AI-103 Concept Map (Mind Map): Skill area 1

Every topic from skill area 1 of the exam on one page, with the pairs the exam swaps to trick you. Skill area 1 counts for 25-30% of your score.

1.1 Choose the right Foundry services

Models, retrieval, memory and tools for each job

Pick the model

Large language model (LLM): A large, general-purpose model for complex reasoning, long content and open-ended tasks.
Small language model (SLM): A compact model, such as Microsoft’s Phi family, for focused tasks with low latency and cost, even on edge devices.
Multimodal model: Takes in and reasons over several types of input together, such as text with images or audio.
Embedding model: Turns text or images into vectors so they can be searched by meaning.
Foundry Tools: Prebuilt, task-specific AI services: Vision, Language, Speech, Translator, Document Intelligence, Content Understanding. Formerly Azure AI services, and Cognitive Services before that.
Model router: A trained model you deploy once; it analyzes each prompt and sends it to the most suitable underlying model. Routing modes: Balanced, Cost or Quality.

Pick the service for the job

Foundry Models: The catalog for deploying and calling models from Microsoft, OpenAI and other providers, including Azure OpenAI.
Foundry Agent Service: Managed service to build, host and run agents that use models, tools, knowledge and conversation state.
Azure AI Search: Search engine for indexing your content with keyword, vector, hybrid and semantic search; the backbone of most RAG on Azure.
Foundry IQ: A knowledge layer, built on Azure AI Search, that grounds agents in enterprise data from several sources while respecting permissions.
Content Understanding: A Foundry Tool that turns documents, images, audio and video into structured output, for multimodal processing.

Retrieval and indexing methods

Keyword search: Full-text search that matches exact terms, ranked with BM25.
Vector search: Compares embeddings to find content with similar meaning, even with different words.
Hybrid search: Runs keyword and vector search together and merges the results (Reciprocal Rank Fusion).
Semantic ranker: Re-ranks the top results with a language-understanding model, and can return captions and answers.
Integrated vectorization: Azure AI Search chunks and embeds content automatically while indexing, and vectorizes queries too.
Indexer: A crawler that pulls data from a source such as Blob Storage into an index on a schedule.
Agentic retrieval: Uses a model to break a complex question into subqueries, runs them in parallel and merges the results.

Memory, tools and knowledge for agents

Conversation state: The running history of a conversation, kept by Agent Service so the agent follows the current exchange.
Long-term memory: Memory in Foundry Agent Service (preview): a managed store that extracts, keeps and recalls facts and preferences across separate conversations.
File search tool: Lets an agent search files you upload; Agent Service chunks and indexes them for you.
Azure AI Search tool: Connects an agent to an existing Azure AI Search index.
Grounding with Bing Search: Lets an agent ground answers in public, up-to-date web results.
Function, OpenAPI and MCP tools: Let agents call your own code, REST APIs described with OpenAPI, or tools exposed through MCP servers.
Code interpreter: Lets an agent write and run Python in a sandbox, for calculations, data analysis and charts.

1.2 Set up AI solutions in Foundry

Resources, deployment types and pipelines

Foundry infrastructure

Foundry resource: The top-level Azure resource that holds model deployments, security, networking and billing.
Foundry project: A workspace inside the Foundry resource where you build agents, run evaluations and organize files.
Hub-based project: The older project type, hosted by a hub. Still needed for features like prompt flow, managed compute deployments and Azure Machine Learning capabilities.
Connections: Saved, secured links from a project to other resources like Storage, Azure AI Search or Key Vault.
Basic vs standard agent setup: Basic setup lets Agent Service manage storage for files, conversations and vector stores. Standard setup uses your own Azure Storage, Azure AI Search and Azure Cosmos DB, so all agent data stays in your tenant.

Deployment options

Standard types: Pay per token with no commitment: Global Standard, Data Zone Standard, or Standard in a single region. Best-effort capacity.
Provisioned (PTU): Reserve throughput in provisioned throughput units, at global, data zone or single-region level, for predictable latency.
Global deployment: Requests can be processed in any Azure region worldwide, giving the highest capacity. Global Standard is the usual starting point.
Data Zone deployment: Prompts and responses are processed only within a Microsoft-defined data zone, such as the US or the EU, with higher default quotas than single-region.
Single-region deployment: Requests are processed in the deployment region only. The pay-per-token version of this type is simply called “Standard”.
Batch deployment: Global Batch or Data Zone Batch: asynchronous jobs with separate quota, a 24-hour target turnaround, at 50% less cost than Global Standard.
Serverless API: The deployment option where Microsoft hosts the model. The Standard, Provisioned and Batch types all apply to it.
Managed compute: Deploy open-source or custom models onto dedicated virtual machines, billed for compute time. Uses hub-based projects (Foundry classic).

Configuring deployments and pipelines

Deployment settings: Choose the model version, upgrade policy, quota (tokens per minute) and the guardrail/content filter applied.
CI/CD integration: Automate deployments of models, agents and prompts with GitHub Actions or Azure DevOps, with infrastructure as code (Bicep or Terraform).
Evaluation as a quality gate: Run automated evaluations in the pipeline and block a release if quality or safety scores drop.

1.3 Manage, monitor and secure

Capacity, cost, health and access

Quotas, scaling and cost

Tokens per minute (TPM): The token quota assigned to a deployment, out of your subscription’s regional quota.
Requests per minute (RPM): The number of calls a deployment accepts per minute, set in proportion to its TPM.
429 errors and retries: “Too Many Requests” when you exceed limits; handle with retries using exponential backoff.
Cost footprint: Track token usage per deployment and agent, set budgets and alerts in Azure Cost Management.

Monitoring

Azure Monitor and Application Insights: Collect metrics, logs and traces for deployments and agents: latency, errors, token use.
Performance and drift: Watch quality over time as data, users or model versions change.
Safety events: Track how often content filters and prompt shields trigger.
Grounding quality: Measure whether answers are supported by the retrieved sources, with groundedness evaluations.
Index health and relevance: Check indexer runs and errors, ingestion quality, index size, query latency and whether top results are relevant.

Security

Managed identity: An identity Azure manages for your app, so it can access other resources without stored secrets.
User-assigned identity: A standalone managed identity you create once and attach to several resources.
Keyless credentials: Authenticate with Microsoft Entra ID tokens instead of API keys.
Private networking: Private endpoints and disabled public access keep traffic on Azure’s private network.
Role policies (RBAC): Assign least-privilege roles, such as Foundry User (formerly Azure AI User) to build and use agents and models in a project.
Azure Key Vault: Stores secrets, keys and certificates when you can’t avoid them.

1.4 Responsible AI in practice

Filters, evaluations, audits and agent controls

Safety filters and guardrails

Content filters: Classify prompts and responses for hate, sexual, violence and self-harm, at severity levels safe, low, medium and high.
Blocklists: Custom lists of terms that are always blocked.
Prompt Shields: user attacks: Detect jailbreak attempts typed by the user to bypass the system’s rules.
Prompt Shields: document attacks: Detect indirect prompt injection hidden in documents, emails or web pages the model processes.
Protected material detection: Flags outputs that reproduce known copyrighted text or public code.
Groundedness detection: Checks whether a response is supported by the provided source material.

Responsible AI instrumentation

Quality evaluators: Scores like groundedness, relevance, coherence and fluency, run on test datasets or live traffic.
Groundedness evaluator: Scores whether the answer is supported by the retrieved context.
Relevance evaluator: Scores whether the answer addresses the user’s question.
Safety evaluations: Measure how often a model or agent produces harmful content or falls for attacks.
AI red teaming: Deliberately attacking your own system, manually or with the AI Red Teaming Agent, to find weaknesses first.

Auditing

Trace logging: Record each step of a model or agent run, such as prompts, tool calls and latency, with OpenTelemetry into Application Insights.
Provenance metadata: Records where content came from, such as Content Credentials attached to AI-generated images.
Approval workflows: Require a person to sign off before sensitive actions, deployments or releases go ahead.

Governing agent behavior

Oversight modes: Choose how autonomous an agent is: fully autonomous, or pausing for human approval before certain actions, such as MCP tool calls.
Constraints: Rules that bound an agent: instructions, allowed scope, limits on steps or spending.
Tool-access controls: Give an agent only the tools and permissions it needs, with identities scoped by RBAC.

Part of the Azure AI-103: AI Apps and Agents Developer Study Kit: Concept Maps, Flashcards and practice questions. Also free: the Skill area 1 flashcards. New to the exam? Read the AI-103 study guide.