You have six knowledge stores. Why not query all six an...

Querying all stores is not merely a latency or cost issue; it introduces significant functional failures. Compliance requirements may forbid querying specific stores entirely. Additionally, the tail error rate compounds as more sources are added. Finally, cross-encoders are pointwise, meaning they score near-copies independently and promote duplication into the context rather than filtering it. Empirical data shows that routing can significantly reduce query volume with negligible impact on accuracy.