LLMs
Context window
A context window is the maximum token span available to one model invocation.
How it works
Prompt tokens, retrieved text, tool outputs, and generated tokens compete for this finite budget. Exceeding it requires truncation, compaction, or another request.
A concrete example
Illustrative: a 32,768-token window with 28,000 input tokens leaves at most 4,768 tokens for output.
Do not confuse it with Knowledge cutoff
context window is per-request capacity; knowledge cutoff is the date boundary of training knowledge.