mxbai-rerank-large-v2: A Powerful Reranker by Mixedbread.ai

mxbai-rerank-large-v2 is Mixedbread.ai's larger, 1.5 billion parameter reranker model, built for high-performance and efficient information retrieval tasks. It stands out for its state-of-the-art capabilities and broad applicability.

Key Capabilities

• State-of-the-art performance: Achieves strong results across various benchmarks, notably a BEIR Avg of 57.49.
• Multilingual support: Supports over 100 languages, with exceptional performance in English and Chinese (84.16 on Chinese benchmarks).
• Code support: Capable of handling code-related search and re-ranking tasks, scoring 32.05 in Code Search benchmarks.
• Long-context support: Designed to process and re-rank documents with extended context lengths.
• Efficient: Demonstrates competitive latency (0.89s on A100 GPU) compared to previous versions.

Training Methodology

The model was trained using a sophisticated three-step process:

1. GRPO (Guided Reinforcement Prompt Optimization)
2. Contrastive Learning
3. Preference Learning

Good for

• Improving the relevance of search results in RAG (Retrieval Augmented Generation) systems.
• Applications requiring precise document ranking across a wide array of languages.
• Enhancing code search and retrieval functionalities.
• Use cases demanding efficient processing of long documents for relevance scoring.