Cross-Encoder Rerank Top-k

Background

After bi-encoder retrieval returns the top-$N$ (say 50 candidates), a cross-encoder reranker scores each (query, candidate) pair more accurately. The pipeline then takes the top-$k$ by rerank score - typically 5-10 chunks for RAG. This is the simple selection step, not the cross-encoder itself.

Problem statement

Implement rerank_topk(candidates, rerank_scores, k). Reorder candidates by the parallel rerank_scores (descending) and return the top-$k$.

Input

• candidates - sequence of items (any type).
• rerank_scores - parallel array of floats.
• k - int, cutoff.

Output

• list of the top-$k$ candidates in rerank order.

Examples

Input: candidates=["a","b","c"], rerank_scores=[0.1, 0.9, 0.5], k=2
Output: ["b", "c"]

Constraints

• Lengths must match; ValueError otherwise.
• Tie-break by lower original index (stable sort).
• $k$ may exceed len(candidates).

Notes

• Two-stage retrieval. Bi-encoder for speed at scale, cross-encoder for quality on the shortlist. Top-k after rerank is what enters the LLM context.