Long-Context Strategies

Strategy

Part of: The Context Window

You have a 500-page manual and a model that holds maybe 200 pages. The question that decides your whole architecture is deceptively simple: do you cram, compress, or fetch? Each answer is a real strategy with real trade-offs, and picking wrong means slow, expensive, or wrong answers. This lesson ties together everything from the module into three named moves. What it is When the input is bigger than what you want in the window, there are three strategies: 1. Stuff it. Put as much as fits directly into the window. Simple and complete, until the input outgrows even a large window, or the cost and the lost-in-the-middle effect bite. 2. Summarize. Compress the input (or the history) into a shorter form first, then send the summary. Cheap and scalable, but lossy, detail is thrown away. 3. Retrieve. Pull in only the chunks relevant to this question and send just those. Precise and cheap, but it depends on retrieval actually finding the right pieces. How it works The three differ in what they put on the desk. Stuffing sends everything; summarizing sends a shrunken everything; retrieval sends a tiny relevant slice: The decision hinges on two questions: does it fit, and does the question ne

Challenge: Strategy Router