How do you respond to a request to retrain a reranker w...

The reranker remains byte-for-byte the same size and the serving cost is identical. However, one must acknowledge the real risks: increased training FLOPs, the complexity of including the production first-stage in the training loop, and the risk of index staleness. Furthermore, if downstream systems threshold the reranker score, shift-invariance means the new scores may not be comparable across queries, potentially breaking downstream consumers.