Overview

mxbai-rerank-base-v2 is a powerful 0.5 billion parameter reranker model from Mixedbread.ai, built for efficient and high-performance document re-ranking. It is part of a family of reranker models, with a larger variant also available. The model was developed using a sophisticated three-step training process involving Guided Reinforcement Prompt Optimization (GRPO), Contrastive Learning, and Preference Learning, ensuring robust performance across various tasks.

Key Capabilities

• State-of-the-art performance: Achieves strong results across benchmarks, with a BEIR Avg score of 55.57.
• Multilingual support: Supports over 100 languages, demonstrating outstanding performance in English and Chinese (83.70 Chinese score).
• Code support: Capable of handling code-related search and reranking tasks (31.73 Code Search score).
• Long-context support: Designed to process and rerank documents with extended context lengths.
• Efficiency: Offers competitive latency, measured at 0.67 seconds on an A100 GPU.

Good for

• Improving the relevance of search results in information retrieval systems.
• Re-ranking documents in question-answering systems to prioritize the most pertinent information.
• Enhancing the precision of retrieved code snippets or documentation.
• Applications requiring high-quality reranking across a broad spectrum of languages.