Article
CommunityJev Guides & Articles
Jev as an Elasticsearch Search Reranker
Elastic's Dustin Coates scores Elasticsearch hybrid search results with Jev and reranks them with a small Python policy, lifting nDCG@10 from 0.9351 to 0.9565 and Exact MRR from 0.9201 to 0.9616 across 250 Amazon Shopping Queries.
- Format
- Article
- Reading time
- 2 min
- Published by
- Community
- Author
- Dustin Coates (Elastic)
- Last verified
- Oct 2026
- Languages
- Python
Highlights
- The best Jev policy lifted nDCG@10 from 0.9351 to 0.9565 on 250 US English queries from Amazon's Shopping Queries Dataset.
- The same run improved Exact MRR from 0.9201 to 0.9616, closing roughly a quarter of the gap to perfect ranking.
- The benchmark paired jev-1.13.0 with 4,754 labeled query/product pairs and compared BM25, hybrid search, and jina-reranker-v3.
- Elasticsearch retrieved up to 40 candidates with an RRF retriever mixing multi_match and semantic_text search; Jev judged each pair.
- The Choice question sorted pairs into exact, substitute, complement, and irrelevant; application code turned the scores into the final order.
Watch out
Published by Elastic with Jev as the only reranker under test on a single dataset of 250 queries and 4,754 pairs, so the gains are specific to that setup; no licence is stated for the companion notebook. The 70–500 ms latency and $0.042 per million input tokens are TypeSafe claims repeated by the article.
Questions about this entry
Short answers drawn from the linked project or page.
It uses Jev to score query/product pairs returned by Elasticsearch hybrid search and a small Python policy to rerank them, benchmarking against BM25, plain hybrid search, and Elasticsearch's Jina reranker on ecommerce search data.
The best Jev policy raised nDCG@10 from 0.9351 to 0.9565 and Exact MRR from 0.9201 to 0.9616, closing roughly a quarter of the gap between hybrid search and perfect ranking on the benchmark queries.
250 US English queries and 4,754 labeled pairs from Amazon's public Shopping Queries Dataset, jev-1.13.0 for the judgments, jina-reranker-v3 as the comparator endpoint, and a Python 3.11 companion notebook for the run.
Elasticsearch retrieves up to 40 candidates with an RRF retriever combining multi_match lexical search and semantic_text search; Jev judges each query/product pair; application code converts those judgments into the final ranking score.
The first implementation used a Choice question over the dataset's four relationships — exact, substitute, complement, or irrelevant — while the page also reports testing several settings including Choice and Score question types across requests.
Elasticsearch score, prior judgments, and original position were left out so they could not improperly influence the model, and product IDs were excluded too. Jev saw only the query plus the product's title, brand, color, bullet points, and description.
It separates retrieval from reranking by judging the same fixed product groups, so it measures ordering quality only, and the full policy code lives in an external companion notebook. The page states no licence and reports no measured pipeline latency.
Reactions & coverage
Posts, threads, and videos about this entry from around the web.
X: Search and tagging on keep.md
X: Jev reranks station suggestions in TrainLCD
YouTube: Jev in the reranking step of a RAG pipeline: why cosine similarity fails and how a steerable reranker compares. A Colab notebook comes with it.
More Jev resources like this
A step-by-step Chinese guide to putting Jev behind backend approval flows: obtain a JEV_API_KEY, send code, type, and scene to the evaluation endpoint, then route by confidence — 0.90 and above auto-pass, 0.70 to 0.90 human review, below 0.70 reject or retry.
Jev Guides & ArticlesArticle#community#article#api
Community
Kervin Hu's Neo4j Developer Blog experiment runs Jev Noul over 1,822 blocker-filtered candidate pairs from 703 synthetic contact records: F1 0.943 at a 0.5 threshold, about two cents per thousand pairs, and a hybrid reaching 0.972.
Jev Guides & ArticlesArticle#community#article#entity-resolution
Community
Philipp Acsany's tutorial replaces a Y-or-N input check with Jev, then asks Noul, Score, and Choice in one call through OpenRouter.
Jev Guides & ArticlesArticle#community#article#tutorial
Community
Datadog's walkthrough of one Jev rubric that scores every criterion in a single request, then runs as online evals on live spans and as offline experiments.
Jev Guides & ArticlesArticle#community#article#evaluation
Community
A curated community list of public projects and practices built on Jev, grouped by direct Jev application domain; the October 2026 crawl lists 481 repositories plus hundreds of practice links, interviews, and discussions.
Jev Guides & Articles#community#awesome-list#documentation
Community
yibie/awesome-jevPython
yibie/awesome-jevPythonPlugs Jev-style decision models into published rerankers as a fast primitive: one forward pass over k candidates, zero decoded tokens. On T2Ranking it reports up to 22x faster matches than a generative listwise reranker at 2.9x the nDCG@10.
Jev Repos & SDKs#community#catalog-import#yibie
Community
hgqimo/jevrankerPython
hgqimo/jevrankerPythonFrom the Jev guides
Original write-ups on what this entry is for.
From the Jev AI model community
Posts from builders shipping with Jev right now.