jina-reranker-v1-base-en

jina-reranker-v1-base-en is a 137M-parameter English cross-encoder reranker that refines search results by performing token-level analysis of query-document pairs. It delivers a 20% improvement in search accuracy over baseline vector search, making it the workhorse of Jina's first-generation reranker family. It is now superseded by the multilingual v2 and listwise v3 families.

The model employs a BERT-based cross-attention architecture (12 layers, 12 attention heads, 768 hidden dimensions, 137M parameters) that fundamentally differs from embedding-based approaches. Instead of comparing pre-computed document embeddings, it performs dynamic token-level interactions between query and document, capturing contextual nuances that cosine similarity misses. The 1,024-token context window supports automatic chunking and max-pooling across chunks for longer documents. Training used contrastive loss on query-document relevance pairs with hard-negative mining. The model was the first Jina reranker and established the pattern for the turbo and tiny variants through knowledge distillation.

The model achieves an 8% increase in hit rate and a 33% boost in mean reciprocal rank compared to baseline vector search. On BEIR, it scores an average NDCG@10 of 0.5588, outperforming BGE (0.5032), BCE (0.4969), and Cohere (0.5141). On the LoCo benchmark for long-context understanding, it scores 0.873, significantly ahead of competitors. It shows particular strength in technical content: 0.996 on QASPER abstracts and 0.962 on government report analysis, though performance is lower (0.466) on meeting summarization tasks. Latency scales with document length: 156ms for 256 tokens, 7,068ms for 4,096 tokens with a 512-token query.

Implement a two-stage pipeline: vector search (e.g., jina-embeddings-v5-text-small) provides initial candidates, then this model re-ranks the top 100–200 for precision. The model is English-only — for multilingual applications, use jina-reranker-v2-base-multilingual or jina-reranker-v3.5. It requires CUDA-capable hardware for production throughput. For new projects, prefer jina-reranker-v3.5 (multilingual, listwise, 131K context, 63.20 BEIR NDCG@10). The model is available via Jina Reranker API, AWS SageMaker, and Hugging Face. When integrating with RAG systems, tune the number of documents sent for re-ranking based on latency requirements — 100–200 documents typically balances quality and speed.