Back to the wiki

Knowledge cutoff

How far the model knows what's happening in the world.

The analogy

It's like a printed book: it holds everything that was known the day it went to press, but nothing that happened after. A model learns from data up to a certain date; anything later, it simply never lived through.

In detail

Training uses data collected up to a given moment: that's the cutoff date. The model doesn't know later events unless you give them in the prompt or it fetches them with tools (web, RAG). That's why it sometimes answers with outdated facts, fully confident.

An example

You ask about something that happened last week and it replies with two-year-old information, or gets it wrong. If you need the latest, give it the source yourself or use a search tool.

Related concepts