jina-reranker-v3: Last but Not Late Interaction for Document Reranking

2025-09-29

AI Summary

To address the trade-off between efficiency and effectiveness in multilingual document re-ranking, this paper proposes a compact causal self-attention re-ranker. The architecture models query–document interaction in a unified context window using a “last-but-not-late” strategy—preserving full semantic context without early truncation while avoiding the redundant autoregressive decoding overhead of generative models. It further incorporates multilingual pretraining and a context-sensitive end-token embedding extraction mechanism to strengthen cross-lingual semantic alignment. Evaluated on the BEIR multilingual benchmark, the model achieves a state-of-the-art nDCG@10 of 61.94 with only one-tenth the parameters of leading generative listwise re-rankers. This yields substantial improvements in inference latency and deployment feasibility, particularly for resource-constrained multilingual applications.

Technology Category

• Natural Language Processing: Machine Translation, Multilinguality, Cross-Lingual NLP
• Machine Learning: Large Multimodal Models (LMMs)
• Search and Optimization: Learning to Search

Application Category

• Search and Retrieval-Augmented AI: Multilingual and cross-lingual Web search
• User Modeling, Personalization and Recommendation: Fairness-aware retrieval and ranking
• Web Mining and Content Analysis: Mining multimedia, multimodal, multilingual, cross-lingual Web data

Abstract

jina-reranker-v3 is a 0.6B parameter multilingual document reranker that introduces a novel last but not late interaction. Unlike late interaction models such as ColBERT that perform separate encoding followed by multi-vector matching, our approach conducts causal self-attention between query and documents within the same context window, enabling rich cross-document interactions before extracting contextual embeddings from the last token of each document. This compact architecture achieves state-of-the-art BEIR performance with 61.94 nDCG@10 while being ten times smaller than generative listwise rerankers.

Problem

• Develops compact multilingual reranker with novel interaction mechanism
• Enables cross-document interactions before embedding extraction
• Achieves state-of-the-art performance with smaller model size

Innovation

• Causal self-attention between query and documents
• Contextual embeddings from last token extraction
• Compact multilingual reranker with state-of-the-art performance