GLOSSARY

Neural Re-ranking (Cross-encoder)

Neural re-ranking rescoring top candidates with a deep model greatly improves ordering. It reads the query and document together for fine-grained relevance.

Example H2

Example H3

Example H4

Example H5

Example H6

What is Neural Re-ranking?

Neural re-ranking uses a cross-encoder to jointly encode query and candidate text, producing high-quality relevance scores for the short list returned by recall.

How It Works (quick)

Pipeline: Recall (BM25/vectors) → top-k → cross-encoder scores → order.
Features: Handles phrases, context, negation better than bi-encoders.
Latency: Apply to small k (e.g., 50–200).

Why It Matters in E-commerce

Lifts NDCG/CTR noticeably on ambiguous queries.
Improves snippets/highlighting consistency.

Best Practices

Distill to lighter models for speed.
Use category-specific variants if needed.
Log explanations/attentions cautiously.

Summary

Cross-encoders fix ordering on hard cases. Keep k small and models optimized.