How would you handle information that exceeds an LLM’s context window?
Instruction: Give a clear explanation, a practical example, and the main limitation.
Updated
Example Answer
I would keep the full information outside the prompt and choose what the current task needs. That can mean retrieving relevant document sections, summarizing older conversation turns, and storing important constraints or unfinished work as structured state. It manages the available window; it does not enlarge the model’s native limit.
For a long policy review, I would retain source identifiers and effective dates, retrieve the relevant rule plus its exceptions, and reserve room for the answer. I would test cases where the needed detail is buried in older material. If retrieval or a summary loses a crucial condition, I would fetch the original evidence or ask for clarification before deciding.
Concrete example
A summary saying "travel is approved" loses the condition "only below $500 and before Friday." Keep the limit and deadline in explicit state and test whether later actions still honor both.
Follow-up to practice
Which information would you keep verbatim instead of summarizing?