall and relevance rather than fixing positional attention bias. Option D is wrong because shrinking chunks to 64 tokens fragments context, destroys semantic coherence, and floods the prompt with hundreds of low-value fragments, worsening grounding.

Full explanation →

338

Multi-Selectmedium

When planning the model selection and versioning lifecycle for Azure OpenAI deployments in an enterprise environment, which TWO practices should you follow? (Choose two.)

Select 2 answers

A.Use deprecated model versions indefinitely to maintain identical outputs.

B.Monitor Microsoft documentation and deprecation notices for model version lifecycles.

C.Automatically upgrade all production model deployments to the latest preview version the moment it is released.

D.Test new model versions in a staging or development environment before promoting them to production.

E.Never update model versions once deployed to avoid any future validation work.

AnswersB, D

Monitoring Microsoft documentation and deprecation notices satisfies the lifecycle-planning constraint by surfacing retirement dates and successor model versions before Azure OpenAI retires a deployment. This lets you schedule migrations and pin versions deliberately, avoiding unexpected service disruption when a model version reaches end of life.

Why this answer

Option B is correct because Azure OpenAI model versions have defined lifecycles and Microsoft publishes deprecation and retirement notices; monitoring official documentation lets you plan migrations before a version is retired and avoid unexpected service disruption. Option D is correct because new model versions can change output quality, latency, and behavior, so validating them in a staging or development environment with representative workloads before promoting to production is the standard safe deployment practice. Option A is wrong because deprecated model versions are eventually retired and cannot be used indefinitely, and pinning to them does not guarantee identical outputs across the service.

Option C is wrong because automatically upgrading production to the latest preview version bypasses testing and exposes production workloads to unvalidated, non-GA behavior. Option E is wrong because refusing to ever update model versions leads to forced migrations when retirement occurs and accumulates unmanaged technical debt.

Exam trap

AI-103 often tests the tension between 'always use the latest' and 'validate before promoting', tricking candidates into picking auto-upgrade or never-update extremes instead of the balanced staging-plus-monitoring approach.

Full explanation →

339

MCQeasy

You are evaluating a generative AI application in Azure AI Foundry using automated metrics. You need to measure how well the generated answers are supported by the retrieved source documents. Which evaluation metric should you use?

A.Word Error Rate (WER)

B.Groundedness

C.BLEU score

D.Perplexity

AnswerB

Groundedness measures whether generated answers are supported by the retrieved source documents, directly matching the stem's requirement. It detects hallucinated content unsupported by context, unlike relevance, fluency or coherence, which assess other quality dimensions of the generative response.

Why this answer

Groundedness measures how well a model's generated response is supported by the retrieved source documents, which is exactly the requirement in a RAG-style generative AI application. Azure AI Foundry's evaluation SDK provides a groundedness metric that uses an LLM judge to check whether claims in the answer are entailed by the provided context.

Exam trap

AI-103 often tests the confusion between generation-quality metrics (BLEU, perplexity, WER) and RAG-specific metrics (groundedness, relevance), so candidates who pick a familiar NLP metric miss the grounding requirement.

How to eliminate wrong answers

Option A is wrong because Word Error Rate measures transcription accuracy in speech-to-text scenarios, not answer grounding. Option C is wrong because BLEU compares n-gram overlap between generated and reference text — it measures surface similarity, not factual support by retrieved documents. Option D is wrong because perplexity measures how well a language model predicts a token sequence and is a training-time metric, not a grounding metric.

Full explanation →

340

MCQhard

You need to secure communications between an Azure App Service hosting your web app and an Azure AI Computer Vision service. The web app must authenticate to Computer Vision without storing API keys in application settings. Which feature should you enable?

A.Managed Identity on the App Service and RBAC role assignment on the Computer Vision resource.

B.API Management subscription keys

C.Cross-Origin Resource Sharing (CORS)

D.Storage account shared access signatures (SAS)

AnswerA

Managed Identity gives the App Service an automatically rotated identity in Microsoft Entra ID, and an RBAC role assignment on the Computer Vision resource authorises it. This removes the need to store API keys in application settings, satisfying the stem's constraint.

Why this answer

Enabling a managed identity on the App Service and assigning an appropriate RBAC role on the Computer Vision resource lets the app obtain Azure AD tokens automatically, eliminating the need to store API keys in application settings. The managed identity is trusted by Azure AD, and the RBAC role grants access to the Computer Vision resource. This is the recommended keyless authentication pattern for service-to-service communication in Azure.

Exam trap

AI-103 often tests keyless authentication, and the trap is choosing API Management subscription keys because they sound like a secure API access mechanism, when they are still static secrets that violate the no-stored-keys requirement.

How to eliminate wrong answers

Option B (API Management subscription keys) is wrong because subscription keys are still static secrets that must be stored and rotated, which does not meet the requirement of avoiding stored credentials. Option C (Cross-Origin Resource Sharing) is wrong because CORS is a browser security mechanism controlling cross-origin requests, not an authentication or authorization mechanism between services. Option D (Storage account shared access signatures) is wrong because SAS tokens apply to Azure Storage, not Computer Vision, and they are still time-limited secrets that must be managed.

Full explanation →

341

MCQeasy

You need to extract text lines, words, bounding polygons, and selection marks (such as radio buttons and checkboxes) from arbitrary documents without training. Which prebuilt model should you invoke?

A.Prebuilt Business Card model

B.Prebuilt ID Document model

C.Prebuilt Layout model

D.Custom Neural model

AnswerC

The prebuilt Layout model returns text lines, words, bounding polygons and selection marks (radio buttons, checkboxes) from arbitrary documents without training, satisfying the no-training constraint. Unlike the Read model, which returns only text and no selection marks, Layout adds geometric and mark detection, matching every required output.

Why this answer

The prebuilt Layout model in Azure AI Document Intelligence extracts text lines, words, bounding polygons, and selection marks (checkboxes, radio buttons) from documents without any training. It is designed for general document structure extraction and is the correct choice when you need layout-level elements and selection marks from arbitrary documents. It requires no custom training and works out of the box.

Exam trap

AI-103 often tests whether candidates know that the Layout model is the no-training, general-purpose extractor for text, tables, and selection marks — the trap is selecting a specialized prebuilt model (ID, Business Card) or a custom model when the requirement is arbitrary documents with selection marks.

How to eliminate wrong answers

Option A is wrong because the Prebuilt Business Card model is specialized for extracting contact fields (name, phone, email, company) from business cards — it does not return general text lines, words, bounding polygons, or selection marks. Option B is wrong because the Prebuilt ID Document model targets identity documents (passports, driver's licenses) and extracts specific identity fields, not arbitrary layout elements or selection marks. Option D is wrong because the Custom Neural model requires labeled training data and is built for domain-specific field extraction — it is not a no-training, general-purpose layout extractor.

Full explanation →

342

MCQmedium

You are deploying an Azure OpenAI GPT-4 model in a region experiencing high demand, and you need to guarantee a dedicated throughput level without contending with other tenants. Which deployment type should you select?

A.Serverless deployment

B.Standard deployment

C.Provisioned (PTU) deployment

D.Batch deployment

AnswerC

Provisioned throughput units reserve dedicated model capacity for your deployment, guaranteeing a fixed throughput level that is not shared with other tenants. This removes the contention and variable latency inherent in standard pay-as-you-go deployments during periods of high regional demand.

Why this answer

Provisioned Throughput Units (PTU) deployments reserve dedicated model capacity for your subscription, guaranteeing consistent throughput and latency without contending with other tenants. This is the only deployment type that provides the dedicated capacity and isolation the scenario requires. Standard and serverless deployments share pooled capacity and can be throttled under high regional demand.

Exam trap

AI-103 often tests the distinction between shared-capacity deployments (Standard/Serverless) and dedicated-capacity deployments (PTU), so candidates focused on cost pick Standard and miss the 'guaranteed dedicated throughput' requirement.

How to eliminate wrong answers

Option A is wrong because serverless (pay-as-you-go) deployments draw from shared regional capacity and offer no throughput guarantee, making them unsuitable when demand is high. Option B is wrong because Standard deployments also use shared capacity with token-based billing and can experience throttling or variable latency under load. Option D is wrong because Batch deployments are designed for asynchronous, large-volume jobs with relaxed latency requirements, not for guaranteed interactive throughput.

Full explanation →

343

Multi-Selectmedium

You are designing an asynchronous text analysis workflow using the Azure AI Language service to process large volumes of documents (e.g., hundreds of megabytes of text files stored in Azure Blob Storage). Which TWO features support asynchronous batch processing jobs via the /analyze-text/jobs endpoint?

Select 2 answers

A.Standard sentiment analysis

B.Language detection

C.Custom text classification

D.Key phrase extraction

E.Text summarization (Abstractive)

AnswersC, E

Custom text classification supports asynchronous batch jobs through the /analyze-text/jobs endpoint, submitting documents from Azure Blob Storage and polling for results. This satisfies the stem's requirement for processing hundreds of megabytes without synchronous per-document calls, since long-running jobs handle volume asynchronously rather than blocking on individual requests.

Why this answer

The /analyze-text/jobs endpoint is the Azure AI Language asynchronous batch API, and it is designed for long-running document processing tasks that submit a job and poll for results. Option C (Custom text classification) is correct because custom text classification is explicitly supported as an asynchronous job through /analyze-text/jobs, allowing you to classify large volumes of documents against your trained model. Option E (Text summarization (Abstractive)) is correct because abstractive summarization is also a long-running operation submitted via /analyze-text/jobs, which is appropriate for generating summaries over hundreds of megabytes of text.

Options A (Standard sentiment analysis), B (Language detection), and D (Key phrase extraction) are not correct here because these are synchronous, single-document or small-batch features exposed through the /analyze-text endpoint (or the synchronous analyze-text call), not the asynchronous jobs endpoint.

Exam trap

The trap is assuming all Azure AI Language features support the async jobs endpoint; in reality only long-running features like custom classification, custom NER, and summarization do, while sentiment, language detection, and key phrases are synchronous.

Full explanation →

344

Multi-Selecteasy

Which TWO components are essential for creating an effective system prompt?

Select 2 answers

A.A clear definition of the agent's persona and objective.

B.A list of all users authorized to use the agent.

C.The full training dataset of the base model.

D.Constraints and instructions on how to handle edge cases or disallowed topics.

E.A list of all hardware components in the data center.

AnswersA, D

Defining the agent's persona and objective anchors every subsequent instruction, giving the model a stable role and success criteria to reason against. Without this, responses drift and tone becomes inconsistent, so it satisfies the stem's requirement for an essential system-prompt component.

Why this answer

Option A is correct because an effective system prompt must establish the agent's persona and objective, giving the model a clear role, tone, and goal that shape all subsequent responses. Option D is correct because system prompts need explicit constraints and instructions for edge cases or disallowed topics, which set behavioral boundaries and safety guardrails for the agent. Options B, C, and E are not essential: authorized-user lists are typically enforced by authentication and access-control systems rather than the prompt, the base model's full training dataset is not something the prompt author supplies, and data center hardware details are irrelevant to prompt design.

Exam trap

AI-103 often tests whether candidates understand that system prompts define behavior and boundaries, not access control or infrastructure — candidates may pick 'list of authorized users' because it sounds like a security requirement, but authorization is enforced outside the model.

Full explanation →

345

MCQeasy

You need to extract key concepts and main ideas from long-form technical support articles using the Azure AI Language service. Which feature should you invoke?

A.Personal Personally Identifiable Information (PII) detection

B.Language detection

C.Summarization

D.Key phrase extraction

AnswerD

Key phrase extraction returns the salient terms and concepts from each document, which directly satisfies the requirement to surface key concepts and main ideas from long-form articles. It operates on unstructured text without training, unlike custom classification or summarisation.

Why this answer

Key phrase extraction is the Azure AI Language feature designed to identify the main concepts and important terms from unstructured text. It analyzes the text and returns a list of key phrases that represent the core ideas, making it ideal for extracting key concepts from long-form technical support articles. Summarization, while related, produces a condensed version of the text rather than a list of key phrases, and is typically used for creating abstracts or summaries.

Therefore, key phrase extraction directly fulfills the requirement to extract key concepts and main ideas.

Exam trap

AI-103 often tests the distinction between summarization and key phrase extraction, as both involve extracting important information from text, but key phrase extraction returns a list of key phrases while summarization produces a condensed summary.

How to eliminate wrong answers

Option A is wrong because PII detection identifies and redacts personal information like names, addresses, and phone numbers, not key concepts. Option B is wrong because language detection determines the language of the text, which is unrelated to extracting main ideas. Option C is wrong because summarization generates a concise summary of the text, but the question specifically asks for extracting key concepts and main ideas, which is the purpose of key phrase extraction, not summarization.

Full explanation →

346

MCQmedium

Your development team needs to deploy an Azure AI Language resource that can be accessed from an on-premises datacenter via an ExpressRoute circuit without traversing the public internet. Which Azure feature should you configure?

A.Azure Bastion host subnet

B.Azure Private Endpoint

C.NAT gateway resource attachment

D.Azure Front Door with Web Application Firewall

AnswerB

Azure Private Endpoint assigns a private IP from your VNet to the Azure AI Language resource, so traffic from on-premises over ExpressRoute reaches it via private peering and never traverses the public internet, satisfying the stated no-public-internet constraint.

Why this answer

Azure Private Endpoint assigns a private IP from your VNet to the Azure AI Language resource, so traffic from on-premises over ExpressRoute reaches the service via private IP without traversing the public internet. This is the standard Private Link pattern for PaaS services.

Exam trap

AI-103 often tests whether candidates confuse outbound internet connectivity (NAT gateway) or public ingress (Front Door) with private inbound connectivity (Private Endpoint), so they pick a service that actually routes traffic over the public internet.

How to eliminate wrong answers

Option A is wrong because Azure Bastion provides secure RDP/SSH to VMs, not private connectivity to PaaS AI services. Option C is wrong because a NAT gateway provides outbound internet SNAT for subnets, which is the opposite of avoiding the public internet. Option D is wrong because Front Door with WAF is a public-facing global entry point and does not provide private on-premises connectivity.

Full explanation →

347

MCQeasy

You are configuring an Azure AI Language resource and want to ensure that it can be accessed securely from an Azure Virtual Network (VNet) without traversing the public internet. Which Azure networking feature should you configure?

A.Private endpoints

B.Azure Front Door

C.Azure Bastion

D.VPN Gateway

AnswerA

Private endpoints assign a private IP from the VNet to the Azure AI Language resource, so traffic stays on the Microsoft backbone and never traverses the public internet. Service endpoints do not provide this private addressing.

Why this answer

Private endpoints create a private IP address for an Azure AI Language resource within your VNet, enabling secure access over the Microsoft backbone network without exposing traffic to the public internet. This is the standard way to privately connect to Azure PaaS services. It uses Azure Private Link to map the service to a private endpoint in your subnet.

Exam trap

The trap here is confusing network connectivity options: VPN Gateway connects networks, Bastion connects to VMs, and Front Door is for global HTTP traffic, but only Private Endpoint provides a private IP for a PaaS service within a VNet.

How to eliminate wrong answers

Option B is wrong because Azure Front Door is a global HTTP load balancer and CDN that operates at the edge and does not provide private VNet connectivity to a PaaS service. Option C is wrong because Azure Bastion provides secure RDP/SSH access to virtual machines, not private connectivity to AI services. Option D is wrong because VPN Gateway connects on-premises networks to Azure VNets, but it does not create a private endpoint for a PaaS service; it would still traverse the public internet unless combined with private endpoints.

Full explanation →

348

MCQmedium

You are implementing a tool call in a Prompt Flow. The tool needs to interact with an external REST API. What is the recommended way to manage the API credentials securely?

A.Pass the key as a parameter in the input prompt

B.Hardcode the API key in the Prompt Flow code

C.Use an environment variable in the flow definition

D.Use an Azure AI Foundry Connection of type Custom

AnswerD

A Custom Connection in Azure AI Foundry stores the API key or token as a managed secret, so Prompt Flow tools reference it by name rather than embedding credentials in code. This satisfies the requirement to manage the REST API credentials securely.

Why this answer

Azure AI Foundry Connections of type Custom are the recommended way to store and manage credentials for external services such as REST APIs. A Custom connection securely stores the API key or token in the Foundry workspace, and the Prompt Flow tool references the connection by name, so credentials are never exposed in code or prompts. This integrates with Azure RBAC and Key Vault for secure secret storage.

Exam trap

The trap is assuming that environment variables are a secure way to store secrets — candidates often pick 'use an environment variable' because it sounds like a best practice, but in Azure AI Foundry the recommended secure mechanism is a Connection backed by Key Vault, not a plain environment variable.

How to eliminate wrong answers

Option A is wrong because passing the key as a parameter in the input prompt exposes the credential in the prompt text, logs, and conversation history — a serious security anti-pattern. Option B is wrong because hardcoding the API key in the Prompt Flow code embeds the secret in source control and flow artifacts, making rotation and auditing difficult and risking leakage. Option C is wrong because while environment variables can hold configuration, they are not a secure secrets store — they can leak through logs, process listings, and flow exports, and they lack the RBAC and Key Vault integration that Foundry Connections provide.

Full explanation →

349

Multi-Selecthard

Your organization is implementing strict network isolation for Azure OpenAI. Direct internet access from worker nodes is prohibited. Which THREE mechanisms must be correctly configured to allow secure communication? Each correct answer presents part of the solution.

Select 3 answers

A.Public network access disabled on the Azure OpenAI resource

B.Public IP addresses assigned directly to the AI service endpoint

C.Azure Bastion host attached to the OpenAI backend storage

D.Private DNS zone integration for openai.azure.com resolution

E.Azure Private Endpoint connected to the application virtual network subnet

AnswersA, D, E

Disabling public network access removes the Azure OpenAI resource's internet-facing endpoint, forcing all traffic through private paths. This satisfies the stem's prohibition on direct internet access from worker nodes, since clients must then reach the service via a private endpoint inside the virtual network rather than the public FQDN.

Why this answer

Option A is correct because disabling public network access on the Azure OpenAI resource ensures that the service is not reachable over the public internet, forcing all traffic through approved private paths. Option D is correct because a Private DNS zone (privatelink.openai.azure.com) must be integrated with the virtual network so that the FQDN <resource>.openai.azure.com resolves to the private endpoint's private IP instead of the public IP. Option E is correct because an Azure Private Endpoint placed in the application virtual network subnet provides the private IP interface that worker nodes use to reach the Azure OpenAI service over the Microsoft backbone.

Option B is wrong because assigning public IP addresses to the AI service endpoint contradicts the strict isolation requirement and would expose the service publicly. Option C is wrong because Azure Bastion is a jump-host service for RDP/SSH access to VMs and has no role in connecting to Azure OpenAI backend storage.

Exam trap

AI-103 often tests whether candidates include irrelevant services like Azure Bastion or public IPs, confusing network isolation with jump-host access — the key is recognizing that Private Endpoint, Private DNS, and disabling public access are the three required pillars.

Full explanation →

350

MCQhard

You need to optimize the retrieval performance of your RAG agent. You find that the search engine returns too many irrelevant documents. What should you adjust?

A.Configure Hybrid Search settings in the search index.

B.Increase the model temperature.

C.Change the model to a newer version.

D.Reduce the system prompt length.

AnswerA

Hybrid search combines keyword and vector queries, letting you tune weighting between semantic similarity and exact term matching. Raising the keyword component or adjusting the fusion ranking filters out documents that match only loosely on embeddings, directly reducing the irrelevant results the stem describes.

Why this answer

Hybrid Search in Azure AI Search combines keyword (BM25) and vector similarity with Reciprocal Rank Fusion (RRF), and lets you tune parameters like k-nearest neighbors count, semantic ranker, and weighting to improve precision. When too many irrelevant documents are returned, adjusting hybrid search settings — such as enabling semantic reranking or reducing top-k — directly addresses retrieval relevance. The other options affect generation, not retrieval.

Exam trap

AI-103 often tests whether candidates can separate retrieval tuning (index, hybrid search, semantic ranker, top-k) from generation tuning (temperature, model version, prompt length) — candidates frequently pick temperature or model upgrades when the problem is retrieval relevance.

How to eliminate wrong answers

Option B is wrong because increasing model temperature makes the LLM's output more random and creative; it does not change which documents the retriever returns and will likely worsen answer quality. Option C is wrong because upgrading the model version changes generation quality, not retrieval precision — irrelevant documents will still be retrieved and passed to the model. Option D is wrong because shortening the system prompt affects instruction-following and token cost, not the relevance of retrieved documents.

Full explanation →

351

MCQmedium

You are implementing a RAG solution where documents contain complex hierarchical headings. To ensure search chunks retain sufficient context about which section they belong to, what chunking strategy should you adopt?

A.Fixed-size character chunking with zero overlap

B.Random token sampling without structural awareness

C.Single-chunk whole document ingestion

D.Hierarchical (parent-child) chunking retaining section headers

AnswerD

Hierarchical parent-child chunking indexes small child chunks for precise embedding matches while linking each to a larger parent containing its section headers and surrounding hierarchy. Retrieved children are expanded to parents, so chunks retain the heading context the stem requires for complex hierarchical documents.

Why this answer

Hierarchical (parent-child) chunking retains section headers and structural context, so each child chunk knows which parent section it belongs to. This preserves the document's hierarchy, ensuring that retrieved chunks carry enough context for the LLM to answer accurately. It is the recommended strategy for complex hierarchical documents in RAG.

Exam trap

The trap is defaulting to fixed-size chunking because it is simple — the exam tests whether you recognize that hierarchical documents require structure-aware chunking to preserve context.

How to eliminate wrong answers

Option A is wrong because fixed-size chunking with zero overlap ignores document structure and can split mid-section, losing header context. Option B is wrong because random token sampling destroys semantic and structural coherence. Option C is wrong because ingesting the whole document as a single chunk exceeds context limits and reduces retrieval precision.

Full explanation →

352

Multi-Selecthard

When establishing a governance strategy for Azure AI resources across multiple cloud teams, which THREE administrative tasks should an administrator perform? (Choose three.)

Select 3 answers

A.Apply Role-Based Access Control (RBAC) following the principle of least privilege.

B.Enforce naming conventions and allowed regions using Azure Policy.

C.Organize subscriptions under Azure Management Groups for hierarchical governance and policy inheritance.

D.Disable Azure Activity Logs to prevent administrative audit trails.

E.Assign the Owner role to all developers to streamline resource creation.

AnswersA, B, C

Assigning RBAC roles at the appropriate scope with least privilege limits each team to the AI resources and operations it needs, preventing over-permissioned access across subscriptions. This is a core governance control for multi-team Azure AI environments.

Why this answer

Option A is correct because applying Role-Based Access Control (RBAC) with least privilege ensures each cloud team member receives only the permissions needed for their role, such as Reader, Contributor, or specific custom roles, reducing the blast radius of misconfiguration or compromise across Azure AI resources. Option B is correct because Azure Policy can enforce organizational standards such as naming conventions and allowed regions by evaluating resource properties during creation and updates, and it can deny noncompliant deployments or flag them for remediation, which is essential for consistent multi-team governance. Option C is correct because organizing subscriptions under Azure Management Groups creates a hierarchy where policies and RBAC assignments inherit down to child subscriptions, giving administrators a single governance control plane across multiple cloud teams without configuring each subscription individually.

Option D is incorrect because disabling Azure Activity Logs would remove the administrative audit trail needed for compliance, security monitoring, and incident investigation, directly undermining governance. Option E is incorrect because assigning the Owner role to all developers violates least privilege and grants excessive control over access and billing, increasing risk rather than establishing sound governance.

Exam trap

AI-103 often tests whether candidates recognize that governance requires least privilege and auditability, tricking those who equate 'streamlining' with granting broad Owner roles or disabling logs.

Full explanation →

353

MCQmedium

Your team needs to receive an email alert if the error rate for your AI endpoint exceeds 5% for more than 15 minutes. What should you create?

A.An Azure Service Health notification.

B.An Azure Monitor Alert rule with an Action Group.

C.A scheduled Azure Function that polls the API.

D.A Log Analytics query saved as a report.

AnswerB

An Azure Monitor alert rule evaluates the endpoint's error-rate metric against a 5% threshold sustained over a 15-minute window, satisfying both conditions in the stem. Its action group then delivers the email notification. Metric alerts natively support this sustained-window evaluation, unlike log-based queries or diagnostic settings.

Why this answer

Azure Monitor Alert rules evaluate metrics or log queries against thresholds and trigger Action Groups, which can send emails, SMS, webhooks, or invoke Logic Apps and Functions. To alert when error rate exceeds 5% for over 15 minutes, you create a metric or log-search alert rule with the appropriate threshold and evaluation window, then attach an Action Group configured with an email receiver. This is the native, supported mechanism for metric-based alerting in Azure.

Exam trap

AI-103 often tests whether candidates confuse Service Health (platform incidents) with Azure Monitor alerts (resource metrics/logs), or mistakenly believe a saved Log Analytics query alone can send notifications without an alert rule and Action Group.

How to eliminate wrong answers

Option A is wrong because Azure Service Health notifications report platform-wide incidents, planned maintenance, and advisories affecting Azure services — they do not monitor your application's error rate metric. Option C is wrong because a scheduled Azure Function polling the API is a custom, brittle solution that duplicates what Azure Monitor already does natively, adds latency, and requires maintaining code. Option D is wrong because a saved Log Analytics query is only a report — it does not evaluate thresholds or send notifications on its own; it must be wrapped in a scheduled alert rule to trigger actions.

Full explanation →

354

Multi-Selecthard

You are using Azure AI Vision Image Analysis 4.0 to analyze a batch of images. You need to extract both the caption and the detected objects with their bounding boxes. You are writing code that calls the Analyze Image API. Which two parameters must you include in your request? (Choose two.)

Select 2 answers

A.visualFeatures=caption

B.details=landmarks

C.visualFeatures=objects

D.language=en

E.model-version=latest

AnswersA, C

The visualFeatures parameter specifies which visual features to extract. Setting it to caption requests a single-sentence description of the image content. This is required to obtain the caption output. Without this parameter, the response will not include caption data, so it is essential for meeting the requirement of extracting captions.

Why this answer

To obtain both captions and object bounding boxes from a single Analyze Image call, the visualFeatures query parameter must include both caption and objects. These features are requested together, and the response will contain the corresponding sections. Other parameters like details, language, or model-version are optional and do not enable these specific outputs.

Exam trap

The trap here is confusing the details parameter with visualFeatures, thinking that details controls object detection or captions.

Full explanation →

355

MCQeasy

You are developing a web application that allows users to upload images and receive a list of detected objects with their bounding boxes. You want to use the Azure AI Vision Image Analysis 4.0 SDK for Python. Which method should you call to analyze an image from a URL and retrieve object detection results?

A.get_area_of_interest

B.tag_image

C.generate_thumbnail

D.analyze_image

AnswerD

The analyze_image method of the ImageAnalysisClient is used to submit an image for analysis and request specific visual features such as objects. By passing the image URL and specifying the objects feature, you receive the detected objects with bounding boxes. This is the correct method for this scenario.

Why this answer

The analyze_image method is the primary entry point for image analysis in the Azure AI Vision SDK. By specifying visual features such as objects, it returns detected objects with their bounding boxes and labels. Other methods like generate_thumbnail, get_area_of_interest, and tag_image serve different purposes and do not provide object detection with coordinates.

Exam trap

The trap here is assuming that tag_image returns bounding boxes, when it only returns descriptive tags without spatial information.

Full explanation →

356

MCQmedium

You are integrating Azure AI Search with Azure OpenAI for grounded chat generation. Users report that the generated answers are accurate, but the citations returned to the user point to raw internal file paths rather than user-friendly web URLs. How should you resolve this?

A.Enable Azure AI Content Safety citation filtering

B.Decrease the search result top_k parameter to 1

C.Store public URL equivalents in index metadata fields and configure citation mapping in your orchestration layer

D.Modify the Azure OpenAI system prompt to guess web URLs based on file names

AnswerC

Storing public URL equivalents in index metadata lets retrieval return the user-facing address alongside each chunk, while citation mapping in the orchestration layer substitutes that metadata into the response instead of the raw file path. This satisfies the stem's constraint: accurate answers with citations pointing to friendly web URLs.

Why this answer

Citations in grounded chat are generated from metadata fields in the Azure AI Search index; to return user-friendly URLs, you must store the public URL in a retrievable metadata field (e.g., 'url' or 'source_url') and configure your orchestration layer to map that field into the citation output. Azure OpenAI's 'on your data' feature and custom RAG pipelines both rely on index metadata for citation rendering. This is a data-modeling and orchestration fix, not a model or safety fix.

Exam trap

AI-103 often tests whether candidates understand that citations come from index metadata, not from the LLM, tricking them into trying to fix citations via prompt engineering or content safety settings.

How to eliminate wrong answers

Option A is wrong because Azure AI Content Safety filters harmful content; it has no role in citation URL formatting or metadata mapping. Option B is wrong because reducing top_k to 1 only limits the number of retrieved documents — it does not change what URL is embedded in the citation, and it degrades answer quality. Option D is wrong because instructing the model to 'guess' URLs is unreliable and will produce hallucinated links, violating grounding principles and potentially exposing incorrect references.

Full explanation →

357

MCQmedium

You are building a RAG application where documents are updated frequently. You need to ensure that the Azure AI Search index reflects document deletions and updates in near real-time without performing full re-indexing. How should you design the ingestion pipeline?

A.Increase the Azure OpenAI model deployment concurrency quota

B.Enable semantic caching on the OpenAI endpoint

C.Configure Azure AI Search indexers with change tracking and scheduled incremental synchronization

D.Delete the entire search service and recreate the index daily using a batch script

AnswerC

Indexers with change tracking detect additions, updates and deletions in the data source, then apply only those deltas on a schedule. This keeps the index synchronised in near real-time without rebuilding it, satisfying the no-full-re-indexing constraint.

Why this answer

Azure AI Search indexers support change tracking via the data source's native mechanisms (e.g., SQL rowversion, Cosmos DB change feed, or blob soft-delete policies) and run on a schedule to detect and apply only changed or deleted documents. This incremental synchronization keeps the index current in near real-time without full re-indexing. It is the designed pattern for frequently updated content.

Exam trap

AI-103 often tests whether candidates confuse model-side tuning (quota, caching) with data-plane synchronization, so choose the option that explicitly mentions indexers with change tracking.

How to eliminate wrong answers

Option A is wrong because increasing Azure OpenAI deployment concurrency quota affects request throughput to the model, not index freshness or document synchronization. Option B is wrong because semantic caching stores prior model responses to reduce latency and cost; it does not update the search index when source documents change. Option D is wrong because deleting and recreating the search service daily is destructive, causes downtime, loses index configuration, and is not near-real-time — it is the opposite of incremental synchronization.

Full explanation →

358

Multi-Selectmedium

Your company wants to track and allocate Azure AI service costs by department. Which TWO mechanisms can assist with this requirement? (Choose two.)

Select 2 answers

A.Using Azure Cost Management filters and cost allocation rules to generate chargeback reports.

B.Sharing a single Azure subscription across all departments without using tags or resource groups.

C.Deactivating Azure Cost Management to reduce management overhead.

D.Provisioning all AI resources in the Free (F0) tier.

E.Applying department-specific cost tracking tags (e.g., Department: Marketing) to Azure AI resources.

AnswersA, E

Azure Cost Management applies filters and cost allocation rules to split shared Azure AI service spend across departments, producing chargeback reports. This directly satisfies the requirement to track and allocate costs by department, since allocation rules redistribute tagged or untagged usage to the correct cost centres.

Why this answer

Option A is correct because Azure Cost Management supports filters (by resource group, tag, service, subscription, etc.) and cost allocation rules that let you split shared costs across departments, enabling accurate chargeback/showback reports for Azure AI services. Option E is correct because applying department-specific tags such as Department: Marketing to Azure AI resources lets Cost Management group and filter spend by tag, which is the standard way to attribute AI service costs to each department. Option B is incorrect because a single untagged, unsegmented subscription provides no dimension to separate or allocate costs by department.

Option C is incorrect because deactivating Azure Cost Management removes the very reporting and allocation capability needed for cost tracking. Option D is incorrect because the Free (F0) tier only affects pricing/limits of certain AI services and does not provide any cost tracking or allocation mechanism.

Exam trap

AI-103 often tests the misconception that simply using a single subscription or free tier solves cost tracking, when tags and Cost Management allocation rules are the actual mechanisms required.

Full explanation →

359

MCQmedium

You are using Document Intelligence to classify documents into different types (e.g., invoices, bank statements). Which model is most appropriate?

A.Prebuilt-invoice

B.Custom classification model

C.Read model

D.Custom extraction model

AnswerB

A custom classification model trains on your labelled samples to assign documents to your own categories, such as invoices versus bank statements. Prebuilt models extract fields from known types, so they cannot classify arbitrary custom document types.

Why this answer

A custom classification model is purpose-built to categorize documents into user-defined types (e.g., invoice vs. bank statement) before or alongside extraction. It is trained on labeled examples of each document class and returns a document type with confidence. Prebuilt-invoice extracts fields from invoices but does not classify arbitrary document types.

Exam trap

AI-103 often tests the distinction between classification (what type of document is this?) and extraction (what fields are in this document?) — candidates pick 'custom extraction' because it sounds more powerful but it does not classify.

How to eliminate wrong answers

Option A is wrong because Prebuilt-invoice only extracts invoice fields — it cannot distinguish invoices from bank statements or other classes. Option C is wrong because the Read model only performs OCR and returns text/lines; it has no classification capability. Option D is wrong because a custom extraction model pulls structured fields from a known document type; it does not classify documents into types.

Full explanation →

360

Multi-Selectmedium

Which TWO factors directly influence the embedding quality and retrieval relevance in an Azure AI Search vector retrieval pipeline? (Choose two.)

Select 2 answers

A.The background color of the source documents in blob storage.

B.The choice of embedding model (e.g., text-embedding-3-large vs text-embedding-3-small) and its dimensionality.

C.The font family used in the original PDF documents.

D.The client browser version used by the end user.

E.The chunking strategy and chunk size used to segment source documents.

AnswersB, E

The embedding model determines the vector space's semantic resolution and its dimensionality sets how much meaning each vector can encode. A richer, higher-dimensional model such as text-embedding-3-large captures finer distinctions, directly improving retrieval relevance in the Azure AI Search pipeline.

Why this answer

Option B is correct because the embedding model determines the semantic quality of the vectors stored in the Azure AI Search index: text-embedding-3-large produces higher-fidelity, higher-dimensional vectors (3072 dimensions) than text-embedding-3-small (1536 dimensions), directly affecting how well semantic similarity matches user intent during vector search. Option E is correct because chunking strategy and chunk size control the granularity of the indexed content; chunks that are too large dilute the embedding with unrelated content, while chunks that are too small lose context, so both the segmentation method (e.g., fixed-size, sentence-aware, or document-layout-aware) and the chunk length materially affect retrieval relevance. Options A and C are incorrect because background color and font family are visual formatting attributes that are not captured by text embedding models and have no bearing on vector similarity.

Option D is incorrect because the end user's browser version affects client-side rendering and request handling, not the embedding quality or the relevance scoring performed by Azure AI Search.

Exam trap

AI-103 often tests whether candidates can distinguish content-affecting factors from cosmetic ones — the trap is selecting superficial attributes (font, background color, browser) instead of the model choice and chunking strategy that actually drive embedding quality.

Full explanation →

361

Multi-Selectmedium

You are configuring an agent in Azure AI Agent Service that needs to utilize tools. Which TWO types of tools are natively supported by Azure AI Agent Service? (Choose TWO)

Select 2 answers

A.Azure SQL Database stored procedure compiler

B.Azure Virtual Machine provisioning extension

C.Function calling (Custom tools/APIs)

D.Azure Stream Analytics tumbling window operator

E.Code Interpreter

AnswersC, E

Function calling lets the agent invoke custom tools and APIs you define, extending it beyond built-in capabilities. Azure AI Agent Service natively supports this tool type, satisfying the requirement that the agent utilise external tools through declared function schemas the model can call.

Why this answer

Azure AI Agent Service natively supports Function calling (option C), which lets the agent invoke custom tools and APIs by defining function schemas that the model can call to retrieve data or perform actions. It also natively supports Code Interpreter (option E), a sandboxed Python execution environment used for data analysis, math, and file generation. Options A, B, and D are not native tool types in Azure AI Agent Service: a 'stored procedure compiler,' a 'VM provisioning extension,' and a 'Stream Analytics tumbling window operator' are unrelated Azure capabilities and cannot be attached as agent tools.

Exam trap

AI-103 often tests whether candidates can distinguish native agent tools from unrelated Azure services that merely sound technical, so eliminate options naming database, VM, or stream-processing features.

Full explanation →

362

MCQhard

You are using an Azure AI model that supports versioning. You want to perform an A/B test by routing 10% of traffic to a new version. Where do you configure this?

A.Azure App Service routing rules.

B.Azure Load Balancer backend pools.

C.Azure Traffic Manager DNS settings.

D.The 'Traffic' configuration in the Azure AI endpoint.

AnswerD

The Traffic configuration on an Azure AI endpoint supports weighted routing, letting you split requests across model versions by percentage. Setting the new version to 10% satisfies the A/B test constraint, while the remaining 90% continues to the existing deployment without code changes.

Why this answer

Azure AI endpoints expose a 'Traffic' configuration where you assign percentage weights to deployed model versions, enabling A/B testing by routing a fraction of requests to a new version. Setting 10% to the new version and 90% to the current one achieves the desired split. This is native to the endpoint and requires no external load balancer.

Exam trap

AI-103 often tests whether candidates know that traffic splitting for model versions lives in the AI endpoint configuration, so the trap is choosing a generic Azure load-balancing service that cannot split by model version.

How to eliminate wrong answers

Option A is wrong because Azure App Service routing rules apply to web app slots and hostnames, not to AI model version traffic splits. Option B is wrong because Azure Load Balancer backend pools distribute Layer 4 traffic across VMs or instances, with no concept of model versions or percentage-based A/B splits. Option C is wrong because Azure Traffic Manager performs DNS-based global routing across endpoints by geography or priority — it cannot split traffic by model version percentage within a single AI endpoint.

Full explanation →

363

MCQeasy

Your company processes business cards from international clients. Which prebuilt Azure AI Document Intelligence model should you use to extract names, company names, emails, and phone numbers from these cards?

A.prebuilt-businessCard

B.prebuilt-invoice

C.prebuilt-layout

D.prebuilt-receipt

AnswerA

The prebuilt-businessCard model is purpose-built to extract contact fields — names, company names, emails and phone numbers — from business card images. It satisfies the stem's requirement for a prebuilt model targeting business cards, avoiding the need to train a custom extraction model.

Why this answer

The prebuilt-businessCard model is specifically trained to extract common fields from business cards, including names, company names, emails, and phone numbers. It is optimized for the typical layout and content of business cards, making it the correct choice for this scenario.

Exam trap

AI-103 often tests the distinction between prebuilt models by their specific field sets; candidates may confuse business card extraction with invoice or receipt models if they overlook the unique fields like emails and phone numbers.

How to eliminate wrong answers

Option B is wrong because prebuilt-invoice is designed for invoices, extracting fields like invoice ID, due date, and vendor details, not business card information. Option C is wrong because prebuilt-layout extracts text and structure but does not provide semantic fields like names or emails. Option D is wrong because prebuilt-receipt is tailored for sales receipts, capturing items, totals, and transaction details, not business card data.

Full explanation →

364

MCQeasy

You are implementing Retrieval-Augmented Generation (RAG) in Azure AI Foundry using Azure AI Search as the grounding data source. Which indexing feature should you enable to ensure the system can perform semantic ranking alongside traditional keyword search?

A.Fuzzy Search

B.Lucene Standard Analyzer

C.Semantic Search

D.Vector Normalization

AnswerC

Semantic Search adds a reranking layer over results, using Microsoft's language models to rescore matches by meaning while retaining keyword search. Enabling it on the Azure AI Search index satisfies the requirement for semantic ranking alongside traditional keyword search.

Why this answer

Semantic Search in Azure AI Search enables the semantic ranker, which re-ranks an initial BM25 keyword result set using deep-learning models to better capture query intent. It is the feature that layers semantic ranking on top of traditional keyword (and optionally vector) search, improving relevance for natural-language queries.

Exam trap

AI-103 often tests the confusion between lexical analyzers/fuzzy matching and true semantic ranking, causing candidates to pick analyzer or vector-normalization options that do not provide semantic re-ranking.

How to eliminate wrong answers

Option A is wrong because Fuzzy Search handles typo tolerance and approximate matching in the lexical layer, not semantic re-ranking. Option B is wrong because the Lucene Standard Analyzer is a text-analysis component that tokenizes and normalizes text at index time; it does not provide semantic ranking. Option D is wrong because Vector Normalization scales vector magnitudes for similarity computation (e.g., cosine similarity) and is unrelated to semantic ranking of keyword results.

Full explanation →

365

Multi-Selecteasy

As part of responsible AI practices in Azure AI Foundry, which TWO actions should you perform to evaluate and mitigate model fairness and safety risks before deployment? (Choose two)

Select 2 answers

A.Bypass security reviews if the application is intended for internal enterprise employees only.

B.Run automated evaluations in Azure AI Foundry to measure metrics such as groundedness, relevance, and protected material presence.

C.Remove all system prompts to allow the model complete autonomy during evaluation runs.

D.Perform adversarial testing (red teaming) to identify potential prompt injections, jailbreaks, and unintended behaviors.

E.Hardcode responses for all possible user inputs to eliminate model non-determinism entirely.

AnswersB, D

Automated evaluations in Azure AI Foundry score model outputs against metrics including groundedness, relevance, and protected material presence. This satisfies the stem's evaluation requirement by quantifying safety and quality risks systematically, providing repeatable evidence before deployment rather than relying on manual spot checks.

Why this answer

Option B is correct because Azure AI Foundry's automated evaluation flows let you quantify safety and quality metrics—groundedness, relevance, coherence, fluency, and protected-material/hate-and-unfairness scores—so you can objectively assess fairness and safety risks before deployment. Option D is correct because adversarial testing (red teaming) in Azure AI Foundry systematically probes the model with prompt injections, jailbreak attempts, and other malicious inputs to surface unintended behaviors and safety gaps that must be mitigated prior to release. Option A is wrong because internal-only audiences do not exempt an application from responsible AI security and safety reviews; risk assessments apply regardless of user population.

Option C is wrong because removing system prompts strips guardrails and grounding instructions, increasing unsafe or off-topic outputs rather than enabling meaningful evaluation. Option E is wrong because hardcoding responses is not a fairness/safety evaluation technique and would defeat the purpose of testing a generative model's behavior.

Exam trap

AI-103 often tests the misconception that internal-only applications can skip responsible AI reviews, or that removing system prompts improves safety, causing candidates to select those options instead of evaluation and red teaming.

Full explanation →

366

MCQhard

You are managing a multi-tenant SaaS application that uses Azure AI Vision. Each tenant has their own isolated data, and you want to track costs per tenant. How should you provision Azure AI services to achieve granular cost attribution?

A.Use a single global Azure AI Services resource and parse tenant IDs from request user-agent headers for billing.

B.Store tenant IDs in Custom Vision project tags.

C.Deploy a dedicated Azure AI Services resource per tenant and use Azure Cost Management for billing attribution.

D.Use Azure Blob Storage cost allocation tags to cover AI API costs.

AnswerC

A dedicated Azure AI Services resource per tenant creates a separate billing boundary, so each tenant's consumption maps to its own resource and cost data. Azure Cost Management then attributes spend per resource, satisfying the requirement for granular per-tenant cost tracking in the multi-tenant SaaS application.

Why this answer

Deploying a dedicated Azure AI Services resource per tenant is the only approach that provides true isolation and granular cost attribution, because Azure Cost Management tracks costs at the resource level. Each tenant's resource appears as a separate line item in cost analysis, enabling per-tenant billing. This also satisfies the data-isolation requirement mentioned in the scenario.

Exam trap

AI-103 often tests cost-attribution design — the trap is picking a shared-resource approach with header parsing or tags on unrelated services, when only per-resource deployment gives Azure Cost Management the granularity it needs.

How to eliminate wrong answers

Option A is wrong because a single shared resource aggregates all costs, and parsing tenant IDs from user-agent headers is not a supported billing mechanism — Azure Cost Management cannot split costs by header values. Option B is wrong because Custom Vision project tags are metadata for organizing models, not a billing or cost-allocation mechanism. Option D is wrong because Azure Blob Storage cost allocation tags apply only to storage resources, not to Azure AI Services API consumption — you cannot attribute AI API costs via storage tags.

Full explanation →

367

MCQmedium

You are using Custom Vision and want to programmatically publish a trained iteration so that prediction endpoints can use it. Which API operation should you call?

A.GET /customvision/v3.3/Training/{projectId}/iterations/{iterationId}/make-default

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

C.PATCH /customvision/v3.3/Training/{projectId}/projects/{projectId}

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

AnswerD

The publish operation activates a trained iteration, making its model available to the prediction endpoint. The stem's constraint is programmatic publication of a specific iteration; POST to the iterations/{iterationId}/publish route satisfies this by promoting that iteration to live status, unlike training or prediction calls.

Why this answer

Publishing a trained Custom Vision iteration so prediction endpoints can consume it is done by calling POST /customvision/v3.3/Training/{projectId}/iterations/{iterationId}/publish. The publish operation makes the iteration available at the prediction URL, optionally under a named prediction resource. This is the required step before any prediction calls can target that iteration.

Exam trap

AI-103 often tests the confusion between 'make-default' (sets the default iteration for portal/API convenience) and 'publish' (exposes the iteration at a prediction endpoint) — candidates frequently pick make-default when the question asks about serving predictions.

How to eliminate wrong answers

Option A is wrong because GET on /make-default is not a valid operation — make-default is a POST action used to set the default iteration, and GET cannot perform state changes. Option B is wrong because the Prediction API path with /activate is not the publish mechanism; publishing is a Training API operation, and the Prediction endpoint consumes already-published iterations. Option C is wrong because PATCH on the project resource updates project metadata (such as name or description), not iteration publication state.

Full explanation →

368

Multi-Selecthard

Your enterprise requires strict security and governance over your Azure AI Vision resources. Which THREE security configurations should you implement? (Choose three.)

Select 3 answers

A.Store API keys in plain text configuration files on public GitHub repositories.

B.Disable TLS encryption to improve API response latency.

C.Configure Azure Private Endpoints to restrict access to your Virtual Network.

D.Enable diagnostic logging and stream audit records to Log Analytics.

E.Implement Microsoft Entra ID role-based access control (RBAC) instead of relying solely on API keys.

AnswersC, D, E

Azure Private Endpoints assign a private IP from your Virtual Network to the Vision resource, removing public internet exposure. This satisfies the strict network isolation requirement, since traffic traverses the Microsoft backbone rather than the public endpoint, and pairs with disabling public network access.

Why this answer

Option C is correct because Azure Private Endpoints create a private IP address for the Azure AI Vision resource inside your Virtual Network, eliminating exposure to the public internet and enforcing network-level isolation. Option D is correct because enabling diagnostic logging and streaming audit records to Log Analytics provides the governance and auditability required for strict security monitoring, retention, and compliance reporting. Option E is correct because Microsoft Entra ID RBAC provides fine-grained, identity-based authorization with conditional access and managed identities, which is more secure and governable than shared API keys alone.

Option A is wrong because storing API keys in plain text on public GitHub repositories exposes credentials to anyone and violates every security best practice. Option B is wrong because disabling TLS removes encryption in transit, exposing data to interception and violating compliance requirements; latency gains do not justify this risk.

Exam trap

AI-103 often tests the confusion between 'encryption in transit' being optional for performance versus mandatory for security, and the misconception that API keys alone provide sufficient access control.

Full explanation →

369

MCQmedium

You are configuring a Custom Vision project to classify images of plant diseases. You notice that some image files uploaded via the API are rejected because they exceed the maximum allowed file size. What is the maximum image file size supported by Custom Vision for training?

A.2 GB

B.500 KB

C.6 MB

D.50 MB

AnswerC

Custom Vision rejects training images larger than 6 MB, so files exceeding that limit fail upload. This is the documented per-image ceiling for training data; prediction images are capped at 4 MB, which is a separate limit.

Why this answer

Custom Vision enforces a maximum training image file size of 6 MB per image. Images larger than this are rejected by the API during upload, regardless of their dimensions or format. This limit applies to training images, not prediction images, which have their own constraints.

Exam trap

AI-103 often tests memorization of service limits, and the trap is confusing the training image size limit (6 MB) with prediction limits (4 MB) or with blob storage limits (much larger).

How to eliminate wrong answers

Option A is wrong because 2 GB is far beyond any Custom Vision limit and is closer to blob storage limits. Option B is wrong because 500 KB is too small — Custom Vision accepts images up to 6 MB, so this understates the actual limit. Option D is wrong because 50 MB exceeds the 6 MB training image cap and reflects a misunderstanding of the service's constraints.

Full explanation →

370

MCQeasy

You are provisioning an Azure AI services account and want to ensure that all telemetry and monitoring metrics are collected. Which built-in monitoring feature is automatically enabled on Azure AI resources without requiring additional agent installation?

A.Azure Monitor platform metrics

B.Log Analytics Windows security agent

C.Application Insights SDK code instrumentation

D.Azure Diagnostics extension for Linux

AnswerA

Platform metrics are emitted automatically by Azure AI resources into Azure Monitor with no diagnostic setting or agent required, satisfying the stem's no-installation constraint. Guest OS metrics and log-based telemetry need the Azure Monitor agent, so they cannot meet that requirement.

Why this answer

Azure Monitor platform metrics are collected automatically for Azure AI services (Cognitive Services, Azure OpenAI, AI Search, etc.) with no agent, SDK, or diagnostic extension required. They appear in the resource's Metrics blade within seconds to minutes and cover request counts, latency, errors, and token usage. This makes them the only truly zero-configuration monitoring feature among the choices.

Exam trap

AI-103 often tests the distinction between 'automatic' platform metrics and 'opt-in' features like diagnostic settings, Application Insights, and VM agents — candidates frequently pick Application Insights because it is the most familiar monitoring tool.

How to eliminate wrong answers

Option B is wrong because the Log Analytics Windows security agent is an optional VM extension for Windows machines, not an automatic feature of Azure AI PaaS resources. Option C is wrong because Application Insights requires explicit SDK instrumentation in application code or auto-instrumentation via a supported agent — it is not enabled by default on AI resources. Option D is wrong because the Azure Diagnostics extension for Linux is a VM-level agent that must be installed and configured, and it does not apply to PaaS AI services.

Full explanation →

371

MCQeasy

You are developing a mobile application that captures business card photos and extracts contact information. Which Azure AI Vision feature is specifically designed to extract printed and handwritten text lines and words?

A.Image Tagging

B.Dense Captions

C.Brand Detection

D.Read

AnswerD

Read uses the Optical Character Recognition (OCR) engine optimised for printed and handwritten text, returning extracted lines and words with bounding boxes and confidence scores. It satisfies the stem's requirement for a feature specifically designed to extract text lines and words from captured business card images, unlike captioning or tagging features.

Why this answer

The Read feature in Azure AI Vision is specifically designed for OCR, extracting printed and handwritten text lines and words from images. It is optimized for text-heavy images like business cards, documents, and receipts. This makes it the correct choice for capturing contact information from business card photos.

Exam trap

AI-103 often tests whether candidates confuse OCR (Read) with image description features (Dense Captions, Tagging), causing them to pick a captioning feature when the requirement is text extraction.

How to eliminate wrong answers

Option A is wrong because Image Tagging identifies objects, scenes, and concepts, not text. Option B is wrong because Dense Captions generates descriptive captions for image regions, not OCR text extraction. Option C is wrong because Brand Detection identifies commercial brands/logos, not general text.

Full explanation →

372

Multi-Selecthard

You are designing a secure, production-grade architecture for an Azure AI Search service integrated with Azure OpenAI. Which THREE security and operational best practices should you implement? Each correct answer presents part of the solution.

Select 3 answers

A.Publish administrative API keys publicly in GitHub repositories for easy developer access.

B.Configure Azure Monitor diagnostic settings to stream query performance and error logs to a Log Analytics workspace.

C.Use managed identities for secure authentication between Azure AI Search and Azure OpenAI without hardcoded keys.

D.Use the Free pricing tier for high-throughput production workloads.

E.Deploy Private Endpoints for both Azure AI Search and Azure OpenAI to keep data traffic within internal virtual networks.

AnswersB, C, E

Streaming query performance and error logs to a Log Analytics workspace satisfies the operational-visibility requirement of a production-grade design. Azure Monitor diagnostic settings capture Azure AI Search resource logs, enabling query latency analysis, throttling detection and alerting, which supports the monitoring and troubleshooting demands the scenario's production constraint imposes.

Why this answer

Option B is correct because Azure Monitor diagnostic settings let you route Azure AI Search query and operation logs, plus metrics, into a Log Analytics workspace, enabling KQL-based alerting, auditing, and performance troubleshooting required for production operations. Option C is correct because managed identities (system- or user-assigned) allow Azure AI Search to authenticate to Azure OpenAI and other resources via Microsoft Entra ID tokens, eliminating hardcoded API keys and enabling RBAC-scoped, credential-rotation-free access. Option E is correct because Private Endpoints create Azure Private Link connections that assign private IP addresses from your VNet to both services, so index/query traffic and Azure OpenAI calls traverse the Microsoft backbone rather than the public internet, and public network access can be disabled.

Option A is wrong because publishing admin API keys in public repositories exposes full control-plane access to the search service and is a severe credential leak. Option D is wrong because the Free tier caps storage, index count, and throughput (no SLA, limited partitions/replicas), so it cannot support high-throughput production workloads.

Exam trap

AI-103 often tests the misconception that API keys are acceptable for production if rotated, when the exam expects managed identity plus Private Endpoint as the keyless, network-isolated answer.

Full explanation →

373

MCQmedium

Your enterprise team is deploying multiple Azure AI Custom Vision resources across different development stages (Dev, Test, Prod). You need to ensure that billing costs for the Dev environment are billed to a specific cost center. Which Azure feature should you use to categorize these costs?

A.Azure role-based access control (RBAC)

B.Azure resource tags

C.Azure Service Health alerts

D.Azure Firewall rules

AnswerB

Azure resource tags apply key-value metadata directly to each Custom Vision resource, letting cost management group and filter spend by cost centre. Tagging the Dev resource satisfies the requirement to attribute its billing to a specific cost centre.

Why this answer

Azure resource tags are key-value pairs applied to resources and resource groups that flow into the Azure Cost Management and Billing data, enabling cost categorization, filtering, and chargeback/showback by dimensions like CostCenter, Environment, or Project. Tagging the Dev Custom Vision resources with a CostCenter tag lets finance attribute those costs to the correct cost center.

Exam trap

AI-103 often tests the confusion between governance/authorization features (RBAC, Policy) and cost-attribution features (tags) — candidates pick RBAC thinking it controls billing, missing that tags are the billing categorization mechanism.

How to eliminate wrong answers

Option A is wrong because Azure RBAC controls who can perform actions on resources (authorization), not how costs are categorized or reported — RBAC has no bearing on billing attribution. Option C is wrong because Azure Service Health alerts notify you about platform incidents, planned maintenance, and advisories; they do not categorize or attribute costs. Option D is wrong because Azure Firewall rules control network traffic (allow/deny, DNAT, application rules) and have nothing to do with cost management or billing categorization.

Full explanation →

374

MCQhard

You are configuring an Azure AI Document Intelligence orchestration model to automatically classify and extract data from incoming customer correspondence, which includes invoices, work orders, and purchase orders. One of the component models is failing to classify documents correctly. How should you resolve this issue?

A.Modify the orchestration model routing threshold in the Azure portal settings.

B.Switch the orchestration model from a template-based classifier to a neural classifier using the REST API.

C.Retrain the failing component custom model with more distinct sample documents that represent the specific document class.

D.Update the API version to bypass classification and force direct extraction via the general document model.

AnswerC

Retraining the failing component custom model directly addresses misclassification by supplying more distinct samples of the specific document class, sharpening the model's learned feature boundaries. Since orchestration routes documents to component models, each classifier's accuracy depends on representative training data; additional distinct samples reduce ambiguity between invoices, work orders and purchase orders.

Why this answer

The most effective way to resolve misclassification by a component model in an orchestration model is to retrain that custom model with more distinct sample documents that clearly represent the specific document class. This improves the model's ability to differentiate between classes.

Exam trap

AI-103 often tests troubleshooting steps for custom models; candidates might incorrectly think they can adjust routing thresholds or switch model types via API, rather than addressing the training data quality.

How to eliminate wrong answers

Option A is wrong because the orchestration model routing threshold is not a configurable setting in the Azure portal; orchestration uses the component models' own confidence scores. Option B is wrong because switching from template-based to neural classifier is not a simple API switch and may not address the root cause of poor training data. Option D is wrong because updating the API version does not bypass classification; orchestration always classifies first, and forcing extraction via general document model would lose the benefits of specialized models.

Full explanation →

375

Multi-Selecthard

You are designing a generative AI application in Azure AI Foundry that supports multi-turn conversations with memory. Which THREE components are required to properly maintain conversational state and context? (Choose THREE)

Select 3 answers

A.Thread or session state storage to persist message history across turns

B.Permanently deleting all messages after every single API turn

C.Setting temperature to 0.0 to erase all memory between calls

D.Message history appending logic in the prompt construction payload

E.Context window management strategy (such as summarization or truncation) to prevent token overflow

AnswersA, D, E

Thread or session state storage persists the ordered message history across turns, keyed to a conversation identifier. Without durable state, each turn starts context-free and the model cannot reference earlier exchanges, so this component satisfies the stem's requirement for multi-turn memory and context retention.

Why this answer

Option A is correct because maintaining conversational state requires a thread or session store that persists the message history across turns, allowing the model to reference earlier exchanges in a multi-turn dialogue. Option D is correct because even with stored history, the application must append prior messages into the prompt payload sent to the model so the context is actually available during inference. Option E is correct because as history grows it will exceed the model's context window, so a strategy such as summarization, truncation, or sliding-window management is needed to prevent token overflow and keep requests valid.

Option B is incorrect because deleting all messages after every turn destroys the memory that multi-turn conversations depend on. Option C is incorrect because temperature controls randomness of sampling, not memory, and setting it to 0.0 does not erase or manage conversational state.

Exam trap

AI-103 often tests whether candidates mistake sampling parameters (temperature) or destructive actions (deleting messages) for memory management, so select the three options that describe storage, prompt assembly, and context-window control.

Full explanation →

Microsoft Certified: Azure AI Apps and Agents Developer Associate (AI-103) (AI-103) (AI-103) — Questions 301–375