Free Microsoft Certified: Azure AI Apps and Agents Developer Associate (AI-103) Sample Questions

Correct answer: A — Indirect Prompt Injection detection

• A. Indirect prompt injection detection is specifically designed to identify and mitigate attempts to manipulate an AI's behavior through malicious instructions hidden in external content like uploaded documents, webpages, or emails. This feature is the primary defense against the scenario where a user embeds hidden prompts within a PDF.
• B. Groundedness evaluators assess how well the AI's responses are supported by the provided source data (answer faithfulness). While they ensure the AI isn't hallucinating facts, they do not scan for or detect malicious security instructions embedded in the source documents.
• C. Provenance metadata tracks the origin, history, and ownership of data to support traceability and auditing. It provides context about where a file came from but does not analyze the content of the file for hidden malicious instructions.
• D. Tool-access controls manage permissions and restrict which tools or APIs an AI agent can execute. While this can limit the damage an injection might cause by restricting capabilities, it does not detect the presence of malicious instructions within the uploaded text itself.

Correct answers: A, C — Text moderation; Prompt Shields

• A. Text moderation is a core feature of Azure AI Content Safety used to detect unsafe, harmful, or policy-violating content in text inputs, including categories like hate, sexual content, violence, and self-harm.
• B. Network security groups (NSGs) filter inbound and outbound network traffic for Azure resources at the transport layer. They do not analyze application-level user content or detect malicious intent in text prompts.
• C. Prompt Shields is a specialized feature within Azure AI Content Safety designed to detect and mitigate adversarial attacks such as prompt injection and jailbreaks, which represent specific malicious intents in user inputs to generative AI.
• D. Role-based access control (RBAC) manages identity and access permissions for Azure resources. It does not inspect user text for harmful content or security risks within a conversation.
• E. Azure Firewall is a network-level security service used to monitor and filter traffic. It does not provide the natural language processing capabilities required for content moderation or prompt-injection detection.

Correct answer: B — Custom blocklists in a content filter

• A. Jailbreak risk detection is designed to identify and mitigate attempts to bypass the model's safety guardrails or system instructions through prompt engineering. It does not provide the capability to block a specific list of proprietary names or terms in the output.
• B. Custom blocklists are a feature of Azure AI Content Safety that allow you to define a specific set of words or phrases that should be blocked if they appear in either the prompt or the model's completion. This is the most direct and appropriate method to ensure specific proprietary drug names are not included in the model's responses.
• C. Protected material detection (for text or code) is used to detect and block the generation of existing copyrighted content. It is a pre-configured filter for public materials and does not support custom, bespoke entries like proprietary trial drug names.
• D. Prompt shields (including User Prompt Shields and Document Shields) are security features designed to detect and block prompt injection attacks. They focus on identifying malicious intent in inputs rather than enforcing a prohibited vocabulary list for model outputs.

Correct answers: A, B, C — Disable public network access on the Azure AI Hub.; Create a Private Endpoint for the Azure AI Hub.; Configure Private DNS zones for the private endpoints.

• A. Correct. Disabling public network access is a fundamental step to block all ingress from the internet to the AI Hub resource, ensuring the service is not reachable via its public IP.
• B. Correct. A Private Endpoint uses a private IP address from your VNet, enabling private connectivity to the Azure AI Hub via Azure Private Link and keeping traffic within the Microsoft backbone network.
• C. Correct. Private DNS zones are essential to ensure that the Fully Qualified Domain Name (FQDN) of the AI Hub resolves to the internal Private Endpoint IP address rather than a public IP.
• D. Incorrect. CORS is an application-level security mechanism used by browsers to manage cross-domain requests; it does not provide network isolation or disable public internet access.
• E. Incorrect. Opening port 80 (HTTP) on an NSG is insecure and typically increases public exposure rather than restricting it for a private environment.
• F. Incorrect. Assigning a public IP address to managed compute resources directly contradicts the requirement for no public internet access and would expose the environment to the internet.

Correct answer: B — Enable a system-assigned managed identity on the Azure OpenAI resource and assign it the 'Search Index Data Reader' role on the Search service.

• A. Incorrect. While Azure Key Vault improves secret management by centralizing storage, it still relies on a secret-based authentication model (using the key itself). This does not meet the requirement for keyless authentication (RBAC).
• B. Correct. Enabling a system-assigned managed identity on the Azure OpenAI resource allows it to authenticate to Azure AI Search using Microsoft Entra ID. Assigning the 'Search Index Data Reader' role provides the necessary permissions to read and query index data without the need for API keys or stored secrets.
• C. Incorrect. Shared Access Signature (SAS) tokens are credentials derived from shared secrets. They are not considered a keyless authentication mechanism and are not the recommended method for service-to-service authentication in this context.
• D. Incorrect. This configuration is reversed. The identity must be enabled on the service initiating the request (Azure OpenAI), and the permissions must be granted on the target resource (Azure AI Search). Additionally, 'Cognitive Services User' is not a role used for querying an Azure AI Search index.

Correct answer: B — Azure Machine Learning Data Drift monitor

• A. Azure AI Content Safety is designed to detect and manage harmful or unsafe content (such as hate speech, violence, or sexual content) in AI-generated or user-provided text and images. It does not monitor statistical changes in input data distributions or detect model drift.
• B. Azure Machine Learning Data Drift monitor is the specific tool designed to detect changes in the statistical properties of input data over time by comparing inference data with the original training dataset. Detecting these shifts is critical for identifying why a model's accuracy may be degrading in production.
• C. Azure AI Studio Groundedness evaluation is used primarily for generative AI and Large Language Models (LLMs) to verify if the model's responses are supported by the source data provided. It is not intended for monitoring feature distribution changes in traditional machine learning models.
• D. Application Insights Profiler is a performance diagnostic tool used to identify execution bottlenecks and capture traces in live applications. It tracks application responsiveness and resource usage but lacks the capability to analyze machine learning data distributions.

Correct answer: A — AzureDiagnostics

• A. Correct. When diagnostic logging is enabled for Azure OpenAI, the platform-level resource logs—specifically the 'RequestResponse' category—are stored in the AzureDiagnostics table. This table contains the 'Properties' column which holds the JSON payload for content filtering results, detailing which specific categories (such as Hate, Violence, or Sexual) were flagged or triggered a block.
• B. Incorrect. AppTraces is a table associated with Azure Application Insights, used primarily for application-level trace data and telemetry captured via an SDK. While it can store content safety data if an application is manually instrumented to send it there, it is not the default table for the automated diagnostic logs generated by the Azure OpenAI Service.
• C. Incorrect. The AzureMetrics table stores aggregated, numeric performance data and usage statistics, such as token counts or latency. It does not provide the granular, trace-level metadata or JSON properties required to identify specific content safety category triggers for individual requests.

Correct answer: B — Use Azure AI Search with a Retrieval-Augmented Generation (RAG) pattern.

• A. Fine-tuning GPT-4 on PDF manuals is computationally expensive and inefficient for knowledge retrieval. Fine-tuning is typically used to adapt a model to a specific tone, style, or task performance rather than for injecting specific factual knowledge from a document set.
• B. Azure AI Search with a Retrieval-Augmented Generation (RAG) pattern is the standard and most efficient approach for question answering over internal document collections. It allows the system to retrieve the most relevant passages from the PDF manuals and provide them as context to the model, ensuring grounded and accurate responses with the ability to cite sources.
• C. The Bing Grounding tool is designed for grounding responses in public web search results. Since HR policies are internal and private documents, they should be indexed securely in your own data source like Azure AI Search rather than queried through a public web tool.
• D. Training a custom Small Language Model (SLM) from scratch is extremely resource-intensive and unnecessary. Pre-trained models combined with a RAG pattern can handle this task much more effectively without the need for the massive compute required for foundational training.

Correct answer: A — GPT-4o

• A. GPT-4o is a state-of-the-art multimodal large language model capable of processing and generating text, images, and audio. It is the primary choice for complex generative AI applications and agentic workflows that require integrated reasoning across different media types within a single model.
• B. Whisper is a specialized automatic speech recognition (ASR) model designed for speech-to-text transcription. While it is excellent for converting audio into written text, it is not a general-purpose multimodal generative model used for building complex agents.
• C. Azure AI Search with vector support is a retrieval and indexing service. It is used to support Retrieval-Augmented Generation (RAG) by providing grounded data to models, but it is not a generative model or an agent itself.
• D. DALL-E 3 is a specialized model for generating high-quality images from text prompts. It is not suitable for general-purpose natural language reasoning, audio processing, or conversational agent tasks.
• E. Text-Embedding-3-small is an embedding model that converts text into numerical vectors for similarity search and classification. It is a utility model used for retrieval and is not capable of generating natural language responses.

Correct answer: B — Use a System-Assigned Managed Identity and assign appropriate RBAC roles.

• A. Incorrect. A connection string is a secret that would need to be stored in configuration settings. This violates the security policy of not storing secrets and is not the most secure modern authentication method compared to managed identity.
• B. Correct. A system-assigned managed identity allows the Azure App Service to authenticate to Azure AI Foundry using Microsoft Entra ID. This eliminates the need to store any credentials, secrets, or keys in the code, configuration, or Key Vault. Access is governed via Role-Based Access Control (RBAC) roles assigned directly to the identity.
• C. Incorrect. Storing an API key in a Dockerfile as an environment variable embeds a static secret into the application artifact or container image, which is insecure and explicitly forbidden by the company policy.
• D. Incorrect. A Shared Access Signature (SAS) token is a signed credential that must be generated, distributed, and refreshed. It still functions as a secret, which violates the requirement to avoid storing secrets or API keys altogether.

Correct answers: A, B — Fluency; Coherence

• A. Fluency evaluates whether the generated text is grammatically correct, natural, and easy to read. It specifically addresses the requirement to measure grammatical ease.
• B. Coherence measures how logically the summary flows from one sentence to the next and whether the overall content is internally consistent and well-structured.
• C. Groundedness measures whether the generated summary is supported by the source content (factuality). It does not assess the readability, grammar, or logical flow of the text.
• D. F1 Score is a statistical metric used to evaluate token overlap between a reference and generated text. It is not designed to measure stylistic qualities like grammatical correctness or logical flow.
• E. Exact Match is a strict metric that checks if the generated output is identical to a reference string. It is not an appropriate tool for evaluating the nuances of grammar or the logical connection between sentences.

Correct answer: A — True

• A. The statement is true because the Groundedness metric specifically assesses whether the model's response is supported by the provided source context, ensuring that the answer remains factually aligned with the input and avoids generating hallucinated or unsupported claims.
• B. The statement is false because Groundedness is precisely the metric used to evaluate the alignment between the generated output and the source information, rather than general answer quality, style, or fluency.

Correct answer: B — Set Top-p to 0.5

• A. Incorrect. Temperature controls the randomness or determinism of the output by scaling the logits before the softmax layer. While a lower temperature makes the output more focused on high-probability tokens, it does not explicitly limit the selection to a specific cumulative probability threshold like Top-p does.
• B. Correct. Top-p (also known as nucleus sampling) restricts the model to only consider tokens whose cumulative probability mass adds up to the specified threshold. Setting Top-p to 0.5 ensures that only the smallest set of tokens comprising the top 50% of the probability mass are considered for the next word generation.
• C. Incorrect. The Frequency penalty reduces the likelihood of repeating the same word or phrase based on its current frequency in the generated text. It is used to reduce verbatim repetition but does not control the probability mass threshold for token selection.
• D. Incorrect. The Presence penalty penalizes tokens based on whether they have appeared in the text at all, which encourages the model to introduce new topics or vocabulary. It does not control the cumulative probability mass filter for initial token selection.

Correct answer: B — Time Between Tokens (TBT)

• A. Time to First Token (TTFT) measures the latency from the moment a user sends a request until the model produces the very first output token. It highlights initial responsiveness and system prompt processing but does not track the generation speed of subsequent tokens.
• B. Time Between Tokens (TBT) specifically measures the average time elapsed between generating consecutive tokens following the first token. This metric is essential for evaluating the perceived smoothness and throughput of a streaming response for the end user.
• C. Total Generation Time represents the cumulative duration from the start of the request until the final token is generated. It includes TTFT and the time for all subsequent tokens, making it a measure of the overall request lifecycle rather than the specific speed between tokens.
• D. Round Trip Time (RTT) is a network-level metric that measures the total time taken for a request to travel from the client to the server and for the response to return. It focuses on network latency rather than the model's internal token generation performance.

Correct answers: A, B — OpenTelemetry; Azure Application Insights

• A. OpenTelemetry is a widely supported open-source standard for generating, collecting, and exporting telemetry data (traces, metrics, and logs). Azure AI Studio natively supports OpenTelemetry for distributed tracing, allowing developers to monitor and debug complex AI application flows and prompt execution.
• B. Azure Application Insights is an extension of Azure Monitor that provides application performance management (APM) and observability. It is the primary native Azure tool for collecting, storing, and analyzing telemetry and logs generated by Azure AI Studio and Prompt flow.
• C. AWS CloudWatch is a monitoring and observability service specific to Amazon Web Services. It is not a natively supported integration for Azure AI Studio's observability framework.
• D. Google Cloud Operations Suite is the monitoring and logging solution for Google Cloud Platform. It is not natively supported in Azure AI Studio, which focuses on Azure-native and industry-standard integrations like OpenTelemetry.
• E. Apache Kafka is a distributed event streaming platform used for building real-time data pipelines. While it can transport data, it is not a tracing standard or an observability tool natively used by Azure AI Studio for system monitoring.

Correct answer: B — Azure Monitor Application Insights

• A. Incorrect. Azure AI Content Safety is used for detecting and filtering harmful or unsafe content within AI applications. It does not provide the telemetry or performance monitoring capabilities required to track API latency or trigger operational alerts.
• B. Correct. Azure Monitor Application Insights is an Application Performance Management (APM) service designed to collect telemetry such as request duration, dependency latency (including external API calls), and custom events. It allows you to configure alerts when specific performance thresholds, such as a 2-second latency, are exceeded.
• C. Incorrect. Azure Cost Management is dedicated to analyzing, tracking, and optimizing cloud spending and budgets. It cannot monitor runtime application performance metrics like API latency.
• D. Incorrect. Azure Policy is a governance and compliance tool used to enforce organizational standards across Azure resources (e.g., ensuring resources are in specific regions). It is not designed for monitoring runtime telemetry or alerting on application performance.

Correct answers: A, B — Hardcode a rule in the custom function's backend API to reject restart requests for production database VMs.; Add explicit instructions in the system prompt to never restart production database VMs.

• A. Enforcing safety-critical rules in the backend API is the most reliable safeguard. It ensures that even if the model attempts to call the function for a protected asset, the integration layer explicitly rejects the action. This 'server-side' validation is a core requirement for secure agentic systems.
• B. Adding explicit instructions in the system prompt serves as the primary steering mechanism for the agent's reasoning. It defines the constraints and behavioral boundaries the model should follow during planning. While prompts are not foolproof on their own, they are a necessary layer of guidance.
• C. Azure AI Content Safety is designed to detect and mitigate harmful content like hate speech, violence, or self-harm. It is not a tool for infrastructure policy enforcement or logic-based filtering of specific technical terms like 'database'.
• D. Disabling function calling would prevent the agent from performing any actions at all, including the necessary log reading and VM restarts it is intended to do. This is an overly restrictive measure that breaks the agent's core functionality.
• E. The top_p (nucleus sampling) parameter controls the diversity and randomness of the model's output. While a lower value makes the model more deterministic, it does not provide any logical or security-based safeguard against specific unauthorized tool calls.

Correct answer: B — JSON Schema

• A. Incorrect. XML is a markup language used for data storage and transmission, but it is not the standard format for defining function parameters or tool schemas in modern LLM integrations like Azure AI Foundry. It lacks the specific typing validation support typically expected by model inference engines for tool calling.
• B. Correct. JSON Schema is the industry-standard format for defining the structure, data types, and required fields of tool parameters. In Azure AI Foundry and OpenAI-compatible agent frameworks, this schema allows the LLM to understand how to correctly format arguments when calling a custom tool.
• C. Incorrect. While YAML is frequently used for configuration files and serializing data, the formal specification for defining the actual parameter types and constraints in tool contracts for agentic workflows is JSON Schema.

Correct answer: A — Train an Azure Custom Vision object detection model on the prohibited symbol.

• A. Correct. Training an Azure Custom Vision object detection model is the most effective way to identify specific visual patterns or symbols, such as a unique logo, that are not included in standard pre-trained safety categories. This approach allows for tailored fine-tuning to recognize the specific visual elements of the logo that generic models might miss.
• B. Incorrect. The Azure AI Content Safety Hate category is designed to detect hateful or offensive symbols and speech associated with hate groups. It is not intended to recognize commercial brand logos or specific proprietary visual symbols.
• C. Incorrect. Prompt Shields for images are intended to protect generative AI models from prompt injection attacks or malicious instructions in multimodal inputs. They do not provide the functionality to detect specific visual brand assets or custom symbols for policy enforcement.
• D. Incorrect. Azure AI Vision OCR (Optical Character Recognition) is used to extract text from images. While it might read text contained within a logo, it cannot identify the logo as a visual symbol or graphic, making it an unreliable tool for enforcing a strictly visual brand policy.

Correct answer: B — Enable Content Credentials (C2PA) provenance watermarking.

• A. Incorrect. Applying a visible text overlay is not a cryptographic method and can be easily removed or altered. Azure AI Vision is primarily used for image analysis and processing, not for embedding verifiable provenance metadata.
• B. Correct. Content Credentials based on the C2PA (Coalition for Content Provenance and Authenticity) standard provide a cryptographic way to sign images. This allows external platforms to verify the origin and provenance of the image, identifying it as AI-generated.
• C. Incorrect. Prompt Shields are a security feature within Azure AI meant to protect against prompt injection and jailbreak attacks. They do not provide cryptographic watermarking or identity verification for generated images.
• D. Incorrect. Azure AI Content Safety is used for content moderation (detecting harmful content like hate speech or violence). It does not append digital signatures for provenance or cryptographic identification of AI-generated assets.

Correct answer: A — Read API (OCR)

• A. The Azure AI Vision Read API (OCR) is the standard tool for extracting both printed and handwritten text from images and documents. In a multimodal moderation workflow, this feature is essential for identifying embedded text that needs to be checked against a custom blocklist or content safety policy.
• B. The Face API is specifically designed for facial detection, recognition, and analysis of human attributes. It lacks the Optical Character Recognition (OCR) capabilities required to extract text from an image.
• C. Spatial Analysis is used to analyze real-time video streams to monitor the movement and presence of people in physical locations (such as counting people or measuring social distancing). It is not designed for document or image text extraction.

Correct answer: B — It relies on the semantic understanding of the text prompt to identify and modify the relevant regions of the source image.

• A. Incorrect. Prompt-driven image editing APIs are designed to interpret text instructions directly. They do not typically require a secondary API call to generate a mask first, as the model handles region identification and image manipulation within the same inference process.
• B. Correct. Modern generative AI models use semantic understanding to align the text prompt with specific features in the source image. By mapping the linguistic context to the visual data (often through mechanisms like cross-attention), the model identifies which regions are relevant to the edit and modifies them while attempting to preserve the rest of the image.
• C. Incorrect. These APIs aim for targeted, prompt-guided edits rather than applying a global, uniform filter. The goal is localized modification based on the prompt's intent rather than a universal transformation across all pixels.
• D. Incorrect. There is no fixed spatial heuristic, such as modifying only the center 50% of the image. The edit area is determined dynamically based on the model's interpretation of the text prompt and the visual content of the source image.

Correct answer: A — `revised_prompt`

• A. The `revised_prompt` field in the Azure OpenAI DALL-E 3 API response contains the expanded and enriched version of the original user prompt. DALL-E 3 automatically rewrites prompts to provide more descriptive detail for higher-quality image generation; this field allows users to see exactly what the model processed.
• B. The `content_filter_results` field contains metadata regarding safety and content moderation. It indicates whether the input or output triggered filters for categories such as violence or hate speech, but it does not store the modified prompt text.
• C. The `system_fingerprint` field identifies the specific backend configuration or version of the model serving the request. It is primarily used for tracking reproducibility and configuration changes, not for viewing prompt modifications.

Correct answer: B — Azure OpenAI GPT-4o

• A. Incorrect. Azure AI Vision Image Analysis 4.0 with the caption feature is designed for generating short, concise captions (typically a single sentence). It does not provide the deep reasoning or the ability to generate the multi-paragraph, detailed descriptions required for complex charts.
• B. Correct. Azure OpenAI GPT-4o is a multimodal large language model (LLM) that can process images and reason over complex data like charts. It is capable of generating detailed, context-aware, multi-paragraph narratives, making it the most suitable service for meeting advanced WCAG accessibility requirements for complex visual data.
• C. Incorrect. Azure Custom Vision Object Detection is used to train models to identify and locate specific objects within images. It lacks the natural language generation and semantic reasoning capabilities needed to explain data trends or produce descriptive text.
• D. Incorrect. Azure Content Understanding pipelines are primarily focused on extracting structured metadata and information from content. While useful for data extraction, they are not optimized for generating extended human-readable narratives and reasoning over chart content in the way a multimodal LLM like GPT-4o is.

Correct answers: A, C — Defining a custom schema to extract specific data points from the charts.; Combining OCR text extraction with visual object detection in a single workflow.

• A. Correct. Pro-mode in Azure Content Understanding supports custom schema design, allowing you to define exactly which fields or data points to extract from complex content such as charts and tables, ensuring structured output tailored to specific document types.
• B. Incorrect. Azure Content Understanding is designed to analyze and extract structured information from documents and images; it does not possess the functionality to generate 3D models.
• C. Correct. The pro-mode pipeline is specifically designed for multimodal understanding, which integrates OCR-based text extraction with visual analysis (such as object detection or layout analysis) in a single workflow. This is ideal for documents containing both text and embedded charts.
• D. Incorrect. Azure Content Understanding does not offer built-in simultaneous translation into 50 languages. Translation is a separate capability handled by the Azure AI Translator service and would require additional API calls.
• E. Incorrect. While Azure offers video analysis services, the pro-mode pipeline in Content Understanding for documents is focused on processing static images or document files rather than real-time video streaming analysis.

Correct answers: A, B — Plain text sentences containing the proprietary product names.; Audio files with matching human-labeled transcripts of users speaking with the regional accents.

• A. Correct. Plain text sentences (Related Text) that include proprietary product names help the Custom Speech service learn domain-specific vocabulary and improve recognition of terms that are not part of the standard language model.
• B. Correct. Audio files paired with human-labeled transcripts are essential for acoustic adaptation. This training data helps the model learn the specific acoustic patterns and phonetic variations associated with heavy regional accents.
• C. Incorrect. Viseme datasets map phonemes to facial expressions for lip-syncing and speech animation; they do not improve audio transcription accuracy or recognize product names.
• D. Incorrect. Speech Synthesis Markup Language (SSML) files are used to control how text is converted to speech (TTS). They are not used for training Custom Speech-to-Text (STT) recognition models.
• E. Incorrect. While verbal consent may be a legal requirement for data collection, these files do not provide the aligned transcript or domain vocabulary necessary to improve the model's transcription accuracy.

Correct answers: A, B — Add a Speech-to-Text plugin/service to convert user audio to text before the kernel processes the prompt.; Add a Text-to-Speech plugin/service to convert the kernel's text response into audio.

• A. Correct. Speech-to-Text (STT) is required to transcribe the user's spoken audio into text so that the Semantic Kernel can process it as a text-based prompt. This enables the agent to 'listen' to queries.
• B. Correct. Text-to-Speech (TTS) is necessary to convert the agent's generated text output back into synthesized audio. This allows the agent to 'speak' its responses to the user, completing the bidirectional flow.
• C. Incorrect. A Custom Speech acoustic model is used to improve recognition accuracy for specific environments or accents; it does not replace the Large Language Model (LLM) which is responsible for reasoning and text generation.
• D. Incorrect. Keyword Recognition is used for wake-word detection (e.g., 'Hey Siri') to trigger an action, but it is not used to generate conversational text responses.
• E. Incorrect. Visemes are visual representations of mouth and facial positions used for lip-syncing animations (avatars). They do not provide language translation or speech-to-text functionality.

Correct answer: B — False

• A. The statement is false because transcribing audio to text using a model like Whisper transforms acoustic signals into alphanumeric characters, which does not encode paralinguistic features such as pitch, volume, or non-verbal sounds. While the resulting text may convey semantic sentiment, the raw acoustic emotional nuances are lost during the transcription process.
• B. The statement is false because the transcription process strips away the audio characteristics that convey emotion through voice. Since GPT-4 only receives the linguistic content as a text string, it has no access to the original audio waveform's prosody, meaning it cannot directly analyze nuances like the sound of shouting or crying.

Correct answer: B — Few-shot prompting providing examples of passive-aggressive transcripts and their labels.

• A. Zero-shot prompting relies on the model's pre-existing knowledge without specific context. While a detailed system message helps define the task, detecting nuanced linguistic styles like passive-aggression often requires specific patterns that zero-shot might miss compared to provided examples.
• B. Few-shot prompting (providing labeled examples) is the most effective method for complex sentiment or tone analysis. By providing specific examples of passive-aggressive language, the model learns the subtle linguistic markers and contextual cues required for accurate classification.
• C. Chain-of-thought prompting is designed for tasks requiring logical reasoning or multi-step processing. Asking the model to translate the text first does not assist in identifying emotional subtext and may even lose the nuances of the original language's tone.
• D. Map-reduce prompting is typically used for summarizing long documents that exceed token limits. Summarizing a transcript before tone detection is counterproductive, as the subtle phrasing and indirect cues necessary to identify passive-aggression are often stripped away during the summarization process.

Correct answer: B — Add a Python tool node that uses regular expressions to extract the codes directly from the text

• A. Incorrect. Increasing the temperature makes the model's output more random and creative. For structured extraction tasks, deterministic behavior is required, and a temperature of 0 is typically preferred. Increasing it would likely decrease consistency and reliability.
• B. Correct. Using a Python tool node with regular expressions (regex) allows for the reliable extraction of fixed patterns like codes or IDs. This programmatic approach is deterministic and avoids the probabilistic inconsistencies or hallucinations that can occur when using an LLM for exact pattern matching.
• C. Incorrect. Jailbreak detection is a security feature within Azure AI Content Safety designed to identify malicious prompt injection attacks. It does not improve the data extraction capabilities or logic of the model.
• D. Incorrect. Azure AI Translator is designed for translating text between different languages. It is not intended for data normalization or pattern extraction and would not resolve the issue of missing or inconsistent code extraction.

Correct answer: B — Use the `tools` parameter to define a function with a JSON schema for its arguments, and set `tool_choice` to require the model to call that function.

• A. Incorrect. While setting `response_format` to `json_object` ensures the output is a syntactically valid JSON, it does not enforce a specific schema. The model will attempt to follow the schema described in the prompt, but this is not guaranteed, making this method less reliable for ensuring consistent structure.
• B. Correct. This is the most robust method for enforcing a specific JSON schema. By defining a tool (or function) with parameters described by a JSON schema and forcing the model to use it via `tool_choice`, you compel the model to generate a JSON object that strictly adheres to that schema. This is the recommended approach for reliable structured data extraction, as mentioned in best practices for using Azure OpenAI.
• C. Incorrect. While Azure AI Language services are effective for specific extraction tasks, the question is about configuring the Azure OpenAI Chat Completions API for a custom task. This option proposes using a different, more specialized service rather than configuring the generative model itself.
• D. Incorrect. This approach does not configure the model's output; it attempts to structure the data after generation. Relying on regular expressions to parse natural language output from a large language model is brittle and prone to errors if the model's phrasing or format changes slightly.

Correct answers: A, B — A memory connector for Azure AI Search; A text completion or chat completion model

• A. Correct. A memory connector (or vector store connector) for Azure AI Search is necessary to integrate search-backed retrieval into Semantic Kernel. This component allows the kernel to interface with the search index to fetch relevant documents or chunks required for grounding.
• B. Correct. A text completion or chat completion model (LLM) is required to process the retrieved context and generate a natural language response. In Semantic Kernel, the model acts as the reasoning engine that consumes the grounded data to answer the user's question.
• C. Incorrect. An Azure Logic App trigger is used for external workflow automation and is not a core component required for integrating Azure AI Search retrieval within a Semantic Kernel agentic workflow.
• D. Incorrect. While a Custom Web API skill can extend an agent's functionality, it is not an essential component for the specific task of connecting Azure AI Search to Semantic Kernel for document retrieval and question answering.
• E. Incorrect. An Azure Machine Learning compute instance is typically used for training models or hosting managed endpoints; it is not a required orchestration component for a Semantic Kernel retrieval pipeline.

Correct answers: A, B, C — An Indexer; A Data Source; A Skillset with an AzureOpenAIEmbedding skill

• A. An Indexer is the central orchestration engine in Azure AI Search. It automates the ingestion process by pulling data from the source, running the enrichment skillset, and writing the final results into the search index.
• B. A Data Source provides the connection details to the Azure Blob Storage container. It allows the indexer to monitor the container for new or updated PDFs to begin the ingestion process.
• C. A Skillset defines the processing pipeline. To handle RAG ingestion without custom code, you use built-in skills like the Split skill (for chunking) and the AzureOpenAIEmbedding skill (for vectorization) to transform the raw PDF content into searchable vectors.
• D. The Semantic Ranker is a feature used during the search/retrieval phase to improve result relevance using language models. It is not used during the data ingestion or indexing phase.
• E. Synonym maps are used to expand queries with equivalent terms at search time. They do not contribute to the automated cracking, chunking, or vectorization of source documents.
• F. While Azure Functions can be used as custom skills, the scenario specifically asks for a solution that does not require writing custom code for orchestration. Azure AI Search's native indexer and skillset features provide this functionality out-of-the-box.

Correct answer: B — False

• A. The statement is false because the markdown output feature in Azure AI Document Intelligence is designed to enhance document readability and structure (such as headings and tables) for LLMs; embedding per-word geometry or bounding box coordinates inline would clutter the text and significantly increase token consumption.
• B. The statement is true because Azure AI Document Intelligence maintains a clear separation between content and metadata; while spatial data and bounding box coordinates are available in the structured JSON response payload under the pages and words collections, they are not automatically inserted into the markdown string.

Correct answers: A, C — Azure AI Document Intelligence Read API to extract the raw text.; Azure OpenAI Service using Structured Outputs to map the text to the JSON schema.

• A. Correct. The Azure AI Document Intelligence Read API is the standard tool for OCR, used to extract raw text and layout information from scanned or digital clinical notes. This serves as the necessary first step in providing the text content to a secondary processing layer.
• B. Incorrect. The Prebuilt Invoice model is specifically trained for financial documents with set fields like vendor name and total amount. It is not designed to handle the variable and domain-specific terminology found in unstructured clinical notes.
• C. Correct. Azure OpenAI Service with Structured Outputs (or JSON mode) is the ideal tool for processing unstructured text and enforcing a strict, predefined JSON schema. It can interpret complex medical contexts and map them into the required fields for a database.
• D. Incorrect. Azure AI Vision Image Analysis focuses on general visual features, object detection (e.g., detecting a stethoscope), and tagging in images. It does not provide the robust document text extraction or schema-based formatting needed here.
• E. Incorrect. Azure AI Search is a retrieval-augmented generation (RAG) and indexing service. While it can store information, it is not used to extract fields from documents or map them into a specific JSON schema.