Glossary: the basic words

These are the words you meet first when you build or deploy something with AI, each explained using only plain words or other words on this page. Terms are in alphabetical order. A word in italics inside a definition has its own entry here.

Jump to: A · C · D · E · F · G · H · I · L · M · O · P · R · S · T · W · Z

A2A (Agent2Agent)

An open set of rules for one agent to find another agent, give it a task and follow its progress, even when the two are run by different companies. It started at Google and is now a Linux Foundation project. → Agents working together

Agent

A program in which a model decides, step by step, which tools to use to finish a task, and keeps going until it is done or stopped. A chatbot answers; an agent acts. → What an AI agent is

Agent SDK

A code library that gives you the loop, tool calls and memory handling of an agent, so you can build your own agent without writing that part. Examples: Claude Agent SDK, OpenAI Agents SDK, Google ADK. → Ways to use AI

API

A way for one program to ask another program to do something over the internet, by sending a structured request and getting a structured answer. A model API lets your app send a prompt and receive the model's answer. → Ways to use AI

API key

A long secret string that identifies your account when your program calls an API. Anyone who has it can use, and spend from, your account, so it is a secret. → Security

Chatbot

An app where a person types messages and a model answers, one turn at a time. It does not take actions on its own; that is what an agent adds.

Cloud

Computers, storage and software rented over the internet, on demand, instead of owned. For AI apps it covers three separate jobs: running your app, running the model your app calls, and sometimes running a model on your own rented GPUs. → The cloud, for AI apps

Coding agent

An agent that works inside a code project: it reads files, edits them, runs commands and checks the result. It can run in a terminal, an editor or on the vendor's servers. → Ways to use AI

Container

A packaged copy of your app together with everything it needs to run (libraries, settings), so it runs the same on your laptop and on a host. A Dockerfile is the recipe for building one. → The cloud

Context window

The amount of text a model can take into account at once: your prompt, earlier messages, documents and its own answer, counted in tokens. Anthropic calls it the model's "working memory" (Anthropic glossary, checked 2026-10-28). What does not fit is not seen.

Deploy

To put your app on a host so it runs there and others can reach it at a web address. Redeploying means replacing the running version with a new one. → From idea to deployed app

Dockerfile

A short text file listing the steps to build a container: which base system to start from, which files to copy, which command starts the app.

Environment variable

A named setting given to a program from outside its code, such as PORT=8000 or MODEL_API_KEY=.... Keeping settings and secrets in environment variables means the same code runs everywhere and no key sits in the code. → Keys and secrets

Fine-tuning

Training an existing model further on your own examples, so it picks up a style, format or specialty. Many apps never need it: good prompts and RAG often do the job.

Generative AI

Software that produces new text, code, images or sound from an instruction, using a model trained on large amounts of existing data. → What generative AI is

GPU

A chip built to do very many small calculations at the same time. Training and running large models needs GPUs, which is why they are the expensive part of AI. You only rent them if you run a model yourself. → The cloud

Hallucination

When a model states something false or made up, in the same confident tone as true statements. It can be reduced but not removed, so important answers need checking. → Where it fails

Host

A company or service that runs your app on its computers and gives it a web address. Also called a hosting provider or platform. → The cloud

Inference

Running a trained model to get an answer. Each call to a model API is an inference; you pay for it in tokens. (Training is the separate, earlier step that creates the model.)

LLM

Large language model: a model trained on very large amounts of text that reads text and writes text. ChatGPT, Claude, Gemini and Le Chat are apps built on LLMs. → What generative AI is

Logs

The lines a running app writes about what it is doing: requests, errors, warnings. They are the first place to look when something goes wrong. → Read the logs

MCP (Model Context Protocol)

An open set of rules for connecting tools and data to an AI app, so one tool server works with many assistants. Created by Anthropic in 2024 and donated in 2025 to a foundation under the Linux Foundation. → MCP

Model

The trained program that turns an input (text, an image) into an output (text, an image). Models come in families and sizes, such as a large slow one and a small fast one. → Providers and models

Open-weight model

A model whose trained files (the "weights") are published, so you can download it and run it on your own computers, under its licence. The opposite is a closed model, which you can only use through its maker's API. → Closed vs open-weight

Prompt

The text you send to a model: your question or instruction, plus any documents or examples. A system prompt is the part that sets the rules for a whole conversation.

Prompt injection

An attack where text the model reads (a web page, an email, a file) contains instructions that make it do something its owner did not intend. The main new security risk of AI apps. → Security

Provider

A company that offers models through an API: either the maker (Anthropic, OpenAI, Google, Mistral…) or a cloud platform that resells many makers' models. → Providers and models

RAG

Retrieval-augmented generation: before asking the model, your app searches your own documents and puts the most relevant passages into the prompt, so the answer is based on them. Used for help bots that answer from your docs.

Rate limit

The maximum number of requests or tokens a service accepts from you per minute or per day. Above it, the service answers with an error (often 429) until the next period.

Secret

A value that must stay private, such as an API key, a password or a token. Secrets live in a host's secret settings or a secret manager, and reach the app as environment variables, never in the code. → Security

Serverless

A way of hosting where you give the host a function and it runs it only when a request arrives, with a time limit per request. You pay per use, but long or always-on work does not fit. → The cloud

Snapshot

A fixed version of a model that does not change behind your back, identified by an exact model ID. An alias, by contrast, can move to a newer version. → How versions work

Streaming

Sending an answer piece by piece as it is produced, instead of all at once at the end. Chat apps stream so you see words appear; this needs a connection that stays open, which affects hosting. → Time limits and streaming

System prompt

Instructions given to a model before the conversation starts, setting its role, rules and limits ("You are a support assistant for…; never…").

Temperature

A setting that controls how varied a model's answers are: low gives more predictable wording, high more varied. Even at zero, answers are not perfectly repeatable.

Token

The small piece of text a model reads and writes: a word, part of a word or a punctuation mark. Prices and limits are counted in tokens. OpenAI's rule of thumb for English: "1 token is approximately 4 characters or 0.75 words" (OpenAI docs, checked 2026-10-28). → What it costs

Tool

A function an agent is allowed to use, such as "search the web", "read a file" or "run a command". The model asks for the tool; your program runs it and returns the result. Also called tool use or function calling. → What an AI agent is

Workflow

A fixed sequence of steps, written in code in advance, in which a model may do some steps. Unlike an agent, the model does not choose what comes next. → Chatbot, workflow, agent

Zero data retention (ZDR)

An arrangement where a provider does not store your requests and answers after it has processed them. Usually only for business customers, sometimes only on approval. → Data protection by vendor

See also

Sources

All checked on 2026-10-28.


First version written with AI assistance and checked against the sources on 2026-10-28. Found a mistake or a missing word? Fix it or open an issue. Text licensed under CC BY 4.0.