Why it matters
Context windows cap how much retrieved knowledge, instructions, and chat history fit in one request. Overflow truncates silently, often dropping safety rules or the most recent user clarification.
How it works
Budget tokens across system prompt, retrieved chunks, tool outputs, and history. Summarize or compress low-value turns. Monitor truncation events in logs.
Example
Nintendo trims older chat turns when Refund Policy excerpts plus Help Center steps approach the model limit, but never drops citation requirements or prohibited-claim instructions from the system block.
Common mistakes
- 1Stuffing maximum chunks because more context feels safer
- 2Truncating from the wrong end and losing fresh user details
- 3No alert when truncation rates spike after prompt changes
Definitions are useful. Governed knowledge is better.
Wiki helps you turn these concepts into a real system for the knowledge behind your AI.
See Wiki in action