Glossary

Token

The unit models read and bill in, roughly three-quarters of an English word. Other languages, code and unusual names use more tokens for the same content.

Context window

How much text a model can hold at once, counting your input and its output together. Quality typically degrades before the stated limit is reached.

Knowledge cutoff

The point after which the model was not trained on new information. It does not know about later events unless you supply them or it has a search tool.

Hallucination

A fluent, confident, incorrect output. Not lying, since there is no intent. It is the model producing a plausible continuation where it has no grounding.

Reasoning model

A model that produces internal working before answering. Usually billed as output, and controlled by a dedicated effort or thinking parameter rather than by prompt phrasing.

Context engineering

Deciding what goes into the model's working memory: which documents, how much history, which tools. Increasingly a bigger lever than prompt wording.

RAG

Retrieval-augmented generation. Fetching relevant material and putting it in the prompt so the model answers from that instead of memory.

Prompt injection

An attack where instructions hidden inside content the model reads are treated as commands. The core reason agents need real permission boundaries.

MCP

Model Context Protocol, an open standard for connecting AI assistants to tools and data. Now governed by a Linux Foundation body rather than any single vendor.

Structured output

Forcing a response into a fixed shape such as JSON. Improves machine-readability, and carries a measurable quality cost on harder tasks.

Evaluation

Testing a prompt against a set of representative examples with a written standard, instead of judging by impression.

Deployer

In EU regulation, the organisation that uses an AI system in its own activity. Most businesses are deployers rather than providers, and several obligations land here.