AI-103 cheat sheet

A multi-service Azure AI services resource exposes one endpoint and one key shared by Vision, Language, Speech, and other services; a single-service resource exposes only its own API.

Setting a resource's local authentication to Disabled rejects account-key calls and forces callers to present a Microsoft Entra ID token.

The built-in role 'Cognitive Services User' grants data-plane inference calls; 'Cognitive Services Contributor' manages resources but is not the minimal role for runtime calls.

A private endpoint places a private IP in a virtual network, but public network access must also be set to Disabled to remove internet exposure.

Diagnostic settings route request logs and metrics to Log Analytics, storage, or event hubs; platform metrics alone never contain prompt or document text.

Standard (pay-as-you-go) deployments bill per token consumed, while provisioned throughput units reserve capacity and bill hourly regardless of traffic.

Quota is enforced per subscription, region, and model as tokens-per-minute and requests-per-minute; exceeding it returns HTTP 429.

Commitment tiers are purchased on a multi-service resource and discount eligible usage for a one-year or three-year term.

A model deployment has a customer-chosen name; applications send requests to the deployment name, so the underlying model can be changed without editing client code.

A chat completions request carries system (developer) messages for behavior, user messages for input, and assistant messages for prior model turns.

temperature ranges from 0 to 2 and controls randomness; top_p performs nucleus sampling, and only one of the two should be changed at a time.

max_tokens limits only generated output; truncated responses end with finish_reason set to 'length'.

Tool calling is a two-step loop: the model returns tool_calls with JSON arguments, the app executes the function, then sends a 'tool' role message holding the result.

Retrieval-augmented generation grounds answers by retrieving passages from a search index and inserting them into the prompt; it does not retrain model weights.

Hybrid search runs BM25 keyword and vector queries in parallel and fuses the ranked lists; the semantic ranker then reorders the top results.

Prompt Shields detects jailbreak attempts in user prompts and hidden instructions inside documents; groundedness detection scores whether an answer is supported by its sources.

Generative answer quality is evaluated with groundedness, relevance, coherence, and fluency scores produced by a judge model or human labels.

Streaming responses use server-sent events, so the client must accumulate incremental deltas to reconstruct the full message.

Image Analysis 4.0 returns the features named in the request, such as caption, denseCaptions, tags, objects, read, and people.

The Read feature returns text as lines and words with bounding polygons; the asynchronous Read API handles large multi-page PDF and TIFF inputs.

Image classification assigns labels to a whole image, while object detection also returns a bounding box, label, and confidence for each instance.

Custom image models are trained on labeled images and must be published to a prediction resource before an application can call them.

The Face service requires limited-access approval; emotion, identity, and other attributes are restricted even when face detection is enabled.

Dense captions return a description plus a bounding box for each detected region, in addition to a single overall caption for the image.

Spatial analysis processes video streams from cameras to count people or detect motion, using a defined zone and a set of operations.

Precision is the share of predicted boxes that are correct, recall is the share of real objects found, and [email protected] uses a 50 percent overlap threshold.

Sentiment analysis returns per-document and per-sentence confidence scores for positive, neutral, and negative, plus mined opinions when opinion mining is enabled.

PII detection returns entity categories and offsets; the redaction policy controls whether text is masked with a character, with the entity type, or left unmasked.

Key phrase extraction returns salient terms, named entity recognition returns typed entities, and entity linking adds a knowledge-base identifier and URL.

Conversational language understanding maps utterances to intents and extracts entities, trained from labeled example utterances per intent.

Question answering builds a knowledge base from FAQ pages or documents and returns an answer with a confidence score and optional follow-up prompts.

Extractive summarization returns existing sentences ranked as most important, whereas abstractive summarization generates new sentences that paraphrase the source.

The Translator service covers more than 100 languages and supports glossaries and custom translator models for domain terminology.

Azure AI Speech provides real-time and batch speech-to-text, text-to-speech, and custom speech models trained on your audio and transcripts.

Document Intelligence prebuilt models include invoice, receipt, identity document, business card, W-2, and health insurance card, each returning a fixed field schema.

The layout model returns text, tables, selection marks, and paragraph structure with bounding polygons but performs no business-field interpretation.

Custom template models need about five labeled samples per field and depend on consistent visual layout; custom neural models tolerate layout variation.

Document Intelligence analysis is asynchronous: submit the request, poll the operation-location URL until status is succeeded, then read analyzeResult.

Composed models bundle several custom models; the service classifies an incoming document and routes it to the best-matching component model.

Azure AI Search index fields can be searchable, filterable, sortable, facetable, and retrievable; only string fields can be searchable, and only filterable fields accept OData filters.

An indexer pulls source data, runs an optional skillset of enrichment skills, and writes enriched output to the index or to a knowledge store.

Integrated vectorization lets the indexer chunk documents and call an embedding skill so vector fields are populated without custom client code.