Practice Microsoft AI-103 Azure AI Apps and Agents Developer Associate with free sample questions, timed mock exams, topic drills, and detailed explanations in IT Mastery.
AI-103 is Microsoft’s Azure AI apps-and-agents route for candidates who build, manage, and deploy AI solutions that use Microsoft Foundry, Azure AI services, generative AI, agents, retrieval, multimodal workflows, and Python-backed implementation patterns.
IT Mastery practice for AI-103 is live now. Use this page to start the web simulator, review the exam snapshot, work through 24 public sample questions, and continue into the full question bank with the same account on web, iOS, iPadOS, macOS, or Android.
Start a practice session for Microsoft Azure AI Apps and Agents Developer Associate (AI-103) below, or open the full app in a new tab. For the best experience, open the full app in a new tab and navigate with swipes/gestures or the mouse wheel—just like on your phone or tablet.
Open Full App in a New TabA small set of questions is available for free preview. Subscribers can unlock full access by signing in with the same account they use on web and mobile.
Prefer to practice on your phone or tablet? Download the IT Mastery – AWS, Azure, GCP & CompTIA exam prep app for iOS or IT Mastery app on Google Play (Android) and use the same account across web and mobile.
| Domain | Weight |
|---|---|
| Plan and manage an Azure AI solution | 25-30% |
| Implement generative AI and agentic solutions | 30-35% |
| Implement computer vision solutions | 10-15% |
| Implement text analysis solutions | 10-15% |
| Implement information extraction solutions | 10-15% |
These sample questions are drawn from the current local bank for this exact exam code. Use them to check your readiness here, then continue into the full IT Mastery question bank for broader timed coverage.
Topic: Plan and Manage an Azure AI Solution
An HR policy agent in a Foundry project uses a retrieval tool to query an existing Azure AI Search index and a function tool to create approval records. The web app uses managed identity web-mi to call the agent, and the agent tools run under the project managed identity project-mi. Developers must update agent instructions and tool definitions but must not grant access. Employees use only the web app. You need keyless, least-privilege role assignments. Which role map should you use?
web-mi: Azure AI User on the project; developers: Azure AI Developer; project-mi: Search Index Data Reader and function invoke-only; employees: no direct roles.web-mi: Azure AI Developer; developers: Azure AI Administrator; project-mi: Search Service Contributor and Function App Contributor; employees: no direct roles.web-mi: Azure AI User; employees: Azure AI User on the project; tool calls use each employee’s Search and function roles.web-mi: Azure AI User; developers: Azure AI User; store Search and function keys in the agent tool configuration.Best answer: A
Explanation: Least privilege separates runtime, authoring, and end-user responsibilities. The web app needs only project invocation, developers need authoring without access management, and the project identity needs only data-plane rights for the Search retrieval tool and function invocation. Employees do not need direct roles on Foundry or data sources when they only use the app. In a Foundry agent solution, assign roles to the identity that performs each action and scope them to the resource needed. The app identity only calls the agent, so a project user/invocation role is sufficient. Developers who edit prompts, tools, and evaluations need a development role, not an administrator role that can manage access. The project managed identity is the runtime identity for tool execution, so it needs read-only access to the Search index and invoke-only access to the approval function. End users who access the agent only through the app should not receive direct Foundry, Search, or function permissions.
Topic: Implement Information Extraction Solutions
An Azure workflow uses a Foundry agent tool to answer questions about contract obligations. After adding a Content Understanding analyzer and Azure AI Search retrieval, evaluators report two symptoms: answers cannot cite source pages, and users sometimes receive obligations from contracts outside their security group.
Trace excerpt:
User groups: Sales-East
Search filter: customer eq 'Contoso'
Top hit metadata: aclGroups=['Legal'], source='contracts/contoso.pdf', page=7
Tool result sent to agent:
customer='Contoso'
renewalNoticeDays=60
chunkId='contoso-0007'
What is the best next fix?
Best answer: C
Explanation: The failure is in the retrieval-to-tool handoff. The workflow retrieves a Legal-only chunk for a Sales-East user and sends only extracted values plus a chunk ID, so the agent lacks both enforceable access context and citation evidence. Retrieval pipelines that feed workflows or agent tools should preserve both evidence and governance metadata. In this trace, the Azure AI Search filter uses only the customer name and does not constrain results by the user’s access-control metadata. The tool output also removes the retrieved passage and source/page citation details. The fix is to apply the user’s ACL constraints during retrieval, then pass the retrieved passage, extracted field values, citation/provenance fields, and relevant ACL metadata to the workflow or agent tool. This gives the model grounded evidence and gives downstream logic enough information to enforce access and cite sources. Increasing recall or relying on prompt instructions cannot repair missing access filters or omitted provenance.
Topic: Implement Computer Vision Solutions
An advertising team builds a Foundry app to create 6-second product videos. Approved footage is stored in Blob Storage and indexed in Azure AI Search with brand, productId, and approvalStatus metadata. The current workflow retrieves an approved clip record but sends only a text prompt to the video model, so outputs sometimes change product geometry and motion. The team must keep the retrieved footage as the visual grounding source while changing only the background. Which platform control should you use?
Best answer: A
Explanation: The requirement is visual grounding, not just text retrieval. Passing the approved clip as the source or reference video for a video editing operation keeps the product appearance and motion anchored while applying the requested background change. For video generation and editing workflows, use the platform control that matches the intended conditioning source. A text-to-video prompt can describe the approved clip, but it does not force the model to preserve the exact product geometry or motion. When the retrieved asset is the trusted grounding source, the workflow should pass that video into a video editing or reference-video operation and use the prompt only to describe the allowed transformation. Retrieval still helps select the right approved asset, but the grounding quality comes from using the retrieved video as the edit source.
Topic: Implement Generative AI and Agentic Solutions
A Microsoft Foundry project contains an HR assistant that uses Azure AI Search for RAG over policy PDFs. The latest evaluation produced this summary:
Groundedness: 0.43 (target: >=0.80)
Retrieval precision@5: 0.31
Failure pattern: current policy PDFs are indexed, but superseded versions appear in top citations.
Indexed metadata: policyId, department, versionStatus, effectiveDate
The assistant must answer only from current policy versions. Which change should you make?
versionStatus=current metadata filter during retrieval.Best answer: C
Explanation: The evaluator results point to a retrieval and grounding problem, not a model capacity or safety problem. Current documents are already indexed, but stale versions are being retrieved and cited, so retrieval must filter to current policy versions. Groundedness evaluation should guide which part of the RAG pipeline to adjust. Here, low groundedness and low retrieval precision@5 are tied to a clear failure pattern: superseded documents are appearing in the top retrieved citations even though current documents exist in the index. Because the index includes versionStatus and effectiveDate, the appropriate fix is to constrain retrieval to current versions before the model generates an answer. Prompt changes or a larger model can improve wording, but they do not reliably remove stale grounding context. The key takeaway is to fix the retrieval evidence when evaluator results show the wrong sources are being supplied.
Topic: Implement Text Analysis Solutions
Your team uses a Microsoft Foundry agent to translate agent-drafted support replies and post them to a CMS. A failed trace shows:
Input: Reset the Contoso KeyVault appliance.
Tool args: {"targetLanguage":"fr"}
Output: "Réinitialisez le coffre-fort..."
CMS error: expected {"locale":"fr-CA","bodyHtml":...}
Reviewer: In hardware docs, "KeyVault" must remain untranslated.
Tickets often omit locale, domain, and output contract. Which change should you make?
Best answer: D
Explanation: The failure is not just a translation quality issue; the agent lacks required workflow context. A pre-call gate should collect or require the target locale, apply domain terminology, and validate the structured output expected by the CMS. For a translation agent, tool calling should be governed by a schema and orchestration step that prevents incomplete requests. If the target locale is fr-CA, the domain is hardware documentation, and the CMS requires HTML in a JSON shape, those values should be explicit inputs or collected through clarification before translation. The agent can then retrieve approved terminology, pass the relevant context to the translation step, and validate the final structured response before posting it. Relying on inference or post-translation cleanup makes the workflow nondeterministic and allows the same failure to recur.
Topic: Plan and Manage an Azure AI Solution
A team deploys a Foundry agent for contract review. The approved security boundary includes private connectivity to the Foundry project, Azure AI Search, Storage, and two internal tool APIs. Before production, the security team asks for an observability check that verifies model, retrieval, and tool traffic do not leave this boundary. Which check should you implement?
Best answer: B
Explanation: The goal is to validate the security boundary, not model quality or user-content safety. Correlating Foundry agent traces with network egress or DNS logs shows which model, retrieval, and tool calls occurred and whether their destinations stayed on the approved private paths. For a private Foundry deployment, observability should prove that runtime calls use only approved resources and routes. Agent traces identify the model, retrieval source, and tool invocation path; network egress, DNS, firewall, or private endpoint logs confirm where traffic actually went. Together, these logs can detect a misconfigured connector, tool endpoint, or fallback route that sends traffic to a public endpoint outside the boundary. Quality evaluations such as groundedness and relevance are useful, but they do not validate the data path. The key takeaway is to monitor both application-level traces and network-level destinations when the requirement is containment.
Topic: Implement Information Extraction Solutions
A team builds a Foundry agent that answers compliance questions from uploaded vendor invoices and inspection forms. The documents include scanned PDFs, tables, check boxes, repeated line items, and labeled values such as vendor, invoice total, and inspection date. The current RAG index uses plain OCR text chunks, and answers often cite the right page but pair values with the wrong labels. Which ingestion change best fixes the grounding problem?
Best answer: A
Explanation: The issue is not only text recognition; it is loss of document structure. Forms and complex documents need OCR, layout analysis, and field extraction combined so the RAG index contains grounded field-value relationships and source references. For scanned forms, invoices, and complex PDFs, OCR alone extracts words but may not preserve how labels, values, tables, check boxes, and repeated line items relate to each other. A Content Understanding analyzer can combine OCR, layout analysis, and field extraction to produce structured or markdown representations that preserve document structure. Indexing those extracted fields, table content, and page/source metadata gives the Foundry agent retrievable evidence that is grounded to the correct document locations. Chunking or ranking can improve retrieval over text, but it cannot reliably reconstruct missing field-value relationships after ingestion.
Topic: Implement Computer Vision Solutions
You are building a moderation step in a Microsoft Foundry project for a multimodal customer-support agent. The step receives an uploaded image and caption, then stores policy evidence with the conversation trace. You must implement the action returned by the policy engine for this item.
Policy:
Evidence:
image.violence.severity = 2
image.violence.evidence = "minor blood on hand"
image.sexual.severity = 0
text.self_harm.severity = 0
text.hate.severity = 0
evaluator.uncertainty = false
caption = "How do I clean this small cut?"
Which implementation should you use?
Best answer: D
Explanation: The policy maps severity 1 or 2 evidence to a flag when no block or review condition applies. The image has low-severity violence evidence, no self-harm request, and no evaluator uncertainty, so the workflow should continue while preserving the safety flag in the trace. The core decision is policy-evidence routing for multimodal content. The evaluator found image violence at severity 2, which is nonzero evidence but below the review and block thresholds. Because uncertainty is false and no category has severity 3, 4, or 5, the item should not be escalated or blocked. Storing the flag with the trace preserves auditability for monitoring and later safety analysis while allowing the customer-support flow to proceed. The key takeaway is to apply the highest applicable policy action from the evidence, not to over-block low-severity content or ignore it entirely.
Topic: Implement Generative AI and Agentic Solutions
You are implementing a Foundry agent that answers support questions by reading CRM records and can create a billing refund when a case is eligible. Policy requires keyless access over private networking, autonomous read-only lookups, human approval before any refund action, and audit evidence showing which tool outputs grounded each response. Which implementation best meets the requirements without blocking legitimate use?
Best answer: B
Explanation: Tool-augmented generation should expose external data and actions through controlled tools, not through prompts or broad credentials. The best design uses managed identity, private connectivity, role-scoped tools, enforced approval for refunds, and trace/provenance logging for auditability. For a Foundry tool-augmented flow, security controls should be enforced at the tool and workflow layer. Read-only CRM lookup can run autonomously with a least-privileged managed identity over private networking. The refund action is higher risk, so the tool invocation should require a human approval workflow before execution. Trace logging and provenance metadata provide evidence of which tool outputs grounded the final response. Safety filters are useful for generated content, but they do not replace authorization, approval, or audit controls for external actions.
Topic: Implement Text Analysis Solutions
You are building a Foundry-based support triage app. Compliance requires detecting hate, self-harm, sexual, violent, and sensitive personal data before messages are logged or sent to tools. The app must still accept frustrated customer complaints that are negative in tone but not unsafe. Which design should you implement?
Best answer: B
Explanation: Sentiment analysis measures tone or opinion, not whether text violates safety policy or contains sensitive data. The requirement is to detect explicit safety categories and sensitive content without blocking ordinary negative complaints, so dedicated safety and sensitive-content detection is required. The core concept is separating sentiment from risk detection. A customer can express strong negative sentiment in a legitimate complaint, and a positive or neutral message can still contain hate, self-harm intent, sexual content, violence, or sensitive personal data. Use text safety/content moderation classifiers for the safety categories and sensitive-content or PII detection before logging or tool invocation. Sentiment can still support analytics, prioritization, or customer-experience reporting, but it should not be the control that blocks, redacts, or routes risky content. The key takeaway is that moderation should be based on policy-risk signals, not emotional tone.
Topic: Plan and Manage an Azure AI Solution
A legal department is building a Microsoft Foundry agent that answers employee questions from policy documents in a private storage account. New approved documents can be published several times per day, and answers must reflect the latest source content with citations without retraining or redeploying the model. Which implementation should you choose?
Best answer: C
Explanation: The scenario requires fresh content and citations, which is a retrieval-time grounding problem. Keeping the model deployment unchanged while refreshing an Azure AI Search index lets the Foundry agent retrieve current policy chunks at answer time. Retrieval-time grounding is the right design when source facts change frequently. In a Foundry agent, policy files can be ingested into an Azure AI Search index and connected as agent knowledge or a retrieval tool. At runtime, the agent retrieves relevant chunks from the current index and uses them to ground the response, including citations or provenance metadata. Fine-tuning is better for changing model behavior, tone, or task patterns, not for repeatedly loading fast-changing facts that must remain traceable to source documents. The key distinction is that retrieval updates the external knowledge source, while fine-tuning changes the model itself.
Topic: Implement Information Extraction Solutions
Your Foundry project uses Azure AI Search to ground a document-review agent over native PDFs and scanned PDFs. Offline answer evaluation reports low groundedness for questions whose answers are in scanned invoices. You need an observability approach that identifies whether the relevance problem is caused by missing OCR, weak enrichment, wrong search mode, or poorly populated index fields. Which approach should you use?
Best answer: B
Explanation: Retrieval relevance problems must be measured at the retrieval layer, not only at the generated-answer layer. A retrieval evaluation set with known relevant documents plus trace data can show whether scanned content was extracted, enriched, indexed, and searched correctly. The core concept is retrieval observability for an Azure AI Search-backed RAG workflow. Because the failures are concentrated in scanned invoices, the evaluation should compare queries to expected source documents and capture retrieval traces before generation. Useful evidence includes top-k retrieved chunks, scores, query/search mode, OCR text coverage, enrichment outputs, and populated index fields. Segmenting results by scanned versus native files helps isolate whether the issue is missing OCR, weak enrichment, an unsuitable semantic/vector/hybrid search mode, or poor field design. Final-answer evaluation can confirm a symptom, but it cannot reliably identify which retrieval pipeline stage failed.
Topic: Implement Computer Vision Solutions
A Foundry project processes uploaded product images and user captions. The safety policy says: block high-severity visual sexual content or graphic violence; send to human review when age-related uncertainty is detected with sexual content at medium or higher; flag-only medium hate or harassment without other violations; allow only low or none with no uncertainty. A monitoring trace shows sexual=medium, graphic violence=none, age-related uncertainty=detected, and hate/harassment=low. Which disposition should the safety gate emit?
Best answer: D
Explanation: The disposition should follow the logged policy evidence. Medium sexual content combined with age-related uncertainty triggers human review under the stated policy. A hard block is reserved for high-severity visual sexual content or graphic violence. Responsible AI monitoring for multimodal workflows should record the action that matches the applicable policy rule. The trace shows sexual content at medium severity and age-related uncertainty, which the policy explicitly routes to human review. There is no high-severity sexual content or graphic violence, so blocking would exceed the stated rule. Flag-only is reserved for medium hate or harassment without other violations, and allow requires low or none across the evidence with no uncertainty. The key is to apply the policy evidence directly and choose the least-permissive matching disposition.
Topic: Implement Generative AI and Agentic Solutions
Your team is deploying a Foundry project for an internal policy assistant. The agent uses an LLM deployment grounded by Azure AI Search over policy PDFs, runs through private endpoints with managed identity, and is promoted by CI/CD. Before a response is shown, the business wants a visible quality-control record for groundedness, citation coverage, and safety; failed checks must route to a human approver. The record can include a concise critique and scores, but must not expose raw hidden reasoning to users. Which architecture is the best fit?
Best answer: D
Explanation: A reflection or self-critique stage is the best fit because it produces structured, visible quality signals before the answer is released. Combining that stage with Foundry evaluations, trace logging, CI/CD gates, and human approval meets the runtime and operational requirements without exposing hidden reasoning. The core concept is operationalizing generative AI quality control with visible evaluation artifacts. For a RAG-backed agent, a reflection stage can ask a model or evaluator to score the proposed answer against a rubric such as groundedness, citation coverage, completeness, and safety. Those scores and the concise critique can be logged as traces, used in CI/CD regression evaluations, and used at runtime to route failed responses to human approval. The important boundary is that the system should expose evaluation summaries and decisions, not raw hidden chain-of-thought. Safety filtering alone is useful, but it is not a substitute for groundedness and citation-quality checks.
Topic: Implement Text Analysis Solutions
You are building a Foundry app that triages customer emails in French, Spanish, and Japanese. The Azure AI Content Safety text check supports these source languages. The validated extraction and summarization prompts use English-only examples and return canonical English labels. Policy requires unsafe user content to be blocked before any transformation or domain LLM call. Which TWO ordering choices should you implement?
Best answer: D
Explanation: The pipeline must preserve the safety policy first, then satisfy the language assumptions of the domain prompts. Because safety detection supports the incoming languages and must happen before transformations, scan the original text first. After content is allowed, translate non-English input to English for the validated extraction and summarization prompts. Translation order depends on which component has the stricter language or governance requirement. Here, safety detection is both multilingual for the workload and required before any transformation, so it should run against the original user message. Once the message is allowed, translation to English is appropriate because the extraction and summarization prompts were validated only in English and produce canonical English labels. This reduces prompt drift and keeps downstream routing consistent. The key distinction is that safety policy controls should not be delayed by translation when the safety detector can process the source language.
Topic: Plan and Manage an Azure AI Solution
You are planning a Microsoft Foundry customer support agent. The agent already retrieves troubleshooting articles from Azure AI Search. A new workflow lets customers reschedule repair visits. Available slots change during the day, and the agent must book the selected slot only after the customer confirms. Which integration choice should govern this workflow?
Best answer: A
Explanation: Tool calls are the right grounding mechanism when the agent must access live system state and take an external action. The confirmation requirement also fits a safeguarded function-calling pattern before committing the booking. For agent planning, choose the grounding method based on the work being done. Retrieval and knowledge integration ground answers in documents. Conversation memory helps preserve user preferences or prior context. A typed tool call is used when the agent must query a live operational system or change state, such as checking current appointment availability and booking a selected repair visit. Because the slots change during the day, indexed exports can become stale. Because the workflow books an appointment, the agent should call a scheduling API through a constrained tool schema and require explicit customer confirmation before the write operation.
Topic: Implement Information Extraction Solutions
A legal department is building a Foundry-based assistant that answers employee questions from 80,000 scanned contracts. Users ask in natural language and often do not use exact clause titles. Answers must be grounded with citations to source pages. Ingestion must extract text from scans, and runtime access must use managed identity over private endpoints. Which architecture should you recommend?
Best answer: D
Explanation: The scenario requires more than keyword matching because users ask natural-language questions that may not share exact terms with the contracts. OCR/layout enrichment plus Azure AI Search hybrid vector and semantic retrieval provides semantically relevant chunks, page metadata for citations, and secure Foundry connectivity. For scanned contract Q&A, the ingestion pipeline should extract text and layout, preserve page metadata, chunk the content, generate embeddings, and index the chunks in Azure AI Search. Vector search handles semantic similarity when users use different wording, while hybrid and semantic ranking can combine exact term matching with meaning-based relevance. The Foundry app can then ground generative answers on retrieved chunks and cite source pages from stored metadata. Managed identity and private endpoints meet the access constraint without exposing keys or public network paths. Keyword-only retrieval is the closest distractor, but it is too brittle for semantic similarity and grounded generative answers in this scenario.
Topic: Implement Computer Vision Solutions
In a Microsoft Foundry project, a marketing app accepts uploaded images and images generated by a model. Before publishing, an automated gate must enforce the visual policy: block inappropriate visual content, require the approved brand watermark, and reject prohibited symbols. The gate must inspect the image itself for every asset.
Which TWO actions should you include?
Best answer: A
Explanation: The gate needs two kinds of visual checks: built-in safety moderation and organization-specific visual policy validation. Azure AI Content Safety is appropriate for inappropriate image content, while a custom Azure AI Vision detector can enforce watermark, brand, and prohibited-symbol rules. For a pre-publish visual policy gate, inspect every candidate image after upload or generation. Azure AI Content Safety image analysis is the right control for harmful or inappropriate visual content categories. Organization-specific rules, such as required watermarks, approved brand marks, or banned symbols, require visual detection against the actual image pixels, which can be implemented with a custom Azure AI Vision detector or equivalent visual analysis step. This is stronger than relying only on generation-time filters because uploaded images bypass those filters.
Topic: Implement Generative AI and Agentic Solutions
You are designing a Microsoft Foundry workflow for warranty claims. The agent must ground answers in internal policy documents, check groundedness and policy compliance before replying, and create a replacement order only after the claim passes checks or receives human approval for an exception. Which implementation best fits the requirements?
Best answer: C
Explanation: The workflow needs control points between evidence gathering, response creation, quality checks, and business actions. Splitting retrieval, generation, evaluation, and action lets Foundry enforce groundedness and approval requirements before the order tool runs. For agentic workflows, split steps when each stage has a different responsibility or risk level. Retrieval should collect trusted policy evidence, generation should draft the response from that evidence, evaluation should verify groundedness and compliance, and the action step should be gated by the evaluator result and human approval rules. This design supports traceability, safer tool invocation, and clearer failure handling. A single prompt can be convenient for prototypes, but it does not provide reliable enforcement points for checks before a replacement order is created.
Topic: Implement Text Analysis Solutions
A company is building a multilingual support agent in a Microsoft Foundry project. User messages are summarized, translated, stored in Azure AI Search for retrieval, and sometimes sent to a refund tool. Compliance requires phone numbers, email addresses, and payment card numbers to be removed before any storage, summarization, translation, or tool call. The solution must use private networking and managed identity. Which architecture should you use?
Best answer: A
Explanation: The deciding requirement is that sensitive content must be detected and removed before any downstream processing. An ingress PII redaction step using Azure AI Language keeps raw sensitive text out of storage, summaries, translations, retrieval indexes, and agent tool calls. For this workload, sensitive-content detection must be a pre-processing gate, not a later audit step. Azure AI Language PII detection/redaction is the Azure-native text analysis capability for identifying and redacting items such as phone numbers, email addresses, and payment card numbers. Placing it at ingestion means downstream components receive only sanitized text. The same component can authenticate with managed identity and access the service through private endpoints, aligning with the networking and credential requirements. The key takeaway is to prevent exposure before persistence or model/tool execution, rather than trying to clean up after the agent has already processed the text.
Topic: Plan and Manage an Azure AI Solution
In a Foundry project, you need to build an internal support assistant. It must hold a multi-turn conversation, retrieve policy passages from Azure AI Search, choose between HR and IT ticket APIs based on user replies, and require manager approval before submitting high-risk requests. The exact sequence cannot be predetermined. Which implementation should you choose?
Best answer: C
Explanation: The scenario requires dynamic decision-making across turns, retrieval grounding, tool selection, and approval controls. An agent-based solution is the best fit because it can use a model to reason over conversation context and invoke governed tools when needed. Use an agent-based solution when the task requires goal-directed behavior across a conversation, especially when the next action depends on user responses or retrieved knowledge. In this case, the assistant must maintain conversation context, ground responses with Azure AI Search, select between different tool calls, and apply a manager approval safeguard before high-risk actions. A deployed model supplies language capability, and Foundry Tools expose callable actions, but the agent coordinates them under policies and state. A fixed workflow is better when the process path is predictable and deterministic.
Topic: Implement Information Extraction Solutions
A Foundry project uses Azure Content Understanding to process scanned customs forms. The forms contain tables, checkboxes, handwritten notes, and fields whose meaning depends on nearby labels. Before replacing manual review, the team must validate that the analyzer handles the complete document structure, not only visible text. Which evaluation best measures this goal?
Best answer: B
Explanation: Complex forms require more than OCR. The evaluation should compare the analyzer output with labeled ground truth for recognized text, document layout, and extracted fields because the business goal depends on all three working together. For forms and complex documents, OCR reads the text, layout analysis identifies structure such as tables, selection marks, and label-value relationships, and field extraction maps that content to required business fields. An evaluation based on labeled ground truth should verify all three layers because a document can have accurate text but still fail if table rows, checkboxes, or nearby labels are misinterpreted. This is the best fit before replacing manual review, where both structure and field values must be reliable. Monitoring only one layer can hide failures in the others.
Topic: Implement Computer Vision Solutions
An inspection agent in a Microsoft Foundry project receives equipment photos and short video clips. The app must identify required safety objects and warning-label regions in the original media, return labels with bounding regions and video timestamps as JSON, and keep provenance for human review. It must not generate, edit, or synthesize visual media. Which TWO implementation choices meet the requirement?
Best answer: C
Explanation: The requirement is to understand and extract evidence from existing media, not to create media. Content Understanding and Azure AI Vision are appropriate because they analyze images or frames and can return structured visual metadata such as regions, text, labels, and timestamps. Visual inspection workloads should use visual understanding or extraction capabilities that analyze the original media and return evidence metadata. Content Understanding analyzers can be designed for image and video inputs to produce structured outputs, such as identified elements, regions, and timestamps, for downstream review. Azure AI Vision image analysis can detect objects and read text; applying it to photos and sampled video frames lets the app persist bounding boxes and frame times without generating new media. Image generation and editing workflows are for creating or modifying pixels, not for authoritative extraction from uploaded evidence.
Topic: Implement Generative AI and Agentic Solutions
A finance company is adding a generative workflow to a Python claims application. A Foundry project already contains an approved chat model deployment and a connection to an Azure AI Search index of policy documents. Requirements: ground answers in the index, open a Dynamics 365 case only after an adjuster approves the draft, use managed identity over private networking, and collect trace logs for evaluation. Which architecture is the best fit?
Best answer: B
Explanation: The best design integrates the application with the existing Foundry project by using the Foundry SDK. A project-hosted agent or workflow can use Azure AI Search for grounding, call an approved connector only after human approval, and preserve managed identity, private networking, and trace observability. The core concept is integrating a generative workflow through Foundry rather than rebuilding the AI stack or bypassing governance. The Python application should call the Foundry-hosted workflow or agent by using the Foundry SDK. The agent can use the existing model deployment, the Azure AI Search connection for retrieval-grounded responses, and the Dynamics connector as a controlled action. The application or workflow should gate the connector call until the adjuster approves the draft. Managed identity and private networking satisfy keyless secure access, while Foundry tracing captures prompts, retrieval, and tool calls for evaluation. Direct endpoint calls or custom model hosting miss the managed Foundry workflow and connector model.