Reranking Models
The Mixedbread rerank family is a collection of state-of-the-art, open-source reranking models designed to significantly enhance search accuracy across various domains. These models can be seamlessly integrated into existing search systems, offering best-in-class performance and easy implementation for improved user satisfaction in search results.
Boost your search with our crispy reranking models! The Mixedbread rerank family offers state-of-the-art performance across a large variety of domains and can be easily integrated into your existing search stack.
What's new in the Mixedbread rerank family?
We recently finished baking a fresh set of rerank models, the mxbai-rerank-v2 series. These models feature reinforcement learning training, multilingual support for 100+ languages, and extended context handling up to 32K tokens. After receiving a wave of interest from the community, we're now happy to provide access to the models with the highest demand via our API:
| Model | Status | Context Length (tokens) | Description |
|---|---|---|---|
| mxbai-rerank-large-v2 | API available | up to 32k | Delivers the highest accuracy, performance and supports multiple languages |
| mxbai-rerank-base-v2 | API unavailable | up to 32k | Strikes balance between size and performance with better accuracy |
| mxbai-rerank-large-v1 | API available | 512 | Delivers the highest accuracy and performance |
| mxbai-rerank-base-v1 | API unavailable | 512 | Strikes balance between size and performance |
| mxbai-rerank-xsmall-v1 | API unavailable | 512 | Focuses on capacity-efficiency while retaining performance |
We are currently investigating finetuning and domain adaptation with a limited number of beta-testers. Please contact us if you are interested in using a reranking model tailored to your data.
Why Mixedbread rerank?
Not only are the Mixedbread rerank models powerful and fully open-source, they're also extremely easy to integrate into your current search stack. All you need to do is give the original search query as well as your search system's output to our reranking models, and they will tremendously boost your search accuracy - your users will love it!
Training Methodology
Our v2 models were built using a three-step reinforcement learning process:
- GRPO (Guided Reinforcement Prompt Optimization) - Teaching the model to output clear relevance scores
- Contrastive Learning - Developing fine-grained understanding of query-document relationships
- Preference Learning - Tuning the model to prioritize the most relevant documents
This layered approach yields a richer query understanding whether you're reordering text results, code snippets, or product listings.
Performance Benchmarks
We evaluated our models by letting them perform the reranking step on the top 100 lexical search results on a subset of the BEIR benchmark, a commonly used collection of evaluation datasets. Specifically, we used the NDCG@10 metric, which measures the overall relevance of the search results compared to the order in which they are ranked by the model, and the accuracy@3 metric, which measures the likelihood of a highly relevant search result appearing in the top 3 results - in our opinion, this is the most important metric to anticipate user satisfaction.
For illustrative purposes, we also included classic keyword search and a current full semantic search model in the evaluation. The results make us confident that our models show best-in-class performance in their size category:
| Model | BEIR Accuracy |
|---|---|
| mxbai-rerank-base-v2 | 55.57 |
| mxbai-rerank-large-v2 | 57.49 |
Latency Comparison
Below is latency per query (seconds) on the NFC dataset, tested on an A100 (80GB) GPU:
| Model | Latency (s) |
|---|---|
| mixedbread-ai/mxbai-rerank-xsmall-v1 | 0.32 |
| mixedbread-ai/mxbai-rerank-base-v2 | 0.67 |
| mixedbread-ai/mxbai-rerank-base-v1 | 0.76 |
| mixedbread-ai/mxbai-rerank-large-v2 | 0.89 |
| mixedbread-ai/mxbai-rerank-large-v1 | 2.24 |
| BAAI/bge-reranker-v2-m3 | 3.05 |
| BAAI/bge-reranker-v2-gemma | 7.20 |
Our 1.5B model is 8x faster than bge-reranker-v2-gemma while delivering higher accuracy.
Beyond Traditional Search
Our reranking models excel at numerous specialized tasks:
- Code and SQL Snippets: Perfect for developer docs or internal codebases
- LLM Tool Selection: Identify the right function from thousands of definitions
- E-Commerce: Combine product metadata to push the most relevant items to the top
- Multilingual Applications: Support for 100+ languages enables global use cases
Why should you use our API?
To get started, you can easily use our open-source version of the models. However, the models provided through the API are trained on new data every month. This ensures that the models understand ongoing developments in the world and can identify the most relevant information for any questions they might be asked without a knowledge cutoff. Naturally, our quality control ensures that the models' performance always remains at least similar to previous versions.
Last updated on
Overview
Reranking is a powerful technique that enhances search systems and RAG (Retrieval-Augmented Generation) setups by intelligently reordering initial search results based on semantic relevance. By integrating reranking into your workflow, you can significantly improve result quality.
mxbai-rerank-large-v2
mxbai-rerank-large-v2 is the flagship model in Mixedbread's second-generation rerank family, delivering state-of-the-art performance across 100+ languages. This reinforcement-learning enhanced 1.5B-parameter model excels at handling long contexts, complex query reasoning, and specialized use cases from code search to e-commerce, all while maintaining impressive processing speed.