Microsoft Certified: Azure AI Apps and Agents Developer Associate (beta) — Practice Set 1

Question 1

A team is building a new Azure AI solution that requires large language model (LLM) inference, vector search for RAG grounding, and agent workflow orchestration — all from a single unified platform. Which Azure service should they use as the primary hub for these capabilities?

Accepted Answer

Microsoft Foundry is the unified platform for building AI apps and agents on Azure, integrating LLM inference, vector search, grounding, and agent workflow orchestration. Azure Machine Learning Studio focuses on training and MLOps. Azure Cognitive Services (multi-service) provides pre-built AI APIs but not the full agent/workflow orchestration hub. Azure Synapse Analytics is primarily a data analytics platform.

Question 2

A developer needs to choose between deploying a large language model and a small language model for a customer-facing chatbot that runs on edge devices with limited compute. Which consideration most directly drives the selection of a small language model?

Accepted Answer

Small language models are designed to operate within tighter compute and memory budgets and therefore have lower latency and lower resource requirements — the key factor for edge devices. LLMs generally support longer context windows, not small models. Accuracy depends on the task and the specific models compared; neither category is universally more accurate. Multimodal capability is a feature of specific multimodal models, not a general property of small language models.

Question 3

A company's AI agent is responding slowly during peak hours because the Azure OpenAI model deployment is hitting its tokens-per-minute (TPM) limit. The agent handles customer service requests. The team wants to maintain response quality while reducing throttling. Which action should the team take first?

Accepted Answer

When a deployment is hitting its TPM (tokens per minute) quota limit, the correct remediation is to request a quota increase for the model deployment through Foundry — this directly addresses the rate-limit constraint without sacrificing quality. Switching to a smaller model (A) would reduce quality. Disabling logging (C) has no effect on TPM limits and removes important monitoring. Removing system prompt instructions (D) degrades agent behavior and only marginally reduces token usage per call.

Question 4

A security architect is reviewing an AI solution deployed in Foundry. They want to ensure that the solution follows the principle of least-privilege and avoids storing long-lived secrets. Which TWO configurations should be applied? Choose 2.

Accepted Answer

Managed identities (A) allow the application to authenticate to Azure services without storing credentials; they are the recommended approach for least-privilege access. Keyless credentials (C) eliminate shared secrets passed at runtime, reducing the attack surface. Storing API keys in container images (B) embeds secrets into artifacts, which is a security anti-pattern. Assigning Contributor at the subscription level (D) violates least-privilege. Private networking (E) is a good defence-in-depth practice but does not on its own address the secret-storage or identity concerns raised in the question.

Question 5

An AI operations team wants to detect when a deployed LLM starts producing responses that significantly differ from an established quality baseline — for example, answers becoming less factual over time. Which monitoring capability in Foundry addresses this?

Accepted Answer

Monitoring model performance and drift is the specific Foundry monitoring capability that tracks degradation in output quality, factual accuracy, or behavioral shift over time relative to a baseline. Data ingestion throughput (A) relates to index health, not response quality. Cost footprints (C) track spending. CI/CD pipeline metrics (D) are DevOps concerns unrelated to response quality drift.

Question 6

An AI governance team wants to ensure that an autonomous agent cannot execute a tool call that transfers funds above a certain dollar threshold without explicit human review. Which responsible AI mechanism in Foundry should be implemented?

Accepted Answer

Agent oversight modes, constraints, and approval workflows are the correct governance controls — they allow fine-grained control over which tool calls require human approval and which can proceed autonomously, directly addressing the fund-transfer scenario. Safety filters (A) target content moderation, not tool-call authorization. Disabling all tool access (C) would render the agent unable to function. Temperature (D) is a generation parameter affecting randomness, not compliance or authorization logic.

Question 7

A team must integrate a Foundry project into an existing Azure DevOps workflow so that model deployments and agent configurations are versioned and deployed automatically on each merge. Which Foundry capability enables this?

Accepted Answer

Foundry supports integration with CI/CD pipelines, allowing teams to automate versioning and deployment of model and agent configurations via Azure DevOps or GitHub Actions. Managed identity (B) is an authentication concern. Drift monitoring (C) is an operational observability feature. Private networking (D) is a security/network configuration, not a deployment automation feature.

Question 8

A developer wants to build a Q&A application over a private document corpus. The documents are ingested, chunked, and stored with vector embeddings. At query time, the most relevant document chunks are retrieved and injected into the LLM prompt. What pattern is being implemented?

Accepted Answer

Retrieval-augmented generation (RAG) is the pattern where relevant document chunks are retrieved from an indexed store and injected into the prompt to ground LLM responses. Fine-tuning (A) adapts model weights, not the inference-time retrieval pattern. Chain-of-thought prompting (C) asks the model to reason step-by-step but does not involve external retrieval. Prompt caching (D) is a cost/latency optimization technique, not a retrieval pattern.

Question 9

A developer is building an AI agent in Foundry that must answer questions about recent financial filings. The agent should be able to call a custom Python function that queries a live financial data API, in addition to using its pre-trained knowledge. The developer wants the agent to decide at runtime when to call the function. Which agent capability is required?

Accepted Answer

Function-calling (also called tool use) allows an agent to decide at runtime whether to invoke a registered external function — exactly the requirement for querying a live API. Fine-tuning (B) bakes knowledge into model weights, which cannot reflect live data. A fixed prompt template (C) can't adapt to different queries and will exceed context limits. Safety filters (D) govern content moderation, not data retrieval logic.

Question 10

A team is evaluating the output quality of their RAG-based generative application. They want to detect when the model produces a response that is not supported by the retrieved documents. Which evaluation metric should they measure?

Accepted Answer

Detecting fabrications — also referred to as measuring groundedness — is the evaluation technique for identifying when the model's response is not supported by the retrieved context. This is a core evaluation task for RAG applications. Latency (A) and token cost (C) measure performance and cost, not factual grounding. Index size (D) is a storage metric unrelated to response quality.

Question 11

A development team is designing a multi-agent solution in Foundry where a primary orchestrator agent delegates sub-tasks to specialized agents. Which TWO capabilities must the team implement to support this architecture? Choose 2.

Accepted Answer

Implementing orchestrated multi-agent solutions (A) and defining agent roles, goals, conversation-tracking approach, and tool schemas (B) are the two foundational requirements for a multi-agent architecture. Without orchestration logic and well-defined roles/schemas, agents cannot coordinate. Prompt caching (C) is an optimization, not a core architectural requirement. Fine-tuning every sub-agent (D) is not required and is not part of a multi-agent pattern. Billing alerts (E) are operational concerns, not agent design concerns.

Question 12

An agent is deployed to autonomously process insurance claims. After a claim is extracted and validated, the agent must request manager approval before making any payment decision. Which design pattern should be applied?

Accepted Answer

Building semiautonomous workflows with safeguards and approval flow controls is the pattern designed precisely for this scenario — the agent processes the claim autonomously but gates payment decisions behind a human approval step. Fully autonomous workflows without safeguards (A) bypass the required human review. Removing tool access (C) prevents any payment action at all. Chain-of-thought prompting (D) only changes the model's reasoning style; it doesn't create a real approval gate involving a human.

Question 13

A developer wants to improve the factual accuracy of an agent's responses by having the model first draft an answer, then critique that draft for errors, and then revise it. Which technique does this describe?

Accepted Answer

Implementing model reflection, chain-of-thought evaluations, and self-critique loops describes the technique where the model evaluates and revises its own output — exactly the draft-critique-revise pattern described. Prompt caching (A) is a latency/cost optimization. Hybrid search (C) and semantic chunking (D) are retrieval-pipeline techniques, not inference-time self-improvement patterns.

Question 14

A production AI system combines an LLM with a rules engine for compliance decisions. The team wants end-to-end observability across both components, capturing which model was called, the token counts, latency at each step, and any safety signals triggered. What should the team implement?

Accepted Answer

Implementing tracing, token analytics, safety signals, and latency breakdowns provides the end-to-end observability needed across the LLM and rules engine components. A cost dashboard (B) does not capture token-level or safety signal data. Disabling logging except safety filters (C) eliminates the tracing and latency data the team needs. CPU/memory alerts (D) are infrastructure metrics, not AI-specific observability.

Question 15

A developer wants to connect a Python application to a Foundry project so it can call LLM deployments and agent endpoints using official SDKs. What is the correct approach?

Accepted Answer

The correct approach is to configure the application to connect to a Foundry project using Foundry SDKs and connectors, which provide typed clients for model inference and agent interactions. Azure Resource Manager APIs (B) manage infrastructure, not inference requests. Azure Data Factory (C) is a data integration service not designed for LLM inference orchestration. Embedding model weights (D) is not how Azure-hosted models are consumed.

Question 16

A marketing team wants to generate product images from text descriptions and then allow users to modify specific regions of the image — for example, changing the background — without regenerating the entire image. Which computer vision workflow should the developer implement?

Accepted Answer

The requirement covers two distinct steps: generating an image from a text prompt, then selectively editing regions via inpainting and mask-based edits — both of which are supported in Foundry's image generation and editing workflows. Azure AI Video Indexer (B) processes video, not static image generation. Custom Vision (C) classifies images rather than generating or editing them. Azure AI Language (D) produces text descriptions, not visual edits.

Question 17

An accessibility team needs a solution that automatically generates detailed alt-text for every product image on an e-commerce site. The alt-text must align with WCAG accessibility guidelines and describe key visual features. Which capability should the developer configure?

Accepted Answer

Configuring multimodal models to generate alt-text and extended image descriptions aligned to accessibility guidelines is the exact Foundry capability for this use case. OCR (B) only extracts text visible in the image, not visual descriptions of the image content. Content safety filters (C) detect harmful content rather than generating descriptions. A Custom Vision classifier (D) would tag images with binary labels, not produce descriptive alt-text.

Question 18

A developer is building a solution that must detect and locate specific product components within images captured on a factory assembly line. Which Foundry capability should be configured?

Accepted Answer

Implementing solutions that identify objects, components, or regions within images is the Foundry capability that addresses part detection and localization on an assembly line. Single-task Content Understanding pipelines (A) are used for extracting visual characteristics in a pipeline configuration context, but the direct answer to component detection and localization is option B. Inpainting (C) is for editing image regions, not detecting them. Azure AI Translator (D) translates text, not image content.

Question 19

A security team wants to ensure that AI-generated images cannot contain unauthorized logos or brand marks from third parties. Which responsible AI capability for multimodal content should be configured?

Accepted Answer

Enforcing visual policy rules — including flagging prohibited symbols and upholding brand usage requirements — is a responsible AI capability for multimodal content that directly addresses unauthorized logo detection and suppression. RAG grounding (B) retrieves documents, it cannot enforce visual output policies. OCR (C) would only catch brand names that appear as readable text, not logos as visual elements. Temperature (D) affects generation randomness and does not reliably prevent specific visual outputs.

Question 20

A developer needs to extract named entities (such as people, organizations, and locations) and generate a concise summary from customer support ticket text using generative AI capabilities in Foundry. Which approach should be used?

Accepted Answer

Using generative prompting and Foundry Tools to extract entities, topics, summaries, and structured JSON outputs is the recommended approach for this text analysis task. Azure Cognitive Search (B) performs keyword or semantic search but does not produce entity extractions or summaries. Training a custom NER model from scratch (C) is unnecessary overhead when Foundry's generative capabilities can accomplish the task. Azure Event Grid (D) is an eventing service, not an NLP processing tool.

Question 21

A multinational company wants to enable its AI assistant to answer users in their native language. The assistant receives queries in English but must reply in the language detected in the user's message. The solution should use Azure services integrated with Foundry. Which implementation is appropriate?

Accepted Answer

Building translation flows using Azure Translator in Foundry Tools or LLM-powered translation is the correct approach for translating assistant responses into the user's detected language. Azure Cognitive Search (B) retrieves indexed content but does not translate it. Speech-to-text (C) converts audio to text, not text to another language's text. Azure Event Hubs (D) is a message streaming service, not a translation service.

Question 22

An AI assistant deployed for a legal firm must detect when users express negative sentiment about a case outcome AND flag any sensitive personal information (PII) in responses before they are sent. Which TWO text analysis capabilities should be configured? Choose 2.

Accepted Answer

Configuring sentiment and tone detection (A) addresses the requirement to identify negative user sentiment. Configuring detection of sensitive content and safety issues (B) addresses PII flagging in responses. Semantic chunking (C) and hybrid search (D) are retrieval-pipeline optimizations, not text analysis capabilities for this scenario. Content Understanding pro-mode (E) is a computer vision/multimodal pipeline, not relevant to text sentiment or PII detection.

Question 23

A developer is building a RAG pipeline that must ingest scanned PDF invoices. The invoices contain both printed text and handwritten annotations. Which ingestion technique is required to make this content searchable?

Accepted Answer

Configuring RAG ingestion flow with OCR is the correct approach for scanned PDFs with printed and handwritten content — OCR converts the visual text to machine-readable characters that can then be chunked, embedded, and indexed. Indexing binary content (B) does not extract text from image-based PDFs. Service Bus (C) is a messaging service, not a text extraction tool. Event Grid (D) is an eventing trigger that could notify a pipeline, but by itself does not perform OCR.

Question 24

A developer needs a search pipeline that can find relevant document chunks using both keyword matching and semantic similarity simultaneously. Which search configuration should be applied?

Accepted Answer

Hybrid search combines keyword search (exact term matching) with vector search (semantic similarity), providing the best of both retrieval approaches for RAG grounding. Keyword-only search (A) misses semantically similar content that uses different words. Semantic-only search (B) can miss exact-match requirements. The Azure Cognitive Services Read API (D) is an OCR tool for extracting text from images, not a search configuration.

Question 25

A developer is building a document processing pipeline that must extract key fields from insurance forms (structured) and also produce clean markdown representations of the extracted content for downstream agent reasoning. Which TWO Azure AI Document Intelligence / Content Understanding capabilities should be used? Choose 2.

Accepted Answer

Extracting information using multimodal pipelines combining OCR, layout analysis, and field extraction (A) handles the structured form field extraction requirement. Implementing analyzers for generating structured or markdown outputs using Content Understanding (B) produces the clean markdown representations needed for agent reasoning. Event Grid (C) can trigger processing but is not an extraction capability. Service Bus (D) routes messages after extraction but is not an extraction tool. Model drift monitoring (E) is an operational concern, not an extraction capability.

AI-103 — Azure AI Apps and Agents Developer Associate