Context window vs Knowledge cutoff

Why they differ

The context window is what the model can read right now; the knowledge cutoff is where its built-in knowledge stops - new facts can only get in through the window.

Context window

Language models

The most text, counted in tokens, that a language model can take in and keep in view at one time, including its own answer.

Formal

The fixed limit on how many tokens a language model can handle in one go; the prompt, earlier turns of the chat, added documents and the answer must all fit inside it.

In plain English

Like the size of a desk - everything the model is working on has to fit on it, and when it is full, something has to be cleared away to make room.

In practice

A clerk in a region works with an AI assistant through a long afternoon; it starts ignoring the format rules given at the start, because the chat app has cut the oldest messages to keep the rest inside the context window.

Why it matters

It limits how much a model can weigh in one answer, and everything inside it - including text from a file nobody checked - shapes that answer.

Knowledge cutoff

Language models

The date after which a language model saw no new text, so on its own it knows nothing about events that came later.

Formal

The last date covered by the training data of a large language model; the model's built-in knowledge stops there, even though it may be released and used for months or years after.

In plain English

Like a printed travel guide - everything in it was true when it went to print, but it cannot tell you about the restaurant that opened last month.

In practice

An IT teacher at a vocational school asks a chat assistant about the latest version of a code library for a lesson; it describes an older version as current, because the new one came out after its cutoff date.

Why it matters

Security advice goes stale fast - new weak spots, new attacks, new versions - so answers about recent events must be checked against current sources.

Shared connections

Large language model (LLM)Training data