context length limits

A model can only attend to a fixed span of text at once — its context window, measured in tokens. Everything you want it to consider, your instructions plus the document plus the conversation so far, has to fit inside that window. Go over and the oldest material falls off the edge: the model simply cannot see it anymore, the way a desk has only so much room before the earliest papers slide onto the floor.

Two consequences follow. In a long chat, early turns can quietly drop out of view, so the model forgets a constraint you set at the start. And feeding a huge document does not guarantee the model truly used all of it — fitting inside the window is necessary but not sufficient for the model to weigh every part fairly. Windows have grown enormous, which helps, but bigger is not the same as reliable; long inputs bring their own failure, where the middle gets neglected.