Does anyone have data on how much better these 1M token context models produce better results than the more limited windows alongside certain RAG implementations? Or how much better in the face of RAG the 200k vs 1M token models perform on a benchmark?