Long Context Windows Do Not Solve RAG: Accuracy Drops 30 to 50 Percent Before You Hit the Limit
中文摘要
长上下文窗口无法替代RAG，因为在达到限制前准确率会下降30-50%。接受海量Token并不等同于有效推理。
English Summary
Long context windows don't replace RAG; accuracy drops 30-50% before limits. Accepting tokens is not the same as reasoning over them.
Original Excerpt
Claude Opus 4.6 will happily accept a million tokens. Accepting them and reasoning over them are two different products. Continue reading on Medium »