7.4. Glossary

Accountability — The responsible-AI principle that humans remain answerable for an AI system's behavior and governance. (See 2.1.6.)

Agent — A generative AI application that reasons about a goal and autonomously calls tools or takes actions to achieve it, beyond generating text. (See 3.3.3, 4.2.)

AIProjectClient — The Foundry SDK client used for project setup and configuration; created with a project endpoint and a credential. (See 4.1.3.)

Analyzer — A configured Content Understanding extractor that defines which fields to pull from content; prebuilt or custom. (See 6.2.1.)

Content Understanding — Azure capability that extracts structured information from unstructured documents, images, audio, and video. (See 6.1.)

Context window — The maximum number of tokens a model can consider at once (prompt plus completion). (See 3.1.1.)

DefaultAzureCredential — Standard Azure authentication that uses your signed-in identity instead of hard-coded keys. (See 4.1.3.)

Deep learning — Machine learning that uses many-layered neural networks; a subset of ML. (See 1.1.2.)

Entity detection — Text analysis that identifies named things (people, places, dates, organizations). (See 3.3.1.)

Fairness — The responsible-AI principle that a system treats all groups equitably and doesn't amplify bias. (See 2.1.1.)

Generative AI — AI that creates new content by predicting likely next tokens, rather than retrieving stored facts. (See 1.2.2.)

Grounding — Supplying a model with trusted, relevant information in the prompt so output is based on it; the main defense against hallucination. (See 3.2.2.)

Hallucination — Fluent but factually incorrect output from a generative model, arising from prediction rather than lookup. (See 1.2.2.)

Image generation — Creating a new image from a text prompt (text in, image out). (See 3.3.2, 5.3.)

Inclusiveness — The responsible-AI principle that AI empowers people of all abilities and backgrounds. (See 2.1.4.)

Information extraction — Pulling meaningful structured fields out of unstructured content; more than OCR. (See 3.3.3, 6.1.)

Large language model (LLM) — A generative model trained on large text corpora to predict text well across many topics. (See 3.1.1.)

Machine learning (ML) — Building AI by learning patterns from data instead of hard-coding rules; a subset of AI. (See 1.1.2.)

Microsoft Foundry — The unified Azure platform for deploying models, building agents, and using AI tools (formerly Azure AI Foundry). (See 4.1.)

Multimodal model — A model that accepts more than one input type (e.g., text and images) in a single call. (See 5.1.2.)

Object detection — Computer vision that locates and identifies multiple objects in an image. (See 3.3.2.)

OCR (optical character recognition) — Reading raw text out of an image; a building block of, but not the same as, information extraction. (See 3.3.2, 6.1.1.)

Privacy and security — The responsible-AI principle covering protection of personal data (privacy) and defense of the system (security). (See 2.1.3.)

Prompt — The input text sent to a generative model; split into a system prompt (persistent rules) and user prompt (the request). (See 3.2.2, 4.1.1.)

PTU (Provisioned Throughput Unit) — Reserved model capacity for predictable, high-volume workloads, versus pay-as-you-go. (See 3.2.1.)

Reliability and safety — The responsible-AI principle that a system behaves predictably and safely, including in unexpected conditions. (See 2.1.2.)

Sentiment analysis — Text analysis that classifies emotional tone as positive, negative, or neutral; measures tone, not truth. (See 3.3.1.)

Speech recognition — Converting spoken audio into text (speech-to-text). (See 5.2.1.)

Speech synthesis — Generating spoken audio from text (text-to-speech). (See 5.2.1.)

Summarization — Text analysis that produces a shorter version preserving key points. (See 3.3.1.)

System prompt — The instruction that sets a model's or agent's persistent role, rules, and constraints. (See 4.1.1.)

Temperature — A runtime parameter controlling output randomness/creativity, not capability. (See 3.2.1.)

Token — A small chunk of text (word-piece) the model reads and predicts; pricing and limits are measured in tokens. (See 3.1.1.)

Transparency — The responsible-AI principle that people can understand how and when AI is used and its limitations. (See 2.1.5.)