Context Window Optimization Guide

1. GPT-4-128K: No information loss when context < 73K tokens. Beyond 73K, information in the middle 7%-50% may be forgotten. 2. Claude-2.1-200K: Best under 24K context, degrades after 24K, very poor after 90K. 3. Keep context short; split long texts manually or automatically. For GPT-4, don't exceed 73K (input + output combined). 4. System Prompt can be at the beginning or end; placing it at the end is better. 5. If System Prompt is at the beginning, for GPT-4 keep it under 5% of context; for Claude, under 1%. 6. GPTs' prompts must be at the very beginning, so keep them under 5% (e.g., 128 * 0.05 = 6.4K tokens).