SCIENCE · RETRIEVAL

What is reranking?

AUTHOR · XERJ documentation team · REVIEWED BY · XERJ engineering team · PUBLISHED · 2026-11-13 · UPDATED · 2026-11-14

SHORT ANSWER · UPDATED 2026-11-14

What is reranking? It is a second stage after retrieval. A fast ranker, such as BM25, first returns a short candidate list. A slower model then reorders those candidates, and does not walk the whole index. XERJ's rerank stage once shipped at 0.3822 nDCG@10, then scored 0.8299 after a fix.

Key number: XERJ's rerank stage once shipped scoring 0.3822 nDCG@10 while plain BM25 scored 0.7750. Fixed, it scored 0.8299. (as of 2026-10-18 · source: /blog/does-xerj-beat-jev)

TL;DR Reranking is a second stage after retrieval. A fast ranker returns a short candidate list. A slower model then reorders those candidates, and does not walk the whole index again.

Try it: two stages

Retrieve first, then reorder

1. First stage
A fast ranker, such as BM25, retrieves a short candidate list from the whole index.
2. Second stage
A slower reranker looks at those candidates again and reorders them. It does not walk the whole index.
3. A published miss, then a fix
XERJ's rerank stage once shipped scoring 0.3822 nDCG@10 while plain BM25 scored 0.7750. One request shape was wrong. Fixed, it scored 0.8299. The loss stayed in the post.
4. When to skip it
If the first stage already puts the right passage first, a second stage may not be worth the cost. The rc.80 gates recorded which features passed and which failed.

The 0.3822, 0.7750 and 0.8299 nDCG@10 figures are from the published Does XERJ beat Jev? post, also told on the 100-contributors page. This walkthrough does not add a new run.

PASTE THIS TO YOUR AI AGENT — IT DOES THE REST

Act as a coding agent. Read https://xerj.org/llms.txt, install XERJ, start a local node, index one folder I point you to, and run one search so I can see a cited passage.

How it works

The first stage has to be cheap enough to search a large index. BM25 is a usual choice. It returns a short list, for example the top tens or hundreds of passages.

The second stage looks only at that list. It can use a slower model, because the list is small. The output is the same passages in a new order. Nothing new is retrieved.

How XERJ uses it

XERJ can rerank a first-stage list. The published post about that feature is honest about a miss. In the Does XERJ beat Jev? post, the rerank stage shipped scoring 0.3822 nDCG@10 while plain BM25 scored 0.7750. One request shape was wrong. Fixed, it scored 0.8299. The loss stayed in the post.

This page does not add a new run. For first-stage retrieval, see BM25 and hybrid search.

History and research

A widely cited early method is Nogueira and Cho, Passage Re-ranking with BERT (arXiv:1901.04085, 2019). That paper shows a pretrained language model can reorder a candidate list for question answering.

XERJ's published numbers above are from the product post, not from that paper.

Guides

See rerank search results with a calibrated judge, Cohere RAG with XERJ and the Cohere integration.

FAQ