An input/output limit bounds how much text a generative AI can read and generate in one interaction. The context window includes both prompt and response tokens, so a long prompt leaves less room for the answer. section 1.1
Track the limit through tokens and model behavior:
- A token may be a word, part of a word, or punctuation mark; tokenization determines the division, averaging about four tokens per three words. section 1.1
- Limits vary substantially across models and implementations, with context windows ranging from several thousand to millions of tokens. section 1.1
- When earlier content falls outside the window, the system may provide no warning; inconsistent replies can reveal that previous exchanges were forgotten. section 1.1
- Summarize the conversation and key findings regularly to keep important information within the available context. section 1.1
A forgotten instruction
After a conversation grows beyond the context window, a response may contradict an earlier instruction or finding. That inconsistency indicates that earlier prompts or responses have been dropped from active context. section 1.1
