AI-103 Cheat Sheet: Last-Minute Revision for Azure AI Apps and Agents

Everything on the AI-103 exam on one page. Read it the day before, then take the AI-103 practice test.

Weighting

Plan and manage

Model choice: SLM for cheap, narrow, high volume · LLM for general chat · reasoning model for hard multi-step problems · multimodal for images and audio · embeddings for vectors · Foundry Tools for predictable prebuilt tasks.

Deployment types:

Setup: resource (deployments, networking) → project (agents, evals, traces, connections) · stable deployment names · versioned agents · Bicep/Terraform + azd + CI/CD with evaluation gates · OIDC for pipelines.

Operate: 429 = rate limit → backoff, fewer tokens, more quota or deployments, an API Management gateway, Provisioned/Batch · monitor latency, tokens, errors, safety events, quality drift, indexer and index health.

Secure: managed identity + Azure AI User role + DefaultAzureCredential · disable key auth · private endpoints + private DNS + public access off · Key Vault for remaining secrets · least privilege, one identity per agent.

Responsible AI: guardrails = harm categories with thresholds, Prompt Shields (jailbreak + indirect attacks), protected material, groundedness detection, blocklists · safety evaluators and red teaming · traces + provenance + logged approvals · agent oversight modes, constraints, tool approvals.

Generative AI and agents

project = AIProjectClient(endpoint=endpoint, credential=DefaultAzureCredential())
openai = project.get_openai_client()
r = openai.responses.create(model="<deployment>", instructions="...", input="...")
# multi-turn: previous_response_id=... or conversation=conv.id
# stream=True · text={"format": {"type": "json_schema", ...}} · {"type": "input_image", ...}

RAG: chunk → embed → index · retrieve (hybrid + semantic ranker) → ground (only sources, cite, say "I don't know") → generate · security trimming with filters · agents use AI Search tool, file search or Foundry IQ (agentic retrieval).

Evaluators:

AI-assisted evaluators need a judge model deployment.

Agents: model + instructions + tools + conversation (+ memory) · tools: file search, AI Search, code interpreter, web, function (your code runs it: function_call → function_call_output with call_id), OpenAPI, MCP · agents.create_version(...) creates versions.

Multi-agent patterns:

Safeguards: approvals, max turns, budgets, scoped identities, validation, escalation, checkpoints.

Optimise: lower temperature for consistency (reasoning models use reasoning effort) · reflection = generate → critique → revise with a round limit · tracing = OpenTelemetry → Application Insights (spans, token analytics, latency, safety signals) · routing, cascades, LLM + rules engines.

Vision

• Generate: images.generate(size, quality, n) · edit with reference images · inpainting with a mask (transparent = change) · video = job → poll → download, remix, image-to-video
• Understand: input_image parts · concise vs detailed captions · grounded visual Q&A ("can't tell from the image") · alt text short and purposeful; extended descriptions for charts
• Content Understanding: visual fields, video segments and key frames, pro mode for multi-file reasoning
• Safety: Content Safety image categories · injected text in images = indirect prompt injection · watermarks, C2PA, brand and symbol policies

Text and speech

• LLM + structured outputs for entities, topics, summaries, tone · Azure Language for sentiment, entities, PII redaction, custom models
• Translator for volume, glossaries, documents · LLM translation for tone and adaptation
• Speech: STT (real-time, fast, batch, diarization) · TTS with SSML · phrase lists / custom speech (measured by WER) · Voice Live / realtime audio for low-latency barge-in · speech translation

Information extraction

• AI Search: data source → indexer → skillset → index, vectorizer for queries · OCR / Document Layout for scans · Text Split + embedding skill = integrated vectorization · Custom Web API skill for your logic
• Content Understanding: analyzer + field schema (extract / classify / generate) · Markdown output for RAG and agents · confidence + grounding → human review · begin_analyze → poll (REST: 202 + Operation-Location)

Commonly confused

Exam-day tips

1. Read the requirement and constraint first: cheapest, least effort, no keys, data residency, real time.
2. Eliminate answers that are less secure (keys in code, public access) or more manual (portal clicks for every environment) when automation is asked for.
3. For code questions, look for the missing step: credential, deployment name, call_id, polling, schema.
4. Flag long case studies and come back to them.

Good luck! When you pass, share the news, and check our AI courses roadmap for what's next.