LongEmbed
A long-document retrieval benchmark. In Voyage’s reranker evaluation it consists of three subsets: NarrativeQA, SummScreenFD and QMSum.
Use in This Vault
rerank-3 and rerank-3-lite - the next generation of Voyage rerankers uses it as the long-document domain for 32K-context rerankers. There, Voyage rerank-3 led rerank-2.5 by 3.35% NDCG@10 and Cohere Rerank v4.0 Pro by 13.84%. The gains came from NarrativeQA and QMSum. Every Voyage reranker scored above 99% on SummScreenFD, so that subset has stopped discriminating between models at the top. An average over the three subsets hides this.
Related Concepts
- Reranking · Text Chunking — long documents are what chunking otherwise avoids