Vectara Multilingual Reranker

Generative AI applications often struggle with ranking the most relevant information, leading to hallucinations and irrelevant responses. The new Vectara Multilingual Reranker V1, also known as Slingshot, is a state-of-the-art reranking model that significantly enhances the precision of retrieved results. Providing advanced neural ranking, it refines the output of initial models like Boomerang, offering even more accurate document scoring and response quality in Retrieval Augmented Generation (RAG) pipelines.

The Vectara Multilingual Reranker operates as a second-pass refinement tool, building on Boomerang’s high-recall capabilities. While Boomerang quickly retrieves a broad set of relevant documents, the Multilingual Reranker delivers more precise results, ensuring that the top-ranked documents are the most relevant. This reranker also excels across both English and multilingual datasets, making it a powerful tool for global use cases.

While more computationally expensive and introducing some additional latency, the multilingual reranker improves neural ranking beyond Boomerang’s initial selection by providing more precise text scoring. Think of the Slingshot reranker as a "better Boomerang" for refining results, with the multilingual capability serving primarily as a differentiator from other rerankers in the market, which are often English-only.

Using this reranker requires both the type and reranker_name in the reranker object. Set the type as customer_reranker and the reranker_name as Rerank_Multilingual_v1.

"reranker": {
      "type": "customer_reranker",
      "reranker_name": "Rerank_Multilingual_v1"
    }

note

The reranker_id and rnk_272725719 have been deprecated. Use reranker_name and Rerank_Multilingual_v1.

The Vectara Multilingual Reranker ensures impressive zero-shot performance on unseen data and domains, and it never trains on customer data. In RAG use cases, this reranker distinguishes the scores of relevant and irrelevant documents in a query-independent manner. For more details about our Multilingual Reranker v1, check out these feature announcement and technical deep dive blogs.

Based on our experimentation we suggest using a cut-off threshold of 0.5 as a good starting point. This threshold value is the relevance score returned by Vectara with each responseAny results that achieve a score of greater than or equal to 0.5 can be considered relevant and anything below that can be considered as non-relevant.