AI-200 key terms

A cloud portal and workspace for building, evaluating, deploying, and governing AI applications on Azure, organized into hubs and projects that share connections, security, and access control.

A browsable library in Azure AI Foundry listing foundation and prebuilt models with their capabilities, providers, and deployment options, so teams can choose a suitable model for a task.

Publishing a model to an endpoint with a chosen hosting option, such as standard, global, or provisioned throughput, so applications can send inference requests and receive responses.

A deployment option that reserves model processing capacity for predictable latency and sustained throughput, billed by capacity units rather than by each token consumed.

A configurable safety layer that inspects prompts and model completions for harmful categories such as hate, violence, sexual content, and self-harm, then blocks or annotates matching content.

An Azure identity created and maintained by the platform and assigned to a resource, allowing that resource to authenticate to other services without storing credentials in code or configuration.

An authorization system that assigns users, groups, and service principals specific roles on Azure resources, determining which management and data actions they are permitted to perform.

A resource setting that sends telemetry such as request metrics, errors, and audit events to Log Analytics, storage accounts, or event hubs for monitoring and troubleshooting.

A service that exposes OpenAI models such as GPT chat models and embedding models through Azure endpoints, adding Azure networking, identity, regional data handling, and safety controls.

A unit of text, roughly a fragment of a word, that a language model reads or generates; billing, context limits, and truncation are measured in these units.

A privileged instruction block placed at the start of a chat conversation that sets the model's role, tone, rules, and boundaries separately from user turns.

A sampling parameter that scales randomness in generated output: lower values make responses more deterministic and focused, while higher values increase variety and creativity.

A numeric vector representing text meaning, produced by a model, so that semantically similar passages land close together and can be compared by distance or similarity.

A pattern that retrieves relevant documents from a search store and inserts them into the prompt so the model answers with grounded, current, source-specific information instead of relying only on training data.

Training a base model further on curated example data to adapt its behavior, style, or task performance, producing a custom model that is deployed and called like other models.

A software component that uses a language model to reason about a goal, choose tools, act, observe results, and iterate toward completion with limited human direction.

A declared capability, such as an API or function, that a model can request with structured arguments so the host application executes it and returns the result to the model.

A persistent conversation state maintained by an agent service that stores messages and context across runs so an agent can continue a dialogue over time.

A single execution of an agent against a thread, during which the service reasons, calls tools, and produces messages or status changes until it finishes or needs input.

The coordination of multiple agents or steps, deciding which agent acts next, passing context between them, and combining their outputs into a final response.

A design where an agent pauses for a person to approve, correct, or supply information before continuing, used for high-risk, regulated, or ambiguous actions.

A vision service capability that extracts descriptions, tags, objects, and text from images through prebuilt or custom models exposed by a single analysis operation.

The process of detecting and converting printed or handwritten text in images and documents into machine-readable characters, often with positional information for each word or line.

A computer vision task that locates instances of trained classes in an image by returning bounding boxes and confidence scores for each detected object.

A computer vision task that assigns one or more trained labels to an entire image based on its visual content, without reporting where those labels occur.

A vision capability that analyzes people and movement in camera video streams, producing events and counts used for occupancy, safety, and flow scenarios.

A service offering prebuilt and customizable natural language features such as sentiment, entity extraction, summarization, and intent classification through APIs and SDKs.

A customizable model that maps user utterances to intents and extracts entities from them, used to interpret commands and requests in a conversational application.

A feature that builds a knowledge base from documents or FAQs and returns precise answers to natural-language questions, optionally using a language model to compose responses.

The extraction of named things in text, such as people, organizations, places, dates, and quantities, with each mention labeled by its category.

A text analysis task that scores whether a passage expresses positive, negative, or neutral opinion, optionally at the sentence level and with confidence values.

The identification, and optional redaction, of personal data such as names, phone numbers, addresses, and account identifiers in text.

A cloud search service that indexes content, runs queries, and supports keyword, vector, and hybrid retrieval plus enrichment pipelines for AI and knowledge mining applications.

A searchable data structure that stores documents as fields, including vector fields, and controls which fields can be searched, filtered, sorted, or returned in results.

A component that connects to a data source, reads documents, runs an optional skillset, and writes enriched content into a search index on a schedule or on demand.

An ordered collection of enrichment steps in Azure AI Search that transforms content, for example by extracting text, detecting language, or generating embeddings before indexing.

A retrieval method that compares a query embedding to stored document embeddings and returns nearest neighbors by similarity rather than by exact term matching.

A query approach that combines keyword scoring and vector similarity, then merges the result sets so exact terms and semantic meaning both influence final ranking.

A target in Azure AI Search enrichment that projects enriched documents into Azure Storage as tables, objects, or files for downstream analytics and reporting.

A service that extracts text, tables, key-value pairs, and document structure from forms and files using prebuilt or custom models.