Context window

A context window is the amount of text a model can consider at once, covering the instructions, retrieved documents, and conversation history supplied with a request. Anything outside it is invisible to the model, so what gets selected into the window determines the quality of the answer.

Larger windows reduce but do not remove the selection problem. Filling a window with marginally relevant material degrades answers and raises cost, which is why retrieval quality still matters when a model could technically accept far more.

In enterprise systems the window is also a permissions boundary. Whatever is placed in it has effectively been disclosed to that request, so retrieval must apply the requesting user's access rights before anything is assembled.