Welcome to Part 1 of our complete study series for the AI-103 Certification (Developing AI Apps and Agents on Azure, replacing the retired AI-102 exam). Selecting the right foundation model and hosting architecture inside Microsoft Foundry (formerly Azure AI Foundry) is the core of Domain 1. This free module covers model selection, serverless pay-per-token vs. provisioned throughput (PTU), managed compute, and agent grounding.
 |
| Microsoft Foundry Model Selection: AI-103 Certification Guide & Practice Questions |
In this free AI-103 practice exam module (Part 1), you will solve realistic scenario questions tackling Microsoft Foundry model selection, Serverless APIs, Managed Compute, and grounding services. Every question comes with verified rationales for both correct and incorrect options.
Microsoft Foundry & AI-103 Practice Questions (Domain 1)
Q1: Grounding Agents on Proprietary Storage (Azure AI Search)
You are designing an AI agent named Agent1 using Azure AI Foundry.
The agent must be able to answer user queries based on proprietary product information documents stored in an Azure Storage account named storage1. You need to recommend a solution that allows the agent to index, search, and retrieve this internal data to ground its responses.
Which service should you recommend?
• A is correct: Azure AI Search is the primary information retrieval system for Azure AI solutions. It can directly index data stored in Azure Storage and provides the vector or hybrid search capabilities required to ground AI agents on internal, proprietary data (Retrieval-Augmented Generation or RAG).
• B is incorrect: Azure AI Translator is used to translate text between different languages. It does not provide indexing or search capabilities for data stored in Azure Storage.
• C is incorrect: Azure AI Document Intelligence is used to extract text, key-value pairs, and structure from documents. While it can be used during the data ingestion phase to prepare files, it does not act as the search index or retrieval engine that the agent queries at runtime.
• D is incorrect: Grounding with Bing Search connects the model to the public internet to answer questions based on current web data. It cannot be used to search private, internal data stored in an Azure Storage account.
Q2: Dedicated GPU Capacity for Open-Weight Models (Managed Compute)
A team wants to deploy an open-weight model from the Hugging Face collection in Microsoft Foundry (for example, a Qwen model). The model isn't offered as a pay-per-token model. It must run on dedicated GPU capacity that Foundry manages, and the team doesn't want to create, size, or patch virtual machines.
Which option should you use?
• D is correct: Managed compute serves open-source, partner, and custom models on dedicated GPUs, and Foundry takes care of sizing, provisioning, and patching. It bills hourly per accelerator SKU, and its quota is separate from your Azure VM quota.
• A is incorrect: Serverless API is the preferred option for Foundry Models (models sold by Azure and select partner models), billed per token or through PTUs. It doesn't give you dedicated GPUs for a Hugging Face model.
• B is incorrect: Instant access calls supported models by name without creating any deployment, so there's no dedicated capacity.
• C is incorrect: AKS could work, but you would own the nodes and the serving runtime, which the requirement rules out.
Q3: Code Generation & Refactoring (Generative Language Models)
An engineering team is designing an autonomous software development extension for an integrated development environment (IDE). The solution requires an underlying language model capable of syntax-aware code generation, multi-token autocompletion, docstring synthesis, and automated unit test refactoring across multiple programming languages.
Which model category from the Azure AI Foundry model catalog should be deployed?
• C is correct: Code-optimized models (like DeepSeek-Coder or CodeLlama) and frontier generative LLMs (like GPT-4o) are trained on massive corpora of source code repositories, abstract syntax trees (ASTs), API documentation, and algorithmic logic. This generative language category is the only one capable of understanding code semantics, auto-completing tokens, synthesizing documentation, and refactoring unit tests.
• A is incorrect: Dense text-embedding models produce numeric vector representations designed for semantic search, similarity matching, and retrieval (RAG). They cannot generate, autocomplete, or refactor code.
• B is incorrect: Neural text-to-speech models convert written text into spoken audio; they lack code comprehension and text generation capabilities.
• D is incorrect: Computer vision segmentation models process visual data to identify pixel-level object boundaries in images; they cannot process or author source code.
Q4: Data Residency & Version Update Policy (Data Zone Standard)
A financial services firm in Frankfurt is deploying a customer-facing chatbot in Microsoft Foundry. The deployment must meet these requirements:
• Prompts and responses are processed only within the Azure EU Data Boundary.
• Traffic is unpredictable, and the firm doesn't want to buy reserved capacity.
• The chatbot was validated against one specific model version. It must stay on that version for as long as Microsoft supports it, but the deployment must not stop working when that version is retired.
How should you configure the deployment?
Configuration 1: Deployment type
Configuration 2: Version update policy
Correct Answer:
• Deployment type: Data Zone Standard
• Version update policy: Once the current version expires
• Data Zone Standard is correct: it's pay-per-token, and the service processes prompts and responses only within the chosen data zone. For the EU, that zone follows the Azure EU Data Boundary.
• Global Standard is incorrect: global types may process data in any Azure region.
• Global Provisioned is incorrect: it reserves PTU capacity, which the firm doesn't want, and it also processes data in any Azure region.
• Once the current version expires is correct: the deployment stays on your version until it's retired, then updates automatically to the next available version, so there's no outage.
• Upgrade once a new default version becomes available is incorrect: it moves the deployment as soon as a new default version is set, so prompts can break before you've retested them.
• Opt out of automatic model version upgrades is incorrect: it never upgrades on its own, but once the version is retired the deployment stops accepting requests and returns errors.
Q5: Query Complexity & Routing Strategy (Model Cascade)
An e-commerce mobile application processes over 300,000 chat inquiries per day. Telemetry shows that 80% of interactions are repetitive queries regarding shipping policies and store hours, while the remaining 20% involve complex multi-order returns and disputes requiring deep logical reasoning.
The engineering team needs an architectural strategy that minimizes per-query costs and Time-to-First-Token (TTFT) latency while preserving high answer quality for difficult cases.
Which model selection and routing strategy should the team implement?
• Why B is correct: A model cascade (or dynamic model router) uses a lightweight classifier or an SLM to evaluate query complexity first. Straightforward, high-volume FAQs are routed to fast, cost-effective models (such as Phi-4 or GPT-4o-mini), dramatically lowering latency and token spend. Only complex, ambiguous requests are routed to resource-intensive frontier models (such as GPT-4o), preserving overall response quality at a fraction of the cost.
• Why A is incorrect: Forcing all queries through a small language model degrades output quality on multi-step reasoning, return disputes, and complex context synthesis, violating the requirement to maintain quality on complex tasks.
• Why C is incorrect: Routing 100% of traffic to a frontier model incurs high financial costs and higher average latency for simple questions. Setting the temperature to 0 merely makes outputs deterministic; it does not reduce token consumption or model execution overhead.
• Why D is incorrect: Fan-out routing with multi-model consensus multiplies API token costs by at least 3x and increases overall latency to match the slowest model in the pool, directly opposing the optimization goals.
Q6: Root-Cause Analysis Agent (LLM vs SLM vs Key Phrase Extraction)
An IT operations team is selecting a foundation model from the Azure AI Foundry catalog to power an automated infrastructure root-cause analysis agent. The agent must parse multi-page system event logs, trace interdependencies across 50,000 tokens of diagnostic context, and write comprehensive step-by-step remediation plans.
For each of the following statements regarding model selection for this workload, select Yes if the statement is true. Otherwise, select No.
1. A Large Language Model (LLM) is the appropriate model type because the workload requires deep, multi-step causal reasoning across long context windows.
2. A prebuilt Key Phrase Extraction model can serve as the primary inference engine to diagnose root causes and author remediation plans.
3. A Small Language Model (SLM) is preferred over an LLM for this task because SLMs demonstrate superior performance on long-context causal chain analysis.
• Statement 1 is Yes: Large Language Models (LLMs) possess the architectural scale, parameter depth, and extended context window support necessary to correlate events across tens of thousands of tokens and generate coherent, multi-step natural language explanations.
• Statement 2 is No: Key Phrase Extraction is a specialized NLP service designed solely to extract prominent keywords and entities from text. It lacks generative text synthesis capabilities and cannot perform causal reasoning or author procedural guides.
• Statement 3 is No: While Small Language Models (SLMs) offer lower latency and smaller compute footprints for targeted tasks, they have constrained reasoning capacity over massive context windows compared to frontier LLMs.
Q7: Reasoning Models Optimization (reasoning_effort Parameter)
A developer is writing code to invoke a reasoning model (such as o3 or o4-mini) in Azure OpenAI within Azure AI Foundry. The application task requires generating a concise, three-bullet summary of a short email notification.
The developer wants to minimize latency and token cost for this simple task while still using a reasoning-capable model.
Which parameter should the developer configure, and to what value?
• Why B is correct: Reasoning models in Azure OpenAI (o1, o3, o3-mini, o4-mini) support a dedicated reasoning_effort parameter, which can be set to low, medium, or high. Setting it to "low" constrains the internal reasoning-token budget the model spends before answering, which is appropriate for a simple, low-complexity task like summarizing a short email — reducing both latency and cost.
• Why A is incorrect: Reasoning models do not support the temperature or top_p parameters — these are automatically ignored or excluded by the API for o-series models, so setting them has no effect on behavior.
• Why C is incorrect: A "high" reasoning effort causes the model to spend a larger share of its token budget on internal reasoning, which is unnecessary and wasteful for a trivial summarization task.
• Why D is incorrect: presence_penalty and frequency_penalty influence word/token repetition in the output; they have no effect on reasoning depth, latency, or the internal thinking-token budget.
Q8: Exact Matching vs Meaning (Hybrid Search + Semantic Ranker)
A support portal indexes 40,000 product manuals in Azure AI Search for a Foundry agent. Users search with exact model numbers such as "XR-4821-B", and also with descriptions such as "the projector that shuts off after an hour". Which retrieval configuration gives the agent the most relevant grounding data?
• A is correct: hybrid search runs a full-text query and a vector query in parallel and merges the results with Reciprocal Rank Fusion. Keyword matching is stronger for exact terms like product codes, and vectors find text that is close in meaning. The semantic ranker then reranks the merged results.
• B is incorrect: vector search matches meaning, but it can miss an exact identifier such as a model number.
• C is incorrect: keyword-only search misses a description that shares no words with the manual.
• D is incorrect: the semantic ranker reranks results from a full-text or hybrid query, so it can't retrieve anything on its own.
Q9: Bursty & Unpredictable Workloads (Global Standard)
A university builds a study-helper bot on Microsoft Foundry. Usage is unpredictable: almost none on weekends and heavy bursts before exams. There's no data residency requirement, the team wants access to new models as early as possible, and it wants to pay only for what it consumes.
Which deployment type should you choose?
• C is correct: Global Standard is pay-per-token with no reserved capacity, and Microsoft recommends starting there for most workloads: new models land on it first, it's the cheapest, and it covers the most regions.
• A and D are incorrect: provisioned types reserve PTU capacity for predictable, high-volume throughput, so you would pay for it on quiet weekends too.
• B is incorrect: Global Batch is about 50% cheaper but asynchronous, with a 24-hour target turnaround, so it doesn't suit an interactive chat.
Q10: High-Volume Predictable Latency (Provisioned Throughput Units - PTUs)
A healthcare portal is deploying a patient intake assistant expected to process sustained high-volume query traffic with millions of tokens per hour. Enterprise SLAs mandate predictable sub-second latency and guarantee that inference calls will never be throttled by noisy-neighbor contention during peak business hours.
Complete the architecture recommendation by selecting the correct option.
To guarantee dedicated compute capacity, eliminate multi-tenant throttling risks, and achieve predictable latency for this high-volume workload, the architect must deploy the model using:
• Why B is correct: Provisioned Throughput Units (PTUs) reserve dedicated processing capacity for the deployment, providing predictable throughput and lower latency variance for sustained workloads. This makes PTUs appropriate when an application requires dedicated capacity and consistent performance at scale.
• Why A is incorrect: Standard pay-as-you-go deployments use shared capacity and can experience variable latency and throttling when demand exceeds available capacity.
• Why C is incorrect: The Batch API is designed for asynchronous processing and is not suitable for interactive, real-time patient interactions.
• Why D is incorrect: Global Standard remains a pay-as-you-go deployment. Although it can dynamically route traffic across Azure infrastructure, it does not provide the same reserved capacity and predictable throughput characteristics as a provisioned deployment.
Q11: Model Catalog Categories (Sold by Azure vs Partner Models)
For each of the following statements about the Microsoft Foundry model catalog, select Yes if the statement is true. Otherwise, select No.
1. The catalog's models fall into two main categories: Foundry Models sold by Azure and Foundry Models from partners and community.
2. Every model sold by Azure is developed by Microsoft or OpenAI.
3. Models sold by Azure are billed through your Azure subscription, covered by Azure service-level agreements, and supported by Microsoft.
• Statement 1 is Yes: this is how the current docs describe the catalog.
• Statement 2 is No: besides all Azure OpenAI models, the sold-by-Azure list includes selected models from providers such as Cohere, DeepSeek, Meta, Mistral AI, and xAI. The category depends on the model, not on the provider's name.
• Statement 3 is Yes: this is what defines the sold-by-Azure category. Partner and community models are supported by their providers, with varying levels of SLA.
Q12: Grounding Strategy for Daily Changing Content (RAG vs Fine-Tuning)
A retail company's internal support agent must answer employee questions using an internal policy document set that is updated almost every day (new promotions, changing return policies, updated store hours). The team wants the agent's answers to always reflect the latest version of these documents without retraining or redeploying the model.
Which customization approach should the team use to ground the agent's responses?
• Why RAG is correct: RAG retrieves relevant content from an external, continuously updated knowledge source (such as an Azure AI Search index over the policy documents) at query time and injects it into the model's context. Because the underlying documents can be re-indexed independently of the model, the agent always reflects the latest content without any model retraining.
• Why Fine-tuning is incorrect: Fine-tuning bakes knowledge into the model's weights at training time. For documents that change daily, this would require constant retraining and redeployment, which is costly, slow, and impractical — fine-tuning is better suited to teaching a model a style, format, or stable domain skill, not fast-changing factual content.
• Why "Prompt engineering alone, with no external data source" is incorrect: Without a live external source, the model can only rely on whatever was in its training data or what's manually pasted into the prompt, so it cannot stay current with daily policy changes at scale.
Q13: Matching Scenarios to Azure AI Services
Match each scenario to the Azure AI service that best fits it. Each service is used exactly once.
1. A moderation pipeline must automatically detect and block hateful, violent, or sexual content submitted by users of a public chat application before it reaches the generative model.
2. A customer feedback platform needs to extract sentiment (positive/negative/neutral) and named entities (product names, locations) from thousands of free-text reviews.
3. A support chatbot needs to retrieve the most relevant passages from a 10,000-document internal knowledge base to ground its answers.
• 1 → Azure AI Content Safety: This service is purpose-built to detect and filter harmful content categories (hate, violence, sexual, self-harm) in text and images, and is designed to sit in front of or alongside a generative model as a safety guardrail.
• 2 → Azure AI Language: This service provides prebuilt natural language processing features including sentiment analysis and named entity recognition (NER), which directly match the need to extract sentiment and entities from review text — no generative model is required for this task.
• 3 → Azure AI Search: This is Azure's information retrieval and indexing service, providing the vector/hybrid search capability needed to find and rank the most relevant passages from a large document set, which is the core retrieval step of a RAG-grounded chatbot.
Q14: Semantic Search & Ranking Without Generation (Text Embeddings)
A media company is building a feature that lets users search their internal video-transcript archive using natural-language queries (e.g., "find clips where the CEO discusses layoffs") and get back the most semantically relevant transcript snippets, ranked by relevance. The feature does not need to generate new text — only to find and rank existing content.
Which type of model is the correct foundation for this feature?
• Why D is correct: Text-embedding models convert text into dense numeric vectors that capture semantic meaning. By embedding both the search query and the transcript chunks into the same vector space, the system can compute similarity scores and rank snippets by semantic relevance — exactly the retrieval task described, with no text generation needed.
• Why A is incorrect: A generative chat model is optimized to produce new text based on a prompt. While it could be combined with embeddings in a RAG pipeline, on its own it is not the component responsible for semantic ranking/retrieval, and using it alone for search would be slower and more expensive than needed.
• Why B is incorrect: Image segmentation models operate on visual pixel data, not text, and cannot process or rank transcript content.
• Why C is incorrect: Text-to-speech models convert written text into audio; they have no role in semantic text search or ranking.
Q15: Deployment Quotas & Capacity (Global Standard vs PTU)
For each of the following statements about deployment quota in Azure AI Foundry, select Yes if the statement is true. Otherwise, select No.
1. Global Standard deployments generally offer the highest default quota compared to single-region Standard deployments.
2. Choosing Global Standard eliminates the need to load-balance requests across multiple resources to reach higher throughput.
3. Quota consumed by a Provisioned (PTU) deployment is measured in tokens-per-minute (TPM), identical to how Standard deployment quota is measured.
• Statement 1 is Yes: Because Global Standard deployments dynamically route traffic across Microsoft's global infrastructure rather than being pinned to one region's capacity, Microsoft documentation states this deployment type provides the highest default quota among the standard tiers.
• Statement 2 is Yes: The same dynamic global routing means a single Global Standard deployment can absorb higher throughput without the architect needing to manually distribute traffic across multiple resource instances for capacity reasons.
• Statement 3 is No: Provisioned (PTU) deployments reserve dedicated capacity measured in Provisioned Throughput Units, a different unit reflecting reserved compute capacity — not the same measurement as the tokens-per-minute (TPM) quota used for pay-as-you-go Standard/Global Standard deployments.
Q16: Foundry Agent Service Tools (OpenAPI vs Code Interpreter vs MCP)
Match each requirement to the Foundry Agent Service tool that fits best. Each tool is used exactly once.
1. The agent must look up customer orders through an existing REST API that already has an OpenAPI 3.0 specification.
2. The agent must analyze an uploaded sales CSV by running Python code and return a chart.
3. Several agents must share the same set of tools, which are exposed by a server endpoint that another team maintains.
• 1 → OpenAPI tool: it connects an agent to external APIs described by an OpenAPI 3.0 or 3.1 specification.
• 2 → Code Interpreter: a built-in tool that writes and runs Python in a sandbox for data analysis, math, and charts.
• 3 → MCP tool: it connects agents to tools hosted on an MCP server endpoint, and it's best for tools shared across agents or maintained by another team.
Q17: Constrained Edge & Offline Inference (Small Language Models - Phi)
A manufacturer wants to run a lightweight language model directly on a factory-floor edge device with limited compute and no reliable internet connectivity, to classify short maintenance log entries into categories.
Which type of model from the Azure AI Foundry model catalog is the best fit for this constraint?
• Why A is correct: Small Language Models such as the Phi family are designed with a much smaller parameter footprint, giving them lower compute and memory requirements, faster inference, and the ability to run on constrained or offline/edge hardware — matching the limited-compute, unreliable-connectivity scenario described.
• Why B is incorrect: Frontier multimodal models like GPT-4o are large, cloud-hosted models that require substantial compute and a reliable connection to Azure's infrastructure; they are not designed to run on constrained edge devices.
• Why C is incorrect: Embedding models produce vector representations for similarity search; they don't perform text classification/generation on their own and don't solve the edge/offline compute constraint by themselves.
• Why D is incorrect: o3 is a large, cloud-based reasoning model requiring significant backend compute; it is not intended for constrained edge or offline deployment.
Q18: Visual & Handwritten Reasoning (Multimodal Input)
An insurance company wants to build a claims-processing agent that can look at a photo of vehicle damage submitted by a customer, read the customer's handwritten description on the same submission, and produce a single combined damage assessment — without chaining together a separate OCR step and a separate captioning step.
Which capability should the underlying model support?
• Why this is correct: A multimodal model can accept both an image and text in the same request and reason over them jointly, producing one combined output. This removes the need to orchestrate a separate image-analysis call and a separate text-generation call, and lets the model correlate what it sees in the photo with what's written in the description.
• Why "Text-only generation" is incorrect: A text-only model cannot process the image input at all — it would require a separate vision service to first convert the image into text, adding an extra pipeline step the requirement is trying to avoid.
• Why "Speech-to-text transcription only" is incorrect: This capability converts spoken audio into text; it doesn't apply to a photo or handwritten text on a form, and produces no image understanding.
Q19: Model Catalog Comparison & Benchmark Leaderboards
For each of the following statements about the Azure AI Foundry model catalog's comparison features, select Yes if the statement is true. Otherwise, select No.
1. The model catalog lets you compare and evaluate models side-by-side using benchmark metrics before choosing which one to deploy.
2. Benchmark data in the catalog is only available for models in the "Azure OpenAI" collection and never for partner or open-source models.
• Statement 1 is Yes: Azure AI Foundry Models lets you explore, compare, and evaluate models side-by-side, including benchmark and leaderboard metrics, to support an informed model-selection decision before deployment.
• Statement 2 is No: Benchmark and comparison metrics are shown for a broad set of models in the catalog, not limited to the Azure OpenAI collection — models curated by Azure AI and other partner models are also included in catalog comparisons where benchmark data is available.
Q20: Multi-Session Customer State (Azure AI Search vs Agent Memory)
A retail support agent in Foundry Agent Service must (1) answer questions from 800 internal policy documents that are indexed in Azure AI Search, and (2) remember each customer's preferred language and open issues across separate chat sessions. Which design meets both requirements?
• Why C is correct: The Azure AI Search tool grounds an agent in an existing search index. Memory in Foundry Agent Service is built for long-term memory: it extracts key information from conversations, consolidates it, and makes it available across sessions. A scope value keeps each user's memories separate.
• Why A is incorrect: File search grounds answers in uploaded files. It doesn't extract customer preferences from conversations and carry them across sessions.
• Why B is incorrect: Fine-tuning bakes static knowledge into the model, so it can't track per-customer facts that change after every chat.
• Why D is incorrect: Short-term memory covers only the current session's conversation.
Microsoft Foundry Architecture Cheat Sheet: Models & Deployments
Mastering Domain 1 of the AI-103 certification requires knowing exactly which deployment tier and foundation model fits each enterprise workload constraint. Review this core comparison cheat sheet before your exam:
| Capability / Service | Architecture & Billing Mechanism | Primary AI-103 Exam Use Case |
|---|---|---|
| Global Standard | Pay-per-token consumption. Dynamic routing across Microsoft global data centers provides the highest default TPM quota. | Default starting point for unpredictable, bursty interactive user traffic with zero idle cost. |
| Data Zone Standard | Pay-per-token consumption. Constrains prompt processing and responses strictly within defined geo-boundaries. | Enterprise workloads mandated to comply with data residency rules like the Azure EU Data Boundary. |
| Provisioned (PTU) | Reserved compute capacity billed per PTU per hour. Provides dedicated processing with no multi-tenant throttling. | Sustained, high-volume production workloads requiring strict SLAs and predictable sub-second latency. |
| Managed Compute (Preview) | Dedicated GPU capacity managed entirely by Foundry. Eliminates VM sizing, cluster provisioning, and host OS patching. | Hosting custom open-weight models from Hugging Face (e.g., Qwen, Mistral) that lack a serverless API. |
| Foundry Model Router | Dynamic classification layer routing prompts based on query complexity to SLMs or frontier LLMs. | Minimizing per-query token expenditure and latency when high volumes of queries are simple FAQs. |
| Azure AI Search Tool | Information retrieval service providing hybrid search (keyword + dense vectors) and semantic reranking. | Grounding agents via RAG against continuously changing documents stored in Azure Blob Storage. |
| Agent Service Memory | Long-term user state storage that extracts and consolidates preferences across separate sessions. | Personalized agent interactions requiring cross-session persistence isolated by user scope. |
Key Takeaways for AI-103 Domain 1 (Choosing Foundry Models)
• Serverless vs. Provisioned: Use Global Standard for cost-effective pay-as-you-go traffic with maximum quota. Upgrade to Provisioned Throughput Units (PTUs) only when sustained volume justifies reserved capacity and predictable low latency is required.
• Data Residency Constraints: When an exam scenario specifies that prompts cannot leave Europe, select Data Zone Standard to adhere to the Azure EU Data Boundary without paying for reserved compute.
• SLM vs. Frontier LLM: Never select a frontier LLM when a task involves lightweight classification, entity extraction, or edge execution on disconnected devices. Phi-4 is the standard recommendation for low-compute, low-latency, or offline edge scenarios.
• Reasoning Model Parameters: Reasoning models (such as o3 and o4-mini) do not accept
temperatureortop_p. Use thereasoning_effortparameter (low, medium, high) to regulate internal reasoning tokens, latency, and cost.• RAG vs. Fine-Tuning: When organizational policies, product manuals, or prices change frequently, implement Retrieval-Augmented Generation (RAG) using Azure AI Search. Fine-tuning is reserved for learning specialized styles or output formats, not dynamic factual knowledge.
Frequently Asked Questions (Microsoft Foundry & AI-103 FAQ)
Review the most common questions regarding Microsoft Foundry model selection and the AI-103 certification exam:
What is the difference between Serverless API and Managed Compute in Microsoft Foundry?
A Serverless API is a pay-per-token consumption model where Microsoft hosts the model with zero infrastructure management. Managed Compute (preview) provides dedicated GPU infrastructure to deploy open-weight or custom models from Hugging Face without having to provision, size, or patch underlying virtual machines.
When should you choose Provisioned Throughput Units (PTUs) over pay-as-you-go?
You should select Provisioned Throughput Units (PTUs) when an application processes sustained high-volume query traffic and mandates strict SLAs with predictable, sub-second latency. PTUs reserve dedicated processing capacity, completely eliminating multi-tenant noisy-neighbor contention and throttling risks.
Which Foundry deployment type keeps prompts inside the European Union?
Data Zone Standard keeps prompt and response processing strictly within the customer's selected geographical boundary, fully adhering to the Azure EU Data Boundary. Unlike Global Standard, which dynamically routes requests across global regions to maximize quota, Data Zone Standard guarantees regional data compliance on a pay-per-token basis.
What is the difference between Azure AI Search and Foundry Agent Service Memory?
Azure AI Search is an information retrieval engine that indexes external knowledge bases to ground models with factual domain data via RAG. Foundry Agent Service Memory provides cross-session conversational state, extracting and persisting user preferences and ongoing issues across independent chat interactions.
How does the Microsoft Foundry Model Router reduce inference costs?
The Foundry Model Router evaluates prompt complexity dynamically. High-volume, straightforward FAQ inquiries are routed to fast, inexpensive Small Language Models (SLMs like Phi-4), while ambiguous, multi-step reasoning queries are escalated to frontier LLMs like GPT-4o, drastically reducing overall token expenditure.
Next Step in Your AI-103 Certification Journey
Congratulations on completing Part 1! By working through these 20 scenario-based practice questions, you have locked down the core architectural skills tested in Domain 1 of the AI-103: Developing AI Apps and Agents on Azure exam. You now have a solid command over navigating the Microsoft Foundry model catalog, balancing cost and latency between SLMs (like Phi-4) and frontier LLMs, configuring Serverless APIs versus Provisioned Throughput Units (PTUs), and grounding agents using Azure AI Search.
