AI-103 - Microsoft Certified Azure AI Apps and Agents Developer Associate Cheat Sheet

Updated 2026-08-06

Next Topic: AI-901 - Microsoft Azure AI Fundamentals Cheat Sheet

🎯Take a practice test on this topic9 practice tests · 329 questions→

AI-103 (Microsoft Certified: Azure AI Apps and Agents Developer Associate) is Microsoft's Foundry-era successor to the retiring AI-102 Azure AI Engineer exam, validating your ability to build, deploy, and govern generative AI apps and autonomous agents on Microsoft Foundry using Python. The exam spans five domains: planning and managing a Foundry-based AI solution, building generative AI and agentic solutions (the largest single domain at 30-35%), and implementing computer vision, text analysis, and information extraction workloads. Microsoft Foundry is the unifying platform across every domain - the same project, connection, deployment, agent, thread, and tool model shows up whether you are picking a model, wiring an agent, grounding a RAG pipeline, or wiring up an evaluator, so a candidate who masters the Foundry object model can transfer that mental model across nearly every exam area. Expect heavy emphasis on choosing the right service for a job (which model tier, which retrieval method, which vision or speech API) and on the responsible-AI guardrails Microsoft now bakes into agentic workflows by default.

What This Cheat Sheet Covers

This topic spans 29 focused tables and 266 indexed concepts, 9 practice tests with 329 questions. Below is a complete table-by-table outline of this topic, spanning foundational concepts through advanced details.

Table 1: Choosing Foundry Models, Services, and Integration OptionsTable 2: Setting Up AI Solution Infrastructure and Deployments in FoundryTable 3: Managing Quotas, Scaling, Cost, and Monitoring for AI WorkloadsTable 4: Securing AI Systems: Identity, Networking, and Access ControlTable 5: Responsible AI: Safety, Auditing, and Agent GovernanceTable 6: Deploying and Consuming Models in FoundryTable 7: Connecting Applications to Foundry: SDKs, Connectors, and ProjectsTable 8: Implementing RAG and Tool-Augmented Reasoning WorkflowsTable 9: Evaluating Generative Models and ApplicationsTable 10: Defining and Building Core Agents: Roles, Retrieval, and MemoryTable 11: Integrating Agent Tools and Knowledge SourcesTable 12: Orchestrating Multi-Agent SolutionsTable 13: Building Autonomous Agent Workflows with SafeguardsTable 14: Monitoring and Evaluating Deployed AgentsTable 15: Tuning Generation Behavior and Reflection TechniquesTable 16: Observability and Hybrid Model OrchestrationTable 17: Generating and Editing Images with FoundryTable 18: Generating and Editing Videos with FoundryTable 19: Multimodal Visual Analysis, Captioning, and Accessibility DescriptionsTable 20: Azure Content Understanding for Visual and Video DataTable 21: Object, Component, and Region Detection in Images and VideoTable 22: Responsible AI for Multimodal and Visual ContentTable 23: Extracting and Structuring Text with Language ModelsTable 24: Sentiment, Safety Detection, and Text TranslationTable 25: Implementing Speech Solutions for Agentic InteractionsTable 26: Ingesting and Indexing Content for SearchTable 27: Semantic, Hybrid, and Vector Search ConfigurationTable 28: Enrichment Pipelines, RAG Ingestion, and Agent-Connected RetrievalTable 29: Extracting and Structuring Document Content with Content Understanding

Quick IndexSubscribe to unlock

A jump-to index of every table row in this cheat sheet.

Mind MapSubscribe to unlock

An interactive map of every table and concept in this topic.

Table 1: Choosing Foundry Models, Services, and Integration Options

Domain 1 opens with picking the right building block for a generative AI or agent scenario: which model family, which Foundry service, which retrieval method, and which memory, tool, or knowledge integration keeps the solution cost-effective and well-grounded.

Concept	Example	Description
Foundry Models Sold by Azure vs. Partner and Community	`Need a guaranteed Microsoft SLA + deep Azure integration → pick a model sold by Azure; need a cutting-edge niche model → pick a partner/community model`	Both live in one Foundry model catalog. •Models sold by Azure carry Microsoft support and enterprise SLAs •Partner/community models trade that for specialized, innovation-led capabilities
Small Language Models (SLMs)	`Classify tickets into 5 fixed categories on a tight budget → Phi-3-mini (3.8B params), not a frontier LLM`	Cheaper and faster to run, best for narrow, well-defined, low-recall tasks. Not to be confused with a full LLM, which suits broad, exploratory, high-recall tasks with more compute budget.
Multimodal Models	`Read a photo's price tag and speak the answer back in one call → a GPT-4o class model with vision + audio`	One model spans more than text, such as image, audio, or video, in a single request. Distinct from chaining a text-only LLM with separate vision or speech services.
Foundry Tools (e.g. Content Understanding)	`Pull vendor, date, and total from scanned receipts into a defined schema → Content Understanding analyzer, no custom prompt engineering`	A ready-made platform capability you call instead of prompting a model yourself. Not to be confused with a custom agent tool function, which is code you author and operate.
Prompt Agents vs. Hosted Agents	`Ship a portal-built FAQ agent this week, no infra to run → Prompt agent; ship your own LangGraph orchestration in a container → Hosted agent`	Both are agent workflow types in Foundry Agent Service. •Prompt agents: config-only, fully managed, zero runtime code •Hosted agents: your own code, Foundry-managed container, full orchestration control
Query Types: Full Text, Vector, and Hybrid Search	`Look up exact order number "ORD-88214" → full text search; a "find similar meaning" query → vector search; best relevance → hybrid`	•Full text: precise on exact codes, jargon, names •Vector: conceptual/semantic similarity across languages and content types •Hybrid: runs both in parallel and merges results with Reciprocal Rank Fusion
Foundry Agent Service Memory	`Remember a shopper's stated shoe size across visits and devices → memory store scoped per user`	Managed, long-term memory for continuity and personalization across sessions and devices. Not the same as a single conversation's chat history; the scope parameter isolates memories per user.
Foundry IQ Knowledge Base	`Agent asks a natural-language HR question → Foundry IQ plans and runs keyword, vector, and hybrid subqueries, then reranks and synthesizes one answer`	A managed knowledge layer over Azure AI Search connected to an agent via MCP. Handles query planning and reranking automatically, instead of you wiring index queries by hand.

Table 2: Setting Up AI Solution Infrastructure and Deployments in Foundry

Domain 1, Task: Set up AI solutions in Foundry, covering the resource/project layout, the deployment options (standard, managed compute, provisioned throughput, batch), and CI/CD for Foundry agent projects.

Concept	Example	Description
Foundry Resource	`Microsoft.CognitiveServices/accounts` (kind `AIServices`)	Top level resource for governance: networking, security, model deployments. • Owns one or more projects • Not to be confused with a Foundry project

The rest of this section is part of the full cheat sheet

Table 3: Managing Quotas, Scaling, Cost, and Monitoring for AI Workloads

Domain 1, Manage, monitor, and secure AI systems: how Foundry meters quota, rate limits, and cost across model and agent workloads, and the Azure Monitor, Foundry Observability, and Azure AI Search surfaces that watch performance, drift, safety, and index health once a system is live.

Concept	Example	Description
Foundry Model Quota (TPM/RPM)	`gpt-4o-mini` • GlobalStandard RPM: 20,000 • TPM: 2,000,000	Quota is assigned per subscription, per region, per model, and per deployment type, in Tokens Per Minute (TPM), never as one tenant-wide pool. • Not to be confused with a dollar cost budget: quota governs throughput, not spend.

The rest of this section is part of the full cheat sheet

Table 4: Securing AI Systems: Identity, Networking, and Access Control

Domain 1, "Manage, monitor, and secure AI systems": configuring authentication, managed identity, private networking, and RBAC role policies for a Microsoft Foundry resource.

Concept	Example	Description
Microsoft Entra ID vs API Keys	`headers = {"Authorization": f"Bearer {token.token}"}` (Entra ID) `headers = {"api-key": api_key}` (API key)	Foundry supports both, but Microsoft recommends Entra ID for production: conditional access, per-principal auditing, managed identities. • API keys are for rapid prototyping only, not accepted by enterprise production workloads and cannot be scoped per user.

The rest of this section is part of the full cheat sheet

Table 1: Choosing Foundry Models, Services, and Integration Options

Concept	Example	Description
Foundry Models Sold by Azure vs. Partner and Community	`Need a guaranteed Microsoft SLA + deep Azure integration → pick a model sold by Azure; need a cutting-edge niche model → pick a partner/community model`	Both live in one Foundry model catalog. •Models sold by Azure carry Microsoft support and enterprise SLAs •Partner/community models trade that for specialized, innovation-led capabilities
Small Language Models (SLMs)	`Classify tickets into 5 fixed categories on a tight budget → Phi-3-mini (3.8B params), not a frontier LLM`	Cheaper and faster to run, best for narrow, well-defined, low-recall tasks. Not to be confused with a full LLM, which suits broad, exploratory, high-recall tasks with more compute budget.
Multimodal Models	`Read a photo's price tag and speak the answer back in one call → a GPT-4o class model with vision + audio`	One model spans more than text, such as image, audio, or video, in a single request. Distinct from chaining a text-only LLM with separate vision or speech services.
Foundry Tools (e.g. Content Understanding)	`Pull vendor, date, and total from scanned receipts into a defined schema → Content Understanding analyzer, no custom prompt engineering`	A ready-made platform capability you call instead of prompting a model yourself. Not to be confused with a custom agent tool function, which is code you author and operate.
Prompt Agents vs. Hosted Agents	`Ship a portal-built FAQ agent this week, no infra to run → Prompt agent; ship your own LangGraph orchestration in a container → Hosted agent`	Both are agent workflow types in Foundry Agent Service. •Prompt agents: config-only, fully managed, zero runtime code •Hosted agents: your own code, Foundry-managed container, full orchestration control
Query Types: Full Text, Vector, and Hybrid Search	`Look up exact order number "ORD-88214" → full text search; a "find similar meaning" query → vector search; best relevance → hybrid`	•Full text: precise on exact codes, jargon, names •Vector: conceptual/semantic similarity across languages and content types •Hybrid: runs both in parallel and merges results with Reciprocal Rank Fusion
Foundry Agent Service Memory	`Remember a shopper's stated shoe size across visits and devices → memory store scoped per user`	Managed, long-term memory for continuity and personalization across sessions and devices. Not the same as a single conversation's chat history; the scope parameter isolates memories per user.
Foundry IQ Knowledge Base	`Agent asks a natural-language HR question → Foundry IQ plans and runs keyword, vector, and hybrid subqueries, then reranks and synthesizes one answer`	A managed knowledge layer over Azure AI Search connected to an agent via MCP. Handles query planning and reranking automatically, instead of you wiring index queries by hand.

Table 2: Setting Up AI Solution Infrastructure and Deployments in Foundry

Concept	Example	Description
Foundry Resource	`Microsoft.CognitiveServices/accounts` (kind `AIServices`)	Top level resource for governance: networking, security, model deployments. • Owns one or more projects • Not to be confused with a Foundry project

Table 3: Managing Quotas, Scaling, Cost, and Monitoring for AI Workloads

Concept	Example	Description
Foundry Model Quota (TPM/RPM)	`gpt-4o-mini` • GlobalStandard RPM: 20,000 • TPM: 2,000,000	Quota is assigned per subscription, per region, per model, and per deployment type, in Tokens Per Minute (TPM), never as one tenant-wide pool. • Not to be confused with a dollar cost budget: quota governs throughput, not spend.

Table 4: Securing AI Systems: Identity, Networking, and Access Control

Domain 1, "Manage, monitor, and secure AI systems": configuring authentication, managed identity, private networking, and RBAC role policies for a Microsoft Foundry resource.

Concept	Example	Description
Microsoft Entra ID vs API Keys	`headers = {"Authorization": f"Bearer {token.token}"}` (Entra ID) `headers = {"api-key": api_key}` (API key)	Foundry supports both, but Microsoft recommends Entra ID for production: conditional access, per-principal auditing, managed identities. • API keys are for rapid prototyping only, not accepted by enterprise production workloads and cannot be scoped per user.