Glossary

The terms used on this site, each in one or two plain sentences. Follow the link for the full explanation.

AI agent

A program that is given a goal, decides its own next steps, uses tools to act, and carries on until the work is done or it needs a person. Full explanation

Chatbot

A program that answers a question and stops. Whatever happens next, the person does. Full explanation

Workflow

A fixed sequence of steps written in advance, in which a model does one or more of the steps but does not choose the path. Full explanation

Language model

The engine inside AI assistants and agents: a program trained on a very large amount of text to predict what comes next. Full explanation

Parameter

One of the internal numbers a model adjusts during training. Large models have billions of them. Full explanation

Reliable knowledge cutoff

The date up to which a model's knowledge can be relied on. For events after it, and anything never made public, the model cannot be relied on. Full explanation

Hallucination

An answer that sounds right but is false, produced by a language model with the same confidence as a correct one. Full explanation

Context window

Everything the model can see while working on a request: instructions, documents, the conversation so far and its own answer. Its working memory. Full explanation

Token

A small piece of text, often part of a word. Context windows are measured in tokens. Full explanation

Context rot

The decline in a model's accuracy as more and more text is placed in its context window. Full explanation

Tool

One specific thing an agent is allowed to ask for, such as searching a mailbox. The model asks; the software carries it out. Full explanation

MCP (Model Context Protocol)

An open standard for connecting AI applications to the systems where data lives, so a system connected once can serve many AI applications. Full explanation

Excessive agency

The risk of giving an agent more functions, permissions or independence than its job needs. Full explanation

Retrieval

Giving a model your own documents at the moment it answers: the agent searches first, then answers from what it found. Full explanation

RAG (retrieval-augmented generation)

The full name of the retrieval method: retrieve the relevant passages first, then generate the answer from them. Full explanation

Fine-tuning

Training a model further on your own examples. A 2023 study found it a weak way to teach a model new facts. Full explanation

Prompt

The instructions that tell a model what to do. For an agent, they work as its job description. Full explanation

Human in the loop

A person takes part at chosen points in an agent's work, usually to approve an action before it happens. Full explanation

Auditor agent

A separate agent whose job is to check the work of other agents, look for what went wrong, and fix only what it is sure of. Full explanation