Microsoft Azure AI Apps and Agents Developer Associate (AI-103) Flashcards¶
51 cards from this guide's fact sheet and notes. Click a question to reveal the answer.
To study with spaced repetition, download flashcards.csv and import it into Anki or Quizlet (comma separated, first row is the header). Back to the study guide.
Exam logistics¶
Which certification does AI-103 earn?
Microsoft Certified: Azure AI Apps and Agents Developer Associate
Which exam did AI-103 replace and when did it retire?
AI-102 (Azure AI Engineer Associate); retired 2026-06-30
What are the five AI-103 domains and weights?
Plan and manage (25-30%); generative AI and agents (30-35%); computer vision (10-15%); text analysis (10-15%); information extraction (10-15%)
How long is the AI-103 exam?
120 minutes
Which programming language does AI-103 expect?
Python
How do you renew an associate certification?
Free online assessment on Microsoft Learn; associate certifications expire annually
Foundry¶
What is the Foundry project endpoint format?
https://<resource>.services.ai.azure.com/api/projects/<project>
Which Python package is the Foundry SDK?
azure-ai-projects 2.x (AIProjectClient)
What do you pass as model in responses.create?
The deployment name
Deployments¶
Which deployment type is Microsoft's default starting point?
Global Standard
Which deployment type keeps inference in the EU?
Data Zone (Standard; Provisioned or Batch) in an EU region
Which deployment type gives predictable throughput and low latency variance?
Provisioned (reserved PTUs)
Which deployment type is 50% cheaper with a 24-hour target?
Global Batch or Data Zone Batch
What is the Developer deployment type for?
Evaluating fine-tuned models only; 24-hour lifetime and no SLA
Operations¶
What HTTP status means you hit TPM or RPM limits?
429 (retry with backoff; honour retry-after)
Security¶
What replaces API keys for keyless auth?
Microsoft Entra ID tokens via managed identity or DefaultAzureCredential
What were Foundry User and Foundry Owner called before?
Azure AI User and Azure AI Owner (role IDs unchanged)
Responsible ai¶
What does a Foundry guardrail consist of?
Controls: a risk; intervention points; and an action (annotate or annotate and block)
Name the four guardrail intervention points
User input; tool call (agents); tool response (agents); output
Content harm blocking level High blocks what?
Low, medium and high severity (most restrictive)
Content harm blocking level Low blocks what?
High severity only (least restrictive)
Agent and its model deployment have different guardrails. Which applies to the agent?
The agent's guardrail; it fully overrides the model's
What are the two kinds of Prompt Shields attack?
User prompt attacks (jailbreaks) and document (indirect) attacks
What does spotlighting do?
Marks documents as lower-trust by base-64 transforming them so the model is less likely to follow embedded instructions (preview)
Evaluation¶
Groundedness vs response completeness?
Groundedness = precision (no content outside the context); completeness = recall (nothing missing versus ground truth)
What scale do Foundry RAG evaluators use and what is the default pass threshold?
1-5; threshold 3
Agents¶
What three things define a prompt agent?
A model; instructions; tools
What keeps multi-turn state for an agent client?
A conversation (openai.conversations.create) passed to each responses.create
What must your app send back after a function_call?
A function_call_output with the same call_id
How long before a function-calling run expires?
10 minutes
Memory vs knowledge for agents?
Memory = user context learned across sessions; knowledge (Foundry IQ / AI Search) = curated organisational content
Name the three Foundry memory types
User profile memory; chat summary memory; procedural memory
Which Agent Framework builder chains agents in order?
SequentialBuilder (agent_framework.orchestrations)
How do you require human approval for a tool in Agent Framework?
@tool(approval_mode="always_require")
Observability¶
Which env var enables Foundry GenAI client tracing in Python?
AZURE_EXPERIMENTAL_ENABLE_GENAI_TRACING=true (before instrumenting)
Are prompts recorded in traces by default?
No; set OTEL_INSTRUMENTATION_GENAI_CAPTURE_MESSAGE_CONTENT=true (dev only)
Vision¶
Which image model was retired on 2026-03-04?
DALL-E 3 (use GPT-image models)
In an image edit mask what do transparent pixels mean?
The area to edit
What does input_fidelity control?
How closely edits preserve the input image's style and features especially faces
What pattern does Sora 2 video generation use?
Asynchronous job: create; poll status; download
How do you edit a generated Sora 2 video?
Remix it with videos.remix and a prompt
Extraction¶
What replaced Content Understanding pro mode?
Agentic mode (preview; 2026-06-01-preview API)
Name the three Content Understanding field methods
Extract; classify (enum); generate
What does begin_analyze return?
A poller; call .result() for the AnalysisResult
Retrieval¶
Which Azure AI Search query type is the agent tool's default?
vector_semantic_hybrid
What must match between indexing and the vectorizer?
The embedding model
What does integrated vectorization add to a skillset?
Chunking (Text Split or Content Understanding skill) and an embedding skill; plus a query-time vectorizer
What are index projections for?
Writing chunks to a secondary index alongside parent documents in one indexer run
Text¶
Which Azure Language call redacts PII?
recognize_pii_entities (returns redacted_text; categories; offsets; confidence)
Speech¶
What handles a single utterance in the Speech SDK?
recognize_once_async (about 30 seconds or until silence)
What is the Voice Live API?
Managed speech-to-speech for voice agents: STT + generative model + TTS with noise suppression and turn detection