VoyageAI by MongoDB: rerank-3-lite

VoyageAI by MongoDB: rerank-3-lite is a reranking model from Voyage AI, a provider whose models are now associated with MongoDB. It sits in the rerank category, which means its job is not to generate text but to take a set of candidate documents and reorder them by relevance to a query. It is positioned as a drop-in upgrade to rerank-2.5-lite, so teams already using that model can move to this one without reworking their pipeline. According to the provider, it is optimized for both latency and quality, improving on rerank-2.5-lite by 0.94% NDCG@10 on average across domain evaluations and by 1.86% on a further measure.

The model accepts text as its input and produces a reranked result as its output. Its context length is 32,000 tokens, which covers the query and the candidate passages you submit together in a single request. The maximum output is listed as 28,800. Because it is a reranker rather than a chat or completion model, it does not write prose or answer questions directly; it returns an ordering that reflects how well each candidate matches the query.

That shape of output makes it useful in retrieval pipelines. A common pattern is to run a fast first-stage search, such as a vector or keyword search, to pull a broad set of candidates, then pass those candidates to rerank-3-lite to sort them before they reach the user or the next stage of an application. This applies to search over documentation, support articles, product catalogs, or internal knowledge bases, and to retrieval augmented generation, where the quality of the top few passages strongly affects the final answer. The emphasis on latency makes it a reasonable fit for interactive search, where the reranking step has to complete quickly.

You can call it through api14's OpenAI-compatible API using the model id "voyageai/rerank-3-lite". Because the endpoint follows the OpenAI-compatible format, you can use existing OpenAI client libraries and point them at api14, which keeps integration work small if you already have such a client in place. The model lists no supported parameters, so there are no extra tuning options to set beyond the model id and the content of your request. Note that the provider release date is 2026-09-30.

If your task is text generation, summarization, or conversation rather than ranking a set of candidates, a chat or completion model will fit better, since rerank-3-lite only outputs an ordering. It is also worth checking that your candidate set and query fit within the 32,000 token context, and splitting or trimming if they do not.