Originally written in French. Translated by AI — the meaning has been preserved, not the prose.
Short definition
The amount of text a language model can receive and process in order to produce an answer — bounded, filled at each request, and emptied afterwards.
Full definition
In the established technical sense, the term designates a limit expressed in tokens: the maximum volume the model accepts as input. Used that way, it describes a capacity, and two models are compared on that figure.
In the articles on this blog, the word keeps that sense and adds a characterization that changes how it is used: the window is a resource whose filling has a cost in quality. What is in it isn't equally available — what is placed at the beginning or the end is better mobilized than what is buried in the middle. Maximum capacity and useful capacity therefore don't coincide, and the second is the one that counts when you decide what to load.
That is also what separates the word from its closest neighbour: the window fills and empties at each request, whereas the working context persists on disk and enters the window only through the share you pour into it.
Usage in the field
Used to explain why a workspace cannot be put back whole into an assistant session, and to situate what a loading rule really arbitrates.
Synonyms and variants
"the window" in speech; "context" on its own in technical teams' vocabulary, which is the main source of confusion.
Not to be confused with
- Context — a durable workspace gathering, around a goal, sources, notes, decisions and drafts. It persists and is loaded selectively; the window is ephemeral and bounded.
- prompt — an instruction addressed to the model at a given moment. It occupies part of the window, alongside everything else poured into it.
- Session memory — a written trace, kept over the course of exchanges, of what was done, decided and left pending. It exists outside the window, and enters it only when reloaded.
Examples
Loading a three-thousand-line file hoping the model will "understand everything": the volume enters the window, and the share actually mobilized stays unknown.
Ambiguities / debates
The word serves both as a sales argument — a bigger window presented as a bigger memory — and as a design constraint. The disagreement bears on what the enlargement settles: capacity increases, degradation by filling remains.