Context window

The context window is the token budget a model can consider in one call. It includes instructions, conversation history, retrieved passages, tool results, the user’s message, and room for the answer.

Tiny example

system instructions  1,000 tokens
conversation history 3,000 tokens
retrieved evidence   8,000 tokens
answer allowance     2,000 tokens
total               14,000 tokens

What it is not

A large context window is not long-term memory, and fitting more text does not ensure the model will use every detail well.

FDE note

Set an explicit budget for every part of the request. Preserve the highest-value evidence and source metadata when trimming.