How do you adjudicate between using a prompted LLM reranker and training an instruction-aware cross-encoder?

Prompting an LLM is excellent for measuring if intent disambiguation is valuable without the overhead of building a training set. However, it shifts costs from training to per-query generation and inherits positional bias and window-stitching artifacts. Adjudicate based on traffic: low QPS favors prompting, while high QPS with stable intents favors training a specialized model.