AI-103 Developing AI Apps and Agents on Azure Questions and Answers

Guardrails: Select User input, Output, Tool response, and Tool call and set Action to Block.

Storage access: A system-assigned managed identity that is assigned the Storage Blob Data Contributor role

The guardrail must be applied to User input, Output, Tool response, and Tool call with the action set to Block . Microsoft Foundry guardrails support four intervention points: user input, tool call, tool response, and output. This scenario includes user-provided screenshots, a ticketing tool that uploads images and returns blob URLs, and final agent responses. Applying blocking controls at all four points ensures harmful image-related content is inspected throughout the agent run and prevented from continuing or being returned to the user. Microsoft’s guardrails guidance also states that tool call and tool response controls are specifically required when harmful content can pass through agent tools.

For storage, configure the Azure AI Content Safety resource with a system-assigned managed identity and grant it Storage Blob Data Contributor on the storage account or container. The Content Safety image moderation quickstart states that images can be supplied by blob storage URL and that the Content Safety resource must be given storage access by enabling its system-assigned managed identity and assigning Storage Blob Data Contributor or Owner; Contributor is the least-privileged valid option shown. Reference topics: Foundry guardrails, agent intervention points, image moderation, managed identity, and Azure Storage RBAC.

The PII analysis is explicitly restricted to Person and PhoneNumber categories. Therefore, only entities belonging to those requested classes are selected for the configured entity-mask redaction.

The word Contact is normal surrounding text and is not a Person or PhoneNumber entity. The SSN is a different PII category and is not included in the configured category list. Therefore, the first statement is No .

For the same reason, the email address and SSN remain in the resulting redactedText because neither category is requested. The second statement is therefore Yes .

John Doe is a Person entity and 312-555-1234 is a PhoneNumber entity. With the entityMask redaction policy, those selected spans are replaced by masks that identify the relevant entity types, so the third statement is Yes .

The code also builds an audit collection from the entities returned by the analysis operation, allowing the application to separately record what was detected while sending only the redacted string to the language model.

Study Guide alignment: configure detection of sensitive content and prepare safe textual inputs for downstream generative models . Microsoft Learn

====================

The correct answer is C. Split the mixed-language segments into single-language segments and translate each segment separately . Azure Translator has a known limitation for text that contains mixed-language input within the same sentence or segment. Microsoft’s Translator known-issues guidance states that the Text translation API does not support translating sentences that contain mixed-language text and that translations can be incorrect or incomplete when a single sentence includes multiple languages. The documented workaround is to specify the intended source language, remove the mixed-language sentence, or split the text into single-language segments.

In this scenario, specifying English for all segments would fail when Spanish content is present. Automatic language detection can detect a dominant language, but it does not reliably solve a segment that contains both English and Spanish. Document translation is also not the right fix because the issue is mixed-language content inside transcript segments, and Microsoft notes similar incomplete results for mixed-source-language documents. Splitting the transcript into language-homogeneous segments ensures each translation request has a clear source language and lets the entire transcript be translated reliably. Reference topics: Azure Translator in Foundry Tools, text translation, language detection, mixed-language input, and transcript preprocessing.

Comprehensive and Detailed 150 to 250 words of Explanation From Azure AI Apps and Agents Developer/Course Guide/topics:

The system message is an example of priming because it establishes contextual expectations before the user supplies the actual mathematical problem.

The first sentence assigns the model a role: an assistant specialized in mathematical puzzles. The second establishes the expected communication style and audience level by requiring explanations suitable for a four-year-old. These instructions influence subsequent responses without modifying model weights.

It is not few-shot learning because no input/output examples are provided. Few-shot prompting would include several representative puzzles together with correctly formatted answers.

It is not chain-of-thought because the prompt does not explicitly provide or request a structured sequence of intermediate reasoning steps. An affordance is a cue that indicates a possible behavior or completion, but role and behavioral context established through a system message is most accurately described by the available option priming .

This technique is commonly used to establish persona, domain boundaries, tone, output style, and operating assumptions for a generative application.

The AI-103 Study Guide includes prompt engineering and adjustment of model behavior as part of optimizing generative AI systems. Microsoft Learn

====================

The correct configuration is D. Enable mask_inpainting and supply both the input image and a mask indicating which part of the image to modify . The requirement is not to generate a new image, but to edit a specific region of an existing image while preserving the surrounding lighting, composition, and style. Azure OpenAI image editing in Microsoft Foundry supports modifying existing images by submitting an input image plus a prompt. For masked edits, the mask explicitly defines the part of the image the model is allowed to change; Microsoft states that the mask parameter defines the area to edit and must match the input image dimensions.

text_to_image would create a new image from a prompt and cannot guarantee preservation of the original image. image_variation generates related variants rather than targeted removals. image_to_image with high strength can regenerate broader areas and may alter unrelated visual details. Mask-based inpainting is the built-in editing control that limits modification to the selected region. Reference topics: Azure OpenAI image editing, mask inpainting, image edit API, input image, mask parameter, and computer vision image generation workflows.

Comprehensive and Detailed 150 to 250 words of Explanation From Azure AI Apps and Agents Developer/Course Guide/topics:

Use WebRTC . Microsoft ' s current Realtime API documentation explicitly recommends WebRTC for low-latency real-time audio streaming and lists it as the preferred connection type for client-side applications such as web and mobile apps. The documented latency is approximately 100 ms . Microsoft Learn

WebRTC is designed specifically for interactive media and provides capabilities appropriate for continuous microphone input and audio playback, including low-latency media transport and handling of real-time network conditions.

Microsoft distinguishes this from WebSocket , which is positioned primarily for server-to-server/backend scenarios and has a higher typical latency of approximately 200 ms in the current guidance. SIP is appropriate when integrating telephony infrastructure such as call centers or IVR systems. RTMP is not the supported GPT Realtime interactive client protocol for this scenario. Microsoft Learn

Therefore, the answer shown in the supplied file ( A, WebSocket ) conflicts with current Microsoft Learn documentation. For a browser/client application targeting approximately 100-ms audio latency, the verified answer is C, WebRTC .

Study Guide alignment: build multimodal/realtime generative applications and integrate speech as an agent modality . Microsoft Learn

====================

The solution does not fully meet the goal. Image moderation is appropriate for one part of the risk: blocking unsafe image content before the image is processed. Azure AI Content Safety provides image APIs that detect harmful content, and its harm categories and severity levels can be used to classify and block objectionable image content. This addresses unsafe photos, but it does not address hidden instructions embedded in images.

The second risk is prompt manipulation through extracted image text. After OCR extracts text from the uploaded image, that text becomes untrusted third-party content supplied to a generative model. Microsoft defines document attacks as malicious instructions embedded in third-party content, where the objective is to cause the model to execute unintended commands or alter intended behavior. Prompt Shields are the control designed to detect user prompt attacks and document attacks, including indirect attacks that come from uploaded or referenced content.

Therefore, image moderation alone is incomplete. A complete mitigation would combine image moderation for harmful visual content with Prompt Shields for document attacks, and optionally Spotlighting, so extracted or embedded text is treated as lower trust. Reference topics: Azure AI Content Safety, image moderation, Prompt Shields, document attacks, indirect prompt injection, and multimodal safety.

Comprehensive and Detailed 150 to 250 words of Explanation From Azure AI Apps and Agents Developer/Course Guide/topics:

Use the Azure Content Understanding skill. Content Understanding combines multimodal document processing, layout understanding, semantic extraction, image/figure handling, and structured document interpretation in a single content-processing capability.

For citation-rich RAG scenarios, the generated output can retain page and spatial information associated with source content, allowing downstream applications to connect returned text and visual content to page-level locations and bounding regions. Its richer document analysis is also designed to handle complex structures such as tables, figures, and content relationships that span document pages.

The ordinary Document Extraction skill primarily performs document cracking and basic extraction. Document Layout focuses on layout structure but does not provide the same complete multimodal semantic processing required by the question. GenAI Prompt applies generative processing to already available content rather than serving as the primary multimodal document-extraction skill.

Content Understanding therefore minimizes the need to compose multiple separate skills for OCR, image extraction, layout, and semantic document processing.

The AI-103 Study Guide explicitly covers multimodal pipelines that combine OCR, layout analysis, and field extraction and producing clean grounded representations for agents and RAG by using Content Understanding. Microsoft Learn

====================

The solution does not meet the goal. Prompt Shields for user prompts are designed to detect direct attempts by a user to manipulate the model through the prompt itself. In this scenario, the malicious instructions are embedded inside uploaded images and then introduced into the model context through extracted image text. That pattern is an indirect prompt injection or document attack, not merely a direct user-prompt attack. Microsoft’s Prompt Shields guidance distinguishes between user prompt attacks and document attacks, and states that document attacks involve harmful instructions embedded in supplied documents or third-party content.

The solution is also incomplete because users can upload unsafe images. Azure AI Content Safety includes image APIs that detect harmful content in images and support moderation across modalities. A complete mitigation would combine image moderation for unsafe visual content with Prompt Shields for document attacks, and optionally Spotlighting, so OCR-derived or embedded image text is treated as lower-trust context. Prompt Shields for user prompts alone would not reliably block unsafe images or hidden instructions extracted from those images. Reference topics: Azure AI Content Safety, Prompt Shields, user prompt attacks, document attacks, image moderation, and multimodal safety.

The correct configuration is advanced data parsing because the issue is not merely OCR; the ingestion job must preserve document structure for reliable RAG retrieval. Microsoft guidance for advanced parsing states that it automatically detects tables across all pages, including tables in scanned documents, merges tables that span multiple pages, restores column headers, and creates table chunks with metadata such as table index, shape, page numbers, section headings, and table previews. This directly satisfies the requirement to retrieve specific rows from multipage tables while retaining source-page context.

Basic parsing with fixed-size chunking would flatten the document into arbitrary text fragments, which is the current failure mode. OCR with page-level chunking improves text extraction from scanned PDFs, but it does not provide structure-aware chunks that preserve headings and table relationships across pages. Storing each page as a single chunk is too coarse for row-level retrieval and can bury relevant table rows in excessive context. Advanced data parsing is purpose-built for RAG ingestion because it produces semantically meaningful, retrievable chunks and enriches them with metadata needed for citations and grounding. Reference topics: RAG ingestion, advanced parsing, OCR, table extraction, structure-aware chunking, page metadata, and Azure Blob Storage document ingestion.

Configure hybrid search , which executes full-text and vector queries within the same Azure AI Search request. The full-text component searches the product-name and product-code fields through the lexical index, providing the precision required for exact or near-exact identifiers. Microsoft specifically identifies product codes and other specialized terms as scenarios that frequently perform better with keyword search.

The vector component compares the embedding of the user’s natural-language query with the embeddings stored for product descriptions. This retrieves semantically similar products even when the query and indexed description do not share the same literal words. Azure AI Search runs the full-text and vector searches in parallel and combines their result sets by using Reciprocal Rank Fusion, returning one unified ranking to the Foundry agent.

Keyword-only search would preserve exact matching but perform poorly for conceptual or paraphrased descriptions. Vector-only search supports semantic similarity but can miss precise product codes and rare identifiers. Semantic search alone reranks text-search results using language understanding; it does not replace the vector query required to use the existing embedding field.

Study Guide alignment: configure semantic, hybrid, and vector search for grounding, choose an appropriate retrieval method, and connect retrieval pipelines to agent tools .

Add the natural-language comment // what does function F do? after the code. The comment gives GPT-35 Turbo a direct and unambiguous instruction to analyze and explain function F. Because it uses syntax that naturally belongs beside source code, the model can distinguish the program being analyzed from the requested task and generate an explanatory completion. Microsoft’s prompt-engineering guidance recommends clearly stating the required outcome and using cues that direct the model toward the desired response.

This approach retains the existing GPT-35 Turbo deployment and adds only a small number of prompt tokens, satisfying the cost-minimization requirement. Changing to GPT-4-32k would be unnecessary because the requirement is a straightforward code-explanation task and does not indicate that a substantially larger context window is required.

Adding function F(explanation) resembles a new function declaration or invocation rather than an instruction to explain existing code. The model could interpret it as source code that must be completed. Setting temperature to 1 changes output randomness but does not tell the model to explain the function; it can also make the answer less consistent.

Study Guide alignment: Use prompt-engineering techniques, construct clear model instructions, tune generation behavior, and select an appropriate model based on quality and cost requirements.

====================

Comprehensive and Detailed 150 to 250 words of Explanation From Azure AI Apps and Agents Developer/Course Guide/topics:

Use Groundedness to detect claims in the generated answer that are not supported by the retrieved source context. Groundedness is fundamental to RAG evaluation because it determines whether the model stays faithful to the evidence supplied to it rather than introducing unsupported information.

Use the Retrieval evaluator to measure whether the search/retrieval stage returns content that aligns with the expected grounding information in the labeled dataset. A failure at this stage means the generation model might receive incomplete or irrelevant evidence even if it behaves correctly.

These two evaluators therefore inspect distinct parts of the RAG pipeline:

Retrieval evaluates the evidence selected by the retrieval system.

Groundedness evaluates whether the generated answer is supported by that evidence.

Response Completeness evaluates whether an answer sufficiently covers expected information. Fluency measures linguistic quality, and Coherence assesses organization and logical readability. None directly covers the two specified failure modes.

This combination is appropriate for a CI/CD quality gate because retrieval regressions and unsupported generation can be detected separately before production deployment.

The AI-103 Study Guide explicitly covers evaluating models and apps for fabrications, relevance, quality, and safety , together with monitoring grounding quality. Microsoft Learn

====================

Azure Language in Foundry Tools provides sentiment analysis and opinion mining , which are the appropriate natural language processing capabilities for evaluating public perception. Sentiment analysis processes unstructured social-media text and assigns classifications such as positive, neutral, or negative, together with confidence scores. Opinion mining extends this analysis by associating expressed sentiment with particular aspects, attributes, products, or services mentioned in the text. This allows the application to determine not only the overall attitude toward the brand but also which specific brand characteristics are receiving favorable or unfavorable reactions.

Azure Document Intelligence is intended primarily for extracting text, fields, tables, and structures from documents. Content Safety detects potentially harmful material rather than measuring consumer opinion. Azure Vision processes visual content and is not the primary service for sentiment classification of written social-media posts.

Azure Language can be accessed through REST APIs, client libraries, or supported containers. The application submits the social-media text and receives document-level and sentence-level sentiment results that can be aggregated into brand-perception metrics, dashboards, or alerts.

Study Guide alignment: Implement text analysis solutions — perform sentiment analysis and opinion mining, process unstructured text, and interpret confidence scores and sentiment classifications.

====================

The first statement is Yes because the conditional expression explicitly requires:

brand.confidence > = 0.75

Only brands meeting or exceeding the 0.75 confidence threshold are printed.

The second statement is Yes . A standard image bounding rectangle exposes x and y as the rectangle ' s origin, corresponding to the upper-left corner. Therefore, printing brand.rectangle.x and brand.rectangle.y provides top-left coordinates.

The third statement is No . The properties w and h represent the width and height of the bounding rectangle, not the absolute x/y coordinates of its lower-right corner.

To obtain the bottom-right coordinates, the application would calculate approximately:

right = x + w

bottom = y + h

Consequently, simply printing (w, h) does not identify the lower-right point despite the wording used in the print statement.

Bounding regions are a common output structure in computer-vision services and must be interpreted correctly when applications overlay detections or validate the physical location of visual entities.

Study Guide alignment: implement solutions that identify objects, components, or regions within images and interpret spatial model outputs . Microsoft Learn

====================

Comprehensive and Detailed 150 to 250 words of Explanation From Azure AI Apps and Agents Developer/Course Guide/topics:

The request invokes the Azure Cognitive Services/Azure AI resource regenerateKey management operation and explicitly specifies " keyName " : " Key2 " .

Azure AI service resources traditionally expose two interchangeable access keys: Key1 and Key2. Key1 is treated as the primary key and Key2 as the secondary key. Regenerating Key2 invalidates the previous Key2 value and produces a new value for that secondary credential.

Only the selected key is regenerated. Key1 remains unchanged, so option D is incorrect.

The request is unrelated to Azure AI Search query keys, which are separate read-only Search credentials. It also does not cause the newly generated key to be written automatically to Azure Key Vault. Such storage would require separate Key Vault configuration or deployment automation.

Maintaining two resource keys enables zero- or low-downtime key rotation: applications can temporarily use one key while the other is regenerated, then switch and rotate the remaining key.

The AI-103 Study Guide includes securing and managing Azure AI systems, including credentials, authentication, least privilege, and operational resource management. Microsoft Learn

====================

request_a calls the embeddings endpoint and specifies text-embedding-ada-002. An embedding request converts the supplied source_text into a numeric vector representation, so the first statement is Yes . Such vectors are used for semantic similarity, vector search, clustering, and Retrieval Augmented Generation.

request_b calls Chat Completions and includes a response_format based on a JSON schema named case_summary. Its schema defines summary, keyRisks, and nextActions, so the requested response is structured JSON. Therefore, the second statement is Yes .

The third statement is No . Although keyRisks is defined as an array and is required in the output shape, an array can legitimately be empty. If the support case contains no relevant risk, the structurally valid value may be " keyRisks " : []. The schema ensures the property exists; it does not guarantee the source content contains a substantive risk.

For current strict structured-output implementations, Microsoft also requires schemas to comply with the supported JSON Schema subset, including appropriate strict object definitions.

Study Guide alignment: integrate generative models into applications, implement vector embeddings, and generate structured JSON outputs . Microsoft Learn

====================

Azure Resource Manager resource paths are scoped through the subscription , not the Microsoft Entra tenant. Therefore, the URL must contain:

subscriptions/8d3591aa-96b8-4737-ad09-00f9b1ed35ad

The tenant ID identifies the Microsoft Entra directory but does not replace the subscription portion of an ARM resource path.

The appropriate resource provider is Microsoft.CognitiveServices . Azure AI service accounts, including multi-service Azure AI resources, are provisioned under the Microsoft.CognitiveServices/accounts resource type.

A multi-service resource is required because the solution needs both Language sentiment analysis and Vision OCR while exposing a consolidated Azure resource with common key/endpoint management. Microsoft.ApiManagement manages API gateways, Microsoft.ContainerService manages Kubernetes services, and Microsoft.KeyVault manages keys and secrets rather than provisioning Azure AI capabilities.

A typical ARM resource path therefore contains:

/subscriptions/ < subscription-id > /resourceGroups/ < rg > /providers/Microsoft.CognitiveServices/accounts/ < name >

This aligns with the AI-103 requirement to design Azure infrastructure for AI apps and choose appropriate Foundry/Azure AI services . Microsoft Learn

====================

Select AI quality (AI assisted) metrics for groundedness and relevance. Groundedness evaluates whether the generated answer is supported by the retrieved RAG context, while relevance evaluates whether the answer appropriately addresses the user’s query. These evaluations use an AI judge model to assess meaning and contextual alignment rather than relying only on exact token matches. Microsoft Foundry categorizes groundedness and relevance as quality evaluators and requires a compatible judge-model deployment for AI-assisted evaluation.

Select risk and safety metrics to evaluate harmful-content categories. These evaluators detect and score content involving violence, sexual material, self-harm, and hate or unfairness. They provide the safety dimension required alongside the RAG quality measurements.

AI quality NLP metrics include mathematically calculated measures such as F1, BLEU, GLEU, ROUGE, and METEOR. These primarily compare token or n-gram overlap with reference answers and do not provide the requested groundedness and content-harm assessments. A fluency evaluator measures language readability, while a similarity evaluator compares a response with ground truth; neither is a complete evaluation category covering the requirements.

Study Guide alignment: evaluate generative AI solutions by using quality, groundedness, relevance, risk, and safety metrics .

Comprehensive and Detailed 150 to 250 words of Explanation From Azure AI Apps and Agents Developer/Course Guide/topics:

The workflow needs the Microsoft Foundry project endpoint so that the evaluation action knows which Foundry project and agent environment it should access.

Microsoft ' s current GitHub Actions documentation lists the Foundry project endpoint as a required evaluation input (azure-ai-project-endpoint). The action uses that endpoint together with model/deployment, dataset, and agent information to execute evaluation against the appropriate project. Microsoft Learn

The tenant ID is used by the Azure login/authentication step but does not identify the Foundry project to evaluate. A model deployment name identifies a particular deployed model but not the complete project or agent resource. Current Microsoft evaluation actions also do not use an evaluation-config input matching option B as the primary project-binding parameter shown in this question.

The GitHub workflow should additionally be configured with a pull-request trigger and use Microsoft Entra/OIDC authentication rather than long-lived secrets wherever possible.

This directly matches the AI-103 Study Guide objective to integrate Foundry projects with continuous integration and continuous deployment (CI/CD) pipelines and automate evaluations before release. Microsoft Learn

====================

credential = DefaultAzureCredential()

agent = project_client.agents.get(agent_name=myAgent)

The correct authentication option is DefaultAzureCredential() because the case study states that API keys must not be used to access Foundry-deployed models and that Contoso developers must authenticate to Microsoft Foundry resources by using Microsoft Entra authentication. It also states that access to Project1 must be assigned to Agent1Dev Team by using the security group SC_Agent1_Dev . Microsoft Foundry authentication guidance recommends Microsoft Entra ID for production workloads because it supports least-privilege RBAC, per-principal auditing, and keyless authentication. AzureKeyCredential() would violate the no-API-key requirement, and None would not provide a valid credential.

The correct agent operation is get because the task is to access an existing agent named Agent1, not create a new version or retrieve a specific published version. Microsoft Foundry SDK examples show AIProjectClient created with DefaultAzureCredential() and then using project agent operations to create, retrieve, or interact with agents by name. To meet the compliance requirement, the group SC_Agent1_Dev must also be granted the appropriate project-scoped Foundry role, such as Foundry User, for Project1. Reference topics: Microsoft Entra authentication, Foundry RBAC, AIProjectClient, and project agent access.

The correct built-in skills are Azure OpenAI Embedding and Text Split. The case study requires an indexing pipeline that enables semantic and vector search over the product sheets stored in Azure Blob Storage, so Agent1 can retrieve relevant product information for natural language customer questions. For a RAG pipeline, long PDF content must first be broken into retrievable chunks, and each chunk must then be vectorized for semantic similarity retrieval.

Microsoft’s Azure AI Search integrated vectorization guidance states that you create a skillset that calls the Text Split skill for chunking and the Azure OpenAI Embedding skill to vectorize the chunks. The Text Split skill breaks text into chunks and provides positional metadata, making it suitable when downstream embedding skills have input-length limits. The Azure OpenAI Embedding skill connects to an embedding model deployed in Azure OpenAI or a Microsoft Foundry project and generates embeddings during indexing.

Language Detection, Entity Recognition, and key phrase extraction can enrich text, but they do not create vector embeddings. Merge is useful for combining OCR text with document text, but it does not satisfy the core vector-search requirement. Reference topics: Azure AI Search skillsets, Text Split skill, Azure OpenAI Embedding skill, integrated vectorization, and RAG indexing.

Call client.videos.create() to submit the prompt and start a video-generation job against the model deployment. Video generation is asynchronous: the initial response represents a job and contains properties such as its unique ID and current status rather than immediately returning the completed video. Microsoft’s official Python example creates the job by using client.videos.create(model=..., prompt=...).

The application must then poll the job. After waiting between requests, client.videos.retrieve(video.id) obtains the latest state of that specific job. Polling continues until the status becomes completed, failed, or cancelled. list would enumerate jobs rather than refresh the selected job, while download_content is used only after successful completion to retrieve the generated video file.

To comply with Contoso’s prohibition on API keys, the client should use a Microsoft Entra bearer-token provider. The OpenAI Python client accepts this provider through its api_key constructor parameter even though no static API key is used.

Study Guide alignment: design and implement video-generation solutions and integrate generative workflows through Foundry SDKs.

The correct answer is B. Modify the system message instructions . The case study states that Agent1 answers general questions about Contoso products and that a business requirement is for Agent1 to answer questions only about products sold by Contoso. This requirement defines the agent’s allowed domain and refusal boundary, so it must be expressed in the agent’s system-level instructions. Microsoft Foundry guidance states that system messages steer Azure OpenAI chat model behavior and are used to define the assistant’s role, boundaries, output format, and safety or quality constraints.

The system message should instruct Agent1 to answer only Contoso-product questions, use Contoso product documentation when available, and decline questions about non-Contoso products. This directly enforces the intended business scope at the highest instruction level. Few-shot examples can reinforce desired behavior but are not the primary control for defining mandatory operating boundaries. Top-p sampling and temperature are decoding controls; they influence randomness and diversity, not whether the agent restricts answers to a specific product domain. Increasing temperature would likely reduce consistency. Reference topics: Microsoft Foundry agent instructions, system message design, prompt engineering, response boundaries, and grounded generative AI behavior.

The correct recommendation is Azure Content Understanding in Foundry Tools . The case study states that Contoso’s finance department must manually review vendor invoices to verify that invoice details match vendor contract terms, and that the invoices contain tables, logos, and varied layouts that make consistent processing difficult. It also states that the planned solution must evaluate both the visual layout and textual content of the invoices.

Azure Content Understanding is designed for this type of multimodal document-processing workload. Microsoft describes Content Understanding as a Foundry Tool that processes unstructured and multimodal content, including documents and images, and transforms it into structured output for AI applications. It can use document analyzers to extract text, layout, tables, fields, and relationships from diverse document types.

Chat completions alone would not reliably extract structured invoice fields from complex layouts. Azure Document Intelligence can extract OCR, layout, and tables, but Content Understanding is the better end-to-end Foundry capability for combining visual and textual understanding with structured extraction for downstream verification. Image Analysis focuses on image-level visual features and is insufficient for invoice field and table review. Reference topics: Content Understanding, document analyzers, multimodal extraction, invoice processing, tables, layout, and structured JSON output.

Agent1 requires memory because the business requirement explicitly states that the agent must retain conversation context and recall relevant information during future interactions. Microsoft Foundry Agent Service memory provides persistent knowledge that can be retained and retrieved across separate conversations, sessions, devices, and workflows. It can store meaningful information extracted from prior interactions, including user preferences, relevant facts, and summarized context, and then retrieve that information when personalizing later responses.

Short-term memory maintains context within the current conversation. Long-term memory supports continuity across future sessions, which directly addresses Contoso’s requirement. Memory scopes should be associated with individual user identities so that one customer’s stored information cannot be retrieved during another customer’s interaction.

Guardrails enforce safety or behavioral restrictions but do not provide conversation recall. Tools enable the agent to access external systems, while instructions define the agent’s role and operating boundaries. Neither tools nor instructions persist user-specific information across sessions.

The AI-103 Study Guide maps this requirement to choose appropriate memory, tool, and knowledge integration services and build agents that integrate retrieval, function calling, and conversation memory .

====================

Deployment type: Standard

Version update policy: Once the current version expires

The correct deployment type is Standard . The case study specifies that Project1 is deployed in an EU Azure region and that model-processed data must remain within the EU. It also requires scalable, high-throughput generative AI workloads that dynamically handle variable customer support traffic without reserved throughput capacity. In Microsoft Foundry Models, Standard is a pay-per-token deployment type that processes data in a single Azure region, while Global Standard can process requests across regions and Global Provisioned uses reserved provisioned throughput. Microsoft’s deployment-type guidance identifies Standard as single-region, pay-per-token, whereas Global Provisioned is cross-region with reserved capacity.

The correct version update policy is Once the current version expires . This keeps Agent1 on the selected model version during its supported lifecycle, which supports stable and consistent responses, but still preserves continuity by automatically moving to a supported replacement when the current version is retired. Microsoft’s model versioning guidance states that this policy updates only when the current model version expires, while upgrading when a new default is available changes the deployment sooner and opting out can cause the deployment to stop working after retirement. Reference topics: deployment types, regional data processing, model versioning, throughput capacity, and stable production deployments.

The correct answer is B. Personally Identifiable Information (PII) Detection . The case study states that Agent1 must never reveal customer information , even if a document containing customer data is added accidentally to the product sheet repository in storage1. This is a privacy and compliance control requirement, so the appropriate capability is PII Detection.

Azure Language PII Detection is a Foundry Tools capability that identifies, classifies, and redacts sensitive information across text, conversations, and native documents. Microsoft states that PII Detection can be used to implement privacy controls, reduce sensitive data exposure, and support compliance requirements. In this scenario, PII Detection should be applied to retrieved product-sheet content and generated responses so customer names, contact details, identifiers, and other sensitive values are not exposed to users.

Prompt Shields are important for a separate requirement: protecting Agent1 from malicious instructions hidden in documents or embedded text. Microsoft describes Prompt Shields for documents as protection against hidden instructions embedded in external content. However, the option that directly satisfies the requirement to prevent disclosure of customer information is PII Detection. Self-harm and violence filters address harmful-content categories, not privacy leakage.

A Retrieval Augmented Generation evaluator is required because Agent1 will retrieve product-sheet content and use it as grounding context when generating answers. Foundry RAG evaluators assess both retrieval quality and the quality of the final grounded response. This directly supports Contoso’s requirement that answers based on the product sheets be relevant, complete, and accurate.

The appropriate evaluation suite can include groundedness , which determines whether the answer is supported by the retrieved product information without fabrication; relevance , which evaluates whether the answer accurately and completely addresses the customer’s question; and response completeness , which measures whether critical expected information is missing. Retrieval evaluators can additionally determine whether the indexing and search pipeline returned context chunks relevant to the query.

A coherence evaluator assesses readability and logical organization but does not establish that claims are supported by storage1. A custom guardrail controls or blocks behavior rather than measuring RAG quality. Fine-tuning changes model behavior and does not evaluate whether retrieved documentation was selected and used correctly.

The AI-103 Study Guide includes evaluating models and applications for fabrications, relevance, quality, and safety , as well as monitoring grounding quality and retrieval relevance.

The correct answer is B. Modify the system message instructions . The case study states that Agent1 answers general questions about Contoso products and that the business requirement is for Agent1 to answer questions only about the products sold by Contoso . This is a behavioral boundary for the agent, so it should be implemented in the highest-priority instructions that define the agent’s role, allowed scope, and refusal behavior.

Microsoft Foundry guidance states that a system message is used to steer model behavior, define the assistant’s role and boundaries, and add safety or quality constraints for the scenario. The system message should instruct Agent1 to answer only when the question concerns Contoso products, use the configured Contoso product documentation as grounding, and politely refuse or redirect questions about non-Contoso products.

Top-p sampling and temperature control randomness, not business-domain scope. Increasing temperature would make responses less deterministic. Few-shot examples can support desired behavior, but examples alone are weaker than explicit system-level instructions for defining operating boundaries. Reference topics: system message design, prompt engineering, agent instructions, response constraints, and grounded generative AI behavior.

The correct recommendation is Azure AI Search. The case study states that the product detail sheets are stored as PDFs in storage1, and that Agent1 must be enabled to retrieve and use detailed product information from those sheets. It also specifies that the indexing pipeline must enable semantic and vector search, and that Agent1 must answer natural language questions about product details by using the product sheet information. Azure AI Search is the Azure service designed to ingest content from sources such as Azure Blob Storage, create searchable indexes, and support keyword, semantic, hybrid, and vector retrieval for Retrieval Augmented Generation (RAG) solutions.

Microsoft’s Azure AI Search guidance states that integrated vectorization can chunk content and generate embeddings during indexing, enabling vector search over source documents. It also states that Azure AI Search supports text and vector queries and can improve raw content for search-related scenarios through enrichment pipelines. Azure Translator is unrelated to retrieval. Document Intelligence can extract document structure, but it is not the retrieval index for Agent1. Grounding with Bing Search retrieves public web content, not Contoso’s private PDFs in storage1. Reference topics: Azure AI Search, RAG, semantic search, vector search, Azure Blob Storage indexing, and agent grounding.