Unlocking the State-of-the-Art Reranker: Introducing the Vectara Multilingual Reranker_v1
In the ever-evolving landscape of RAG and information retrieval, a balance of precision, recall, and latency will make or break the applications. We are excited to introduce our latest innovation, the Multilingual Reranker_v1. This state-of-the-art reranker, significantly enhances the precision of retrieved results across both English and multilingual datasets.
≈ 6 minutes read
Introducing the Vectara Multilingual Reranker_v1
In the complex world of retrieval-augmented generation (RAG) pipelines, the reranker plays a crucial role. After the retrieval model generates a comprehensive list of high-recall results, the reranker steps in to refine these results, boosting the precision of the top results and ensuring more relevant outcomes. For a deeper dive into the importance of reranking, check out our detailed guide on reranking and why it matters.
We are thrilled to announce the launch of our latest reranking model, the Vectara Multilingual Reranker_v1. This cutting-edge reranker excels in handling multilingual datasets.
For English datasets, it delivers an impressive ~10% uplift in Normalized Discounted Cumulative Gain (NDCG). Even more remarkable is its performance on multilingual datasets, where it achieves a ~30% improvement. These advancements underscore the model’s robustness and adaptability in diverse linguistic environments. For an in-depth technical exploration, visit Deep Dive into Multilingual Reranker v1, State-of-the-Art Reranker supporting 100+ Languages.
Benchmarking Against the Best
To validate our model’s performance, we conducted extensive benchmarking against some of the most popular reranker models available today, including Cohere Rerank 3, BGE Reranker Base, BCE Reranker Base, and Mono MT5. Our Multilingual Reranker_v1, emerged as the top performer for nearly all English-based datasets and in the top 2 for nearly all multilingual datasets. These results highlight our model’s superior precision and effectiveness. For detailed benchmark results and analysis, visit our blog post.
Trade-offs
When reranking the top 25 rows of results, the new reranker currently exhibits a latency of approximately 100ms. It’s essential to balance the trade-off between increased precision and added latency. The key parameter you can adjust is the number of results to be reranked. The more results you rerank, the higher the precision as well as latency. We recommend starting with the top 25 results as a default setting for optimal performance.
Getting Started with Vectara Multilingual Reranker_v1
The Multilingual Reranker is available exclusively as a Scale-only feature. To use it, you must be either a Scale-trial customer or a Scale customer. For information on how to start a Scale trial, please contact our sales team.
Once you are a Scale-trial or Scale customer, you can access the new reranker through both the UI and API.
To use the UI, log in to your Vectara console and select the relevant data corpus. Navigate to the “query” tab associated with your data corpus and locate the “retrieval” configurations. Within the “reranking” section, enable the option to “rerank search results” and select “rerank_multilingual_v1” from the reranker selection dropdown. This will activate the new reranker for your queries.
For API access to the new Vectara Multilingual Reranker: To use the Vectara Multilingual Reranker in your queries, set the rerankerID to 272725719.
More details can be found in API documentation.
Sign Up for an Account and Connect With Us!
As always, we’d love to hear your feedback! Connect with us on our forums or on our Discord.
Sign up for a free trial to see how Vectara can help you easily leverage retrieval-augmented generation in your GenAI apps.
