Azure AI Apps and Agents Flashcards & Memory Notes

Exam AI-103. It earns Azure AI Apps and Agents Developer Associate. AI-102 retired on June 30, 2026.

700 or greater on a scale of 1 to 1,000. That scaled score is not a raw percentage.

120 minutes. Microsoft does not publish a fixed question count. Confirm the timer when you schedule.

Generative AI and agentic solutions, at 30–35%. Planning is next, at 25–30%.

Evaluate it on representative labels, ambiguous cases, latency, and total cost. Use an escalation route when the smaller model cannot meet the acceptance criteria.

When one request includes more than text, such as a photograph plus an instruction.

No. It authenticates the workload and still needs appropriate roles. Private connectivity, public-access settings, DNS, and routing are separate controls.

They do not grant document access or authorize tool actions. Test configured content risks and enforce permissions, validation, and approval independently.

Retrieve relevant passages and condition the answer on them. A larger model does not search a private library by itself.

Use supported schema-constrained outputs, handle refusal or incomplete results, and validate field values. JSON mode and lower temperature do not guarantee the required schema or correct facts.

Guide the agent with durable instructions, but calculate and validate the amount in trusted application code before execution. A prompt is not an authorization boundary.

No. The cited source must exist and support the actual claim. A real page about a different product does not establish groundedness.

Whether the answer addresses the request. A grammatical reply about the wrong product fails relevance.

The app calls a tool, then continues from that result. Guessing an order status skips the lookup.

The rules engine owns an exact check, such as a payout threshold. The model explains the result.

Latency by step and token use. That separates retrieval time from model time.

Role, goal, conversation tracking, and tool schemas. A one-line completion is not that definition.

No. Validate and authorize the arguments, execute the permitted tool, and inspect its actual result. A requested call or timeout is not a verified success.

By trusted user or tenant identity and the appropriate conversation, with retention and deletion rules. Remembered account identifiers do not grant access.

Clear roles, handoff contracts, decision ownership, and step or token limits. Add a termination or escalation path for loops and unresolved requests.

The exact action and arguments, such as recipient and amount. Enforce approval in the execution service and require a new decision if approved details change.

Report the outcome as uncertain, reconcile operation status, and use supported idempotency controls before retrying. Do not duplicate the side effect or invent success.

No. Check explicit source support, use independent evaluation examples, and bound the number of critique passes. Two model passes can repeat the same error.

Inspect observable tool choices, arguments, results, citations, and explicit workflow artifacts. The exam names chain-of-thought evaluation, but private internal reasoning is not necessarily exposed.

Generation creates new visuals. Understanding reads images or video that already exist.

Not for every model or workflow. The mask targets an edit; verify the surrounding image and use controlled compositing when exact preservation is required.

Untrusted text inside the image that tries to override instructions. Detect it and do not obey it.

Recognize the term in the April 16 exam guide, but its old preview API is retired. Consult current agentic-mode migration guidance and limitations before building a lab.

An accessibility description of the image. A marketing slogan is not a substitute when the stem requires accessibility guidance.

Speech to text listens. Text to speech speaks.

For a small, changing vocabulary on a supported recognition path, evaluate a phrase list first. Broader domain adaptation may justify Custom Speech with suitable data.

When the task is dedicated translation. Do not answer that stem with image captioning or a video edit.

Searchable representations, useful chunks, embeddings where needed, source references, versions, and permissions. Other grounding designs can use authorized API results or supplied documents.

When scans have no text layer. Indexing the file name does not read the page.

Keyword and vector retrieval together. Use it when an exact SKU and a paraphrase both matter.

Vector retrieval finds candidates using embedding similarity. Semantic ranker reorders retrieved text; reranking alone cannot recover a passage absent from the candidate set.

Tables and reading order, not only a flat bag of words.

When built-in skills miss a form and you already have an extractor for it.

Before retrieved content enters the model context, using supported document authorization or trusted identity filters. Do not rely on the model to hide forbidden passages.

Source grounding, available confidence, and validation of required values and business rules. Missing or conflicting consequential fields need review rather than invented values.