Two-Stage Retrieval with a Reranker

Production RAG retrieves cheaply first (BM25 or vectors), then reranks the top few candidates with an expensive cross-encoder that reads the query and document together.

Implement rerank(query, docs, first_scores, cross_score, top_n, alpha=1.0):

• Take the top_n docs by first_scores (ties: lower index first).
• Score each of them with cross_score(query, doc). Only call it for those top_n docs, because it's expensive.
• Min-max normalise both the first-stage scores and the cross scores within the top_n (if all values are equal, they normalise to 0).
• The final score is alpha * cross_norm + (1 - alpha) * first_norm. Sort the top_n by final score, highest first, breaking ties by first-stage rank.
• Append the remaining docs in first-stage order. Return the list of docs.