AI-103 flashcards

32 cards. Read the term and the definition here. The deck at the top of this page is optional.

Multi-service AI resource

A single Azure resource exposing one endpoint and key for several Azure AI services such as Vision, Language, and Speech, which simplifies billing and credential management.

Disable local auth

A resource setting that rejects account-key authentication and forces every caller to present a Microsoft Entra ID token, enabling managed identities and removing long-lived secrets.

Cognitive Services User role

A built-in Azure role that permits data-plane inference calls against an Azure AI services resource, without granting permission to create resources or regenerate account keys.

Provisioned throughput unit

Reserved model capacity that is billed hourly whether or not requests arrive, delivering predictable latency for steady heavy workloads compared with per-token standard deployments.

HTTP 429 throttling

The response returned when a deployment exceeds its tokens-per-minute or requests-per-minute quota; clients should retry with exponential backoff or request a quota increase.

Private endpoint

A private IP address inside a virtual network that connects to an Azure AI resource; paired with public network access disabled, it removes all internet-facing exposure.

System message

The chat message that sets a model's persona, rules, and output format before user turns, carrying the highest-priority instructions in a chat completions request.

Temperature parameter

A sampling setting from 0 to 2 where lower values make output more deterministic and higher values increase variety; adjust it or top_p, not both at once.

Tool calling loop

The model returns tool_calls containing function names and JSON arguments; the application executes the function and sends a tool-role message with the result so the model can continue.

Retrieval-augmented generation

A grounding technique that retrieves relevant passages from an index and inserts them into the prompt so the model answers from supplied content without retraining weights.

Hybrid search

An Azure AI Search mode that runs keyword and vector queries in parallel and fuses the ranked results, optionally followed by the semantic ranker for better relevance.

Prompt Shields

An Azure AI Content Safety capability that detects jailbreak attempts in user prompts and hidden instructions embedded in documents before the model processes the request.

Agent thread

A conversation state object in Foundry Agent Service that stores the messages exchanged with an agent; runs execute the agent against the messages in that thread.

Dense captions

An Image Analysis feature returning a short description plus a bounding box for each notable region of an image, in addition to one overall caption for the picture.

Read OCR

An Image Analysis feature that extracts printed and handwritten text as lines and words with bounding polygons, synchronously for images and asynchronously for large documents.

Object detection

A custom vision task that locates each instance of a trained class by returning a bounding box, label, and confidence, unlike classification which labels the entire image.

Custom image model

A model trained on your labeled images for classification or object detection; it must be published to a prediction resource before an application can invoke it.

Face liveness detection

A Face service capability that verifies a real person is present before a camera, used under limited-access approval to prevent photo or video spoofing attacks.

mAP metric

Mean average precision, averaged across classes and overlap thresholds, summarizing how well an object detection model places boxes and assigns correct labels.

Opinion mining

A sentiment analysis option in Azure AI Language that returns aspects of a sentence, such as a product attribute, with the sentiment expressed toward each aspect.

PII redaction policy

A setting controlling how detected personal data is replaced in returned text, using character masking, entity-type masking, or no masking, while offsets are still reported.

Entity linking

A text analysis feature returning entities disambiguated against a knowledge base with a Wikipedia identifier and URL, unlike named entity recognition which returns only type categories.

Conversational language understanding

An Azure AI Language feature that trains a model to map user utterances to intents and to extract entities, using labeled example utterances supplied for each intent.

Extractive summarization

A summarization mode that selects and returns the most important existing sentences from a document, preserving original wording instead of generating new paraphrased text.

Custom speech model

A speech-to-text model trained with your own audio and transcripts to improve recognition of domain vocabulary, acronyms, and noisy acoustic conditions.

Custom template model

A Document Intelligence model trained on roughly five labeled examples per field that relies on consistent visual layout and suits standardized forms.

Custom neural model

A Document Intelligence extraction model that handles varied layouts and less structured documents, trained from labeled samples and generally more tolerant of format changes than template models.

Analyze operation polling

Document Intelligence returns a 202 with an operation-location header; the client polls that URL until status reports succeeded, then reads the analyzeResult fields.

Composed model

A Document Intelligence model bundling several custom models, where the service classifies an incoming document and routes it to the best-matching component model.

Layout model

A Document Intelligence prebuilt model returning text, tables, selection marks, and paragraph structure with bounding polygons, without interpreting business-specific fields.

Skillset

A set of enrichment steps in Azure AI Search, such as OCR, entity recognition, or embedding, that an indexer runs over source documents before writing to the index.

Integrated vectorization

An Azure AI Search indexing approach where the indexer chunks documents and calls an embedding skill, so vector fields are populated without custom client code.