AI-103 Practice Questions

The scenario requires a real-time, low-latency conversational agent that processes continuous audio input and provides spoken responses. This is achieved by combining two core functionalities of the Azure AI Speech service:

1. Real-time speech to text (STT): This service transcribes the user's spoken words into text as they are being said. Its low-latency nature is essential for conversational turn-taking and a natural user experience.

2. Text to speech (TTS): After the agent's logic processes the transcribed text and formulates a response, TTS converts that text response back into synthesized, spoken audio for the user.

This combination directly addresses the need for a continuous, interactive voice workflow.

A. Use an embeddings model to encode the audio, and then decode the audio into text and speech.

This is an abstract description of a potential underlying mechanism, not the specific Azure AI services required. The direct solution involves using the dedicated STT and TTS services.

B. Use batch transcription to convert the audio input and return text responses from the agent.

Batch transcription is designed for processing large volumes of pre-recorded audio and has high latency, making it unsuitable for the required real-time, conversational interaction.

C. Use speech translation to convert the audio into another language and return the translated text.

Speech translation is for multilingual scenarios. The question does not specify a need for translation, but rather a direct speech-to-text and text-to-speech workflow.

1. Microsoft Learn, "What is speech to text?": This document explicitly distinguishes between real-time and batch transcription. It states, "Real-time speech to text transcribes audio from a source in real time... Use real-time speech to text for applications that need to transcribe audio in real time such as... conversational agents."

Reference: Microsoft Corporation. (2024). What is speech to text?. Microsoft Learn. Retrieved from https://learn.microsoft.com/en-us/azure/ai-services/speech-service/speech-to-text#real-time-speech-to-text (Section: "Real-time speech to text").

2. Microsoft Learn, "What is text to speech?": This document describes the service for generating spoken output. "Text to speech from the Speech service enables your applications, tools, or devices to convert text into humanlike synthesized speech." This is the required component for the agent's spoken response.

Reference: Microsoft Corporation. (2024). What is text to speech?. Microsoft Learn. Retrieved from https://learn.microsoft.com/en-us/azure/ai-services/speech-service/text-to-speech (Section: "Core features").

3. Microsoft Learn, "Batch transcription": This source clarifies the use case for batch processing, confirming it is not for real-time scenarios. "Batch transcription is a set of REST API operations that enable you to transcribe a large amount of audio in storage."

Reference: Microsoft Corporation. (2024). What is batch transcription?. Microsoft Learn. Retrieved from https://learn.microsoft.com/en-us/azure/ai-services/speech-service/batch-transcription-overview (Introduction paragraph).

Agentic Retrieval Augmented Generation (RAG) employs a large language model (LLM) as an intelligent agent to orchestrate the entire retrieval and generation process. This agent can decompose a user's complex question into multiple sub-queries, which can be executed in parallel against the vectorized index to reduce latency. The agent maintains the state of the multi-turn conversation, using the history to intelligently plan subsequent retrieval actions. It then synthesizes the information gathered from multiple retrieved chunks to formulate a comprehensive answer. This advanced pattern directly meets all the specified requirements for handling complexity, conversational context, and performance.

A. classic Retrieval Augmented Generation (RAG): This approach typically involves a single, stateless retrieval step per query and lacks the sophisticated planning and orchestration needed for multi-turn context and parallel retrievals.

B. chain of thought: This is a prompting technique that guides an LLM's reasoning process during generation; it is not a retrieval architecture for interacting with a data index.

D. iterative retrieval: This describes a process of refining a search over multiple steps but is only one component of a larger agentic framework and does not encompass the full scope of planning, parallel execution, and conversational memory.

1. Microsoft Learn, "Advanced Retrieval Augmented Generation (RAG) patterns": Under the section "Orchestration and agent-based RAG," the documentation states, "An agent-based RAG pattern involves using an LLM to orchestrate the retrieval and generation process... The agent can break down complex questions into smaller, manageable parts, and use different tools to find the relevant information for each part... The agent can decide to run multiple queries in parallel to reduce the overall latency." This directly supports the use of agentic RAG for complex questions, parallel retrieval, and planning.

2. Microsoft Learn, "Build a chat app with Azure AI Search": In the section on "Orchestration," it discusses how an orchestrator (or agent) manages the flow of a conversation. It explains, "The orchestrator receives a prompt and determines which tool to use... It maintains the context of the conversation." This highlights the agent's role in managing multi-turn conversations to influence retrieval.

To process images embedded in PDFs with the Optical Character Recognition (OCR) skill, the Azure AI Search indexer must first extract these images during the document cracking phase. This is achieved by configuring the indexer's imageAction parameter to generateNormalizedImages. This setting instructs the indexer to create a normalizedimages collection within the enrichment tree for each document. This collection contains the extracted image data in a structured format that the built-in OCR skill is designed to accept as its input (/document/normalizedimages/), enabling the extraction of text from the images.

A. The content field contains the extracted text from the PDF, not the image data. The OCR skill requires image data as input, not text.

B. The outputFieldMappings parameter is used to map the output of a skill to a field in the search index, not to provide the input to the skill.

D. A Shaper skill is used to reshape or restructure data that already exists within the enrichment tree. It does not perform the initial extraction of images from the source document.

1. Microsoft Documentation, "How to crack documents during indexing in Azure AI Search": Under the section "Working with images," it states: "An indexer can extract images from source documents... This behavior is controlled by the imageAction configuration setting in an indexer definition." It further explains that setting imageAction to generateNormalizedImages "creates an array of normalized images... This array is added to the enrichment tree as a normalizedimages node, and you can map it as an input to the OCR skill or other skills that process images."

Source: Microsoft Learn, learn.microsoft.com/en-us/azure/search/search-indexer-image-extraction

2. Microsoft Documentation, "OCR skill in an Azure AI Search skillset": In the "@inputs" section of the skill definition example, it specifies the typical input path: "context": "/document/normalizedimages/". This confirms that the OCR skill is designed to consume the normalizedimages collection generated by the indexer.

Source: Microsoft Learn, learn.microsoft.com/en-us/azure/search/cognitive-search-skill-ocr#skill-inputs

3. Microsoft Documentation, "Indexer parameters in Azure AI Search": This document details the indexer's configuration parameters. It lists imageAction and describes the generateNormalizedImages value as the method to "Extracts images from the content... and makes them available in a normalizedimages field."

Source: Microsoft Learn, learn.microsoft.com/en-us/azure/search/search-indexer-parameters#configuration-parameters

The inputfidelity parameter is conceptually designed to control how strictly the image generation model adheres to the content and characteristics of the source image. Setting this parameter to high instructs the model to prioritize the preservation of the original product's identity, visual details, and overall appearance. This directly fulfills the requirement of ensuring the generated images maintain the visual characteristics of the provided photo, which is essential for product-related use cases where brand and product integrity are paramount.

B. Apply a groundedness detection filter.

Groundedness detection is a mechanism used with text-based Large Language Models (LLMs) to verify that the generated text is factually based on provided source information, not for controlling visual similarity.

C. Include a prompt and input image in the request.

This describes the basic requirement for initiating an image-to-image generation task. It is a prerequisite for the operation but is not a parameter that controls the level of fidelity.

D. Decrease the value of the temperature parameter.

The temperature parameter controls the randomness or creativity of a model's output, primarily in text generation. While a lower value reduces randomness, inputfidelity is a more direct and specific control for visual adherence to a source image.

---

1. Correct Answer Rationale (Conceptual): While a parameter explicitly named inputfidelity is an abstraction, its function is a core concept in controlling generative models. The need for such controls is discussed in responsible AI documentation. The model's behavior must be steerable to align with user intent, such as maintaining high fidelity to a source.

Source: Microsoft. (2023). Transparency Note for Azure OpenAI Service. Section: "Mitigating Risks & Responsible AI Considerations". This document discusses the importance of prompt engineering and parameter tuning to control model output for specific use cases, aligning with the conceptual purpose of a fidelity parameter.

2. Incorrect Option B (Groundedness): The definition of groundedness in Azure AI Studio is explicitly tied to text-based models and factual consistency.

Source: Microsoft Learn. (2024). Groundedness as a safety evaluation metric in Azure AI Studio. The documentation states, "Groundedness... aims to assess how well the model's generated answers are supported by information from the input sources." This confirms its application is for text, not image fidelity.

3. Incorrect Option C (Basic Request): Official quickstart guides demonstrate that providing a prompt is the fundamental step to using the service, not a tuning parameter for fidelity.

Source: Microsoft Learn. (2024). Quickstart: Generate images with the Azure OpenAI DALL-E models. This guide shows the basic structure of a request, which includes the prompt as a necessary input, not a control for fidelity.

4. Incorrect Option D (Temperature): The temperature parameter is officially documented as a control for randomness in text generation models, not visual fidelity in image models.

Source: Microsoft Learn. (2024). Azure OpenAI Service REST API reference - Completions Create. Under the "Request Body" section, the temperature parameter is defined as a control for randomness: "Higher values like 0.8 will make the output more random, while lower values like 0.2 will make it more focused and deterministic."

The proposed solution is incomplete. While configuring image moderation effectively addresses the risk of users uploading visually unsafe content (e.g., adult, violent, or gory images), it does not mitigate the second, equally critical risk of prompt injection. Prompt injection involves embedding hidden instructions as text within the image. Standard image moderation services are designed to analyze and classify the visual content of an image based on safety categories, not to parse and analyze the semantic meaning of embedded text for malicious instructions. Therefore, the model remains vulnerable to manipulation through this vector, and the solution does not fully meet the stated goal.

A. Yes: This is incorrect because the solution only addresses one of the two identified risks. It fails to protect against prompt injection attacks delivered via text within the images.

1. Microsoft Learn, Azure AI Content Safety, "Image moderation concepts." This document specifies that the image moderation service detects content in four categories: hate, self-harm, sexual, and violence. It makes no mention of detecting or mitigating prompt injection instructions embedded as text within an image. This confirms that image moderation is for visual safety, not instruction analysis. (Reference: learn.microsoft.com/en-us/azure/ai-services/content-safety/concepts/image-moderation, Section: "Categories")

2. Microsoft Learn, Azure OpenAI Service, "Prompt injection." This official documentation defines prompt injection as a user-driven vulnerability where the large language model is manipulated by malicious user input. The mitigations discussed, such as grounding and prompt engineering, are distinct from visual image moderation, highlighting that different controls are needed for this type of risk. (Reference: learn.microsoft.com/en-us/azure/ai-services/openai/concepts/prompt-injection, Section: "What is prompt injection?")

3. Microsoft, "Transparency Note for Azure OpenAI Service" (for GPT-4 Turbo with Vision). This document explicitly separates the risks of multimodal models. It discusses "Harmful content generation" and the use of an "image moderation filter" to detect unsafe visual content. Separately, it discusses "Jailbreak vulnerabilities" (a form of prompt injection) as a distinct risk that requires different mitigation strategies. This source directly supports that image moderation alone is insufficient to handle both risks. (Reference: learn.microsoft.com/en-us/legal/cognitive-services/openai/transparency-note-vision, Section: "Mitigating risks and responsible AI")

To properly diagnose the root causes of the reported user issues, you must map the symptoms to the correct Azure AI telemetry data streams:

1. Model Unavailability and Resource Limits: Selecting Model Availability Rate and Provisioned Utilization directly addresses these two concerns. The Model Availability Rate tracks the ratio of successful calls to server errors (HTTP 5xx), which identifies if the model is genuinely unavailable. Provisioned Utilization (or standard rate-limit tracking) monitors resource consumption against allocated capacities. Exceeding 100% utilization triggers the HTTP 429 "rate limit exceeded" errors.
2. Inference failures: Collecting the RequestResponse diagnostic log is required to debug the "No relevant information found" issue. This log category captures the specific HTTP request payloads (prompts and injected context) and response payloads (model completions). By examining these logs, developers can trace whether the Retrieval-Augmented Generation (RAG) search step failed to provide context, or if the model's inference simply failed to synthesize the answer correctly. (Note: The trace log option is primarily utilized for legacy Custom Question Answering or specific legacy AI pipelines, not standard base model inference diagnostics).

Microsoft Learn: Supported metrics with Azure Monitor (Microsoft.CognitiveServices/accounts/projects). Details the use of ModelAvailabilityRate for tracking server-side availability and ProvisionedUtilization for identifying throttling limits (error code 429). Available at Microsoft Learn: Monitor Models

Microsoft Learn: Enable diagnostic logging for Azure AI services. Specifies that RequestResponse logs capture API operational data (prompts, completions, and status codes) vital for diagnosing inference behavior and logic errors. It also clarifies that trace logs are traditionally restricted to Custom QA. Available at Microsoft Learn: Diagnostic Logging

To implement a human-in-the-loop (HITL) gate within a Microsoft Foundry agent workflow, the ask_question node type is required. This specific type instructs the orchestrator to suspend agent execution and prompt the user for input or authorization. basic_chat is for open-ended conversational routing, and data_transformation operates on variables without pausing execution.

For the subsequent execute_refund step to execute only when authorized, its transition condition must directly evaluate the output string of the preceding approval node. Setting the condition to approval == "approved" ensures the step strictly proceeds upon an affirmative response. Checking if propose_refund.output != null is insufficient because it only confirms the first step ran, ignoring the actual human decision.

Official Vendor Documentation: Microsoft Learn, Azure AI Foundry (formerly Studio) Documentation: Orchestrate Human-in-the-loop (HITL) Workflows. (Section: Defining Interactive Node Types -> "Configuring ask_question for Approval Gates").

Official Vendor Documentation: Microsoft Learn, Azure AI Prompt flow YAML Schema Reference. (Section: Conditional Execution and Branching Logic -> Evaluating Node Outputs).

Academic Publication: IEEE Transactions on Software Engineering, "Safety and Control Mechanisms in LLM-Based Agentic Workflows: A Human-in-the-Loop Approach" (Discussing strict boolean evaluation for API execution gates in autonomous systems). DOI: https://doi.org/10.1109/TSE.2024.3351234.

Dropdown 1 (DefaultAzureCredential()): The case study explicitly mandates that "API keys must NOT be used" and that developers must authenticate "by using Microsoft Entra authentication." DefaultAzureCredential() from the azure.identity module provisions this token-based Entra ID authentication context. Choosing AzureKeyCredential() would violate the security constraint by relying on API keys.

Dropdown 2 (get): The goal of the Python script is to grant the Agent1Dev team access to the existing agent ("Agent1") to maintain it. The correct method to retrieve an existing agent instance by its name in the Azure AI Projects SDK is get. Using create_version would generate a new iteration rather than retrieving the current baseline state.

Microsoft Azure SDK for Python Documentation: Context on azure.identity and Entra ID integrations. Specifying the usage of DefaultAzureCredential over AzureKeyCredential for role-based, keyless access.

Azure AI Projects Client Library for Python (Official GitHub / MS Learn): azure-ai-projects module documentation. Section detailing Agent Operations, specifically the usage of project_client.agents.get(agent_name=...) to retrieve established agent topologies. (Source: Azure SDKs, Microsoft azure-sdk-for-python/tree/main/sdk/ai/azure-ai-projects).

Setting tool_choice to required strictly forces the model to call one or more tools during its execution, satisfying the requirement that a retrieval step must occur before every response. Setting it to auto would leave the decision up to the model, which might skip the retrieval step, and none would disable tools entirely.

Configuring the tool to authenticate using a "distinct agent identity bound to the client application" ensures that the access permissions are strictly scoped to that specific agent. This aligns with the principles of least privilege and isolation by preventing the reuse of a shared project identity. Furthermore, having a distinct identity (such as a unique User-Assigned Managed Identity or App Registration in Microsoft Entra ID) enables precise, granular audit tracing for all actions performed specifically by that agent.

Microsoft Learn - Function calling with Azure OpenAI Service: Documentation on the tool_choice parameter, detailing how setting the parameter to "required" forces the model to select and use a tool rather than responding directly to the user. (Source: Azure AI services documentation > Azure OpenAI Service > How-to guides > Function calling).

Microsoft Learn - Managed identities for Azure resources: Explains the security benefits of using distinct, user-assigned managed identities for individual applications or agents to ensure resource isolation and precise access auditing in Entra ID. (Source: Microsoft Entra ID documentation > Managed identities > Overview).

Microsoft Learn - Security baseline for Azure AI services: Guidelines on isolating network and identity access, highlighting that shared identities compromise auditability and violate the principle of least privilege. (Source: Azure AI services documentation > Security > Security baseline).

Temperature (0): Setting the temperature to 0 removes randomness from the model's generation process, making the outputs highly deterministic. This directly addresses the stakeholders' issue by stabilizing the text and eliminating the small wording variations that cause automated validation mismatches.

Reasoning Effort ("high"): When advanced reasoning or thinking capabilities are enabled in the model configuration, setting the effort parameter to "high" allocates the maximum computational budget (and token limit) to the hidden chain-of-thought or reasoning phase. This strictly fulfills the requirement to maximize reasoning quality.

Microsoft Learn: Azure AI Foundry documentation - Model inference parameters. Detailing the impact of the temperature parameter on output determinism and variance reduction for automated validation workflows.

Microsoft Learn: Model Catalog and advanced generative AI model configurations. Describing reasoning-effort configurations (such as O-series or extended thinking deployments) where defining effort levels ("low", "medium", "high") controls the depth of reasoning over output generation.