จัดลำดับผลลัพธ์ที่ดึงมาใหม่

Vector search is fast but approximate. Reranking takes the top candidates and reorders them with a more accurate, slower model — the two-stage retrieve-then-rerank pattern.

Bi-encoders embed query and document separately (fast, used for retrieval). Cross-encoders score query+document together (slow, far more accurate) — ideal for reranking a small set.

Step 1: retrieve top 50 with the vector DB. Step 2: rerank those 50, keep the top 5. You get cross-encoder quality at near vector-search speed.

A cross-encoder takes pairs and outputs a relevance score. Higher means more relevant.

Pair each document with its score and sort descending to produce the final order.

Services like Cohere Rerank and Jina offer hosted rerankers via API, removing the need to self-host a model.

Retrieve enough candidates that the true best answer is in the set (high recall), but not so many that reranking gets slow. 20-100 is typical.

Reranking adds latency proportional to candidate count. Cache results, batch the cross-encoder calls, and tune the candidate count to your SLA.

Apply metadata filters during retrieval, then rerank only the filtered set. This keeps the reranker focused on valid candidates.

Compare nDCG or hit rate with and without reranking on a labeled set. Reranking commonly delivers a large accuracy boost on noisy corpora.

If your corpus is small and clean, or latency is critical, plain vector search may suffice. Reranking pays off most on large, diverse datasets.

Test your reranking knowledge.

You learned the retrieve-then-rerank pattern, the difference between bi- and cross-encoders, how to score and sort candidates, and when reranking is worth its latency cost.