AI-103 key terms

A workspace inside Azure AI Foundry that groups model deployments, data connections, evaluations, and agents so a team can build and monitor AI applications together while inheriting security and billing from the parent resource.

Publishing a specific model version to a named endpoint with its own capacity, content settings, and configuration so client applications can send inference requests to it.

A shared secret string that lets a client call an AI service endpoint without signing in with an identity; it is simple but must be rotated and kept out of source code.

An automatically managed Microsoft Entra identity assigned to an Azure resource, allowing that resource to authenticate to other services without storing credentials in configuration or code.

Assigning built-in or custom roles to users, groups, or managed identities so each principal receives only the permissions it needs at a scope such as a resource group or resource.

A configurable safeguard that inspects prompts and model completions for hate, violence, sexual, and self-harm material, and can block or annotate responses before the application receives them.

The capacity, usually measured in tokens or requests per minute, allotted to a model deployment in a region; exceeding it causes throttling errors, so applications need retry logic.

A resource setting that streams request logs and metrics to Log Analytics, storage accounts, or event hubs, enabling auditing, latency analysis, and alerting for an AI resource.

A network interface that assigns an Azure AI resource a private IP address inside a virtual network, so traffic from applications never travels over the public internet.

The instruction block sent ahead of a conversation that sets the model's role, rules, tone, and output format; it outranks user turns in most chat models.

A sampling parameter that controls randomness: low values produce more predictable, repeatable answers, while high values produce more varied and creative text.

The practice of shaping instructions, examples, and supplied context so a language model returns accurate, consistently formatted answers without any retraining of the model.

Placing trusted reference content, such as retrieved documents or database rows, inside the prompt so the model's answer rests on supplied facts rather than only on its training data.

A pattern that searches an index for content related to the user's question, inserts those passages into the prompt, and lets the model compose an answer based on them.

A numeric vector produced by a model to represent the meaning of text, such that passages with similar meaning end up close together in vector space.

A searchable collection of embeddings that answers similarity queries quickly, letting an application find the passages most related to a user's question.

Splitting long documents into smaller, often overlapping passages before embedding, so retrieval returns focused content that fits inside the model's context window.

The maximum number of tokens a model can consider in a single request, covering the system message, conversation history, retrieved content, and the generated reply together.

A model capability in which the model requests that the application run a named function with structured arguments, then uses the returned result to continue composing its answer.

An AI application that pairs a model with instructions, tools, and stored conversation state so it can plan multiple steps and act toward a goal instead of replying once.

The stored sequence of messages and tool outputs for one conversation with an agent, which the service replays to the model on each turn to preserve continuity.

A check that judges whether a generated answer is supported by the source content supplied with it, flagging responses that contain unsupported or fabricated claims.

Extracting machine-readable text, together with its positions, from photographs of documents, receipts, and signs so the text can be searched, indexed, or processed further.

Assigning one or more descriptive labels to an entire image, such as a category of product or scene, using a trained custom model or a prebuilt vision model.

Locating items within an image by returning a bounding box, a label, and a confidence value for each detected instance rather than a single label for the whole picture.

A vision model trained on your own labeled images to recognize classes or objects specific to your business, then deployed to an endpoint that returns predictions.

Scoring text as positive, negative, or neutral, optionally sentence by sentence, to gauge opinion in reviews, surveys, support tickets, or social posts.

Pulling out the main concepts or topics from a document so its content can be tagged, grouped, summarized, or matched against search queries.

Detecting mentions in text and classifying them into categories such as people, places, organizations, dates, quantities, and events for later structuring or filtering.

Identifying personal data such as names, phone numbers, addresses, and account identifiers in text, and optionally redacting it so content can be stored or shared safely.

Condensing a long document or conversation into a shorter abstract or extractive summary that keeps the essential points while sharply reducing length.

A trained model that maps a user's utterance to an intent and extracts relevant entities, typically used to route requests inside a bot or voice application.

A curated knowledge base of question-answer pairs and passages that returns a short answer with a confidence value for a user's natural-language question.

A ready-made extraction model for common document types such as invoices, receipts, or identity cards that returns typed fields with confidence values from an uploaded file.

A Document Intelligence model trained on your own labeled samples, using template matching or neural learning, to pull out fields unique to your organization's forms.

Several extraction models assembled behind one model identifier, so a single request automatically routes each submitted document to the model that fits it best.

A value, usually between zero and one, expressing how certain the service is about a detected label, extracted field, or answer; low values usually warrant human review.

A pipeline of AI steps attached to a search indexer that extracts text, generates embeddings, and derives new fields from each document as it is indexed.

A search component that reads a data source, runs each document through an optional skillset, and writes the resulting fields into an index on a schedule or on demand.