VoyageAI by MongoDB: rerank-2.5-lite

VoyageAI by MongoDB: rerank-2.5-lite is a reranking model from Voyage AI, a provider whose reranking family is now part of MongoDB. It sits in the rerank category, so its job is not to generate text but to take a set of candidate documents and reorder them by relevance to a query. The provider released it on 2026-07-27. If you are building a retrieval pipeline and already have a first pass of results from a vector or keyword search, this model is the second stage that decides which of those results actually belong at the top.

The model accepts text input and produces a reranking output. It has a context length of 32000 tokens and a maximum output of 28800. Only text is supported on the input side, so there is no image or audio handling here. The provider states that rerank-2.5-lite is optimized for both latency and quality, and that it delivers a 7.16% improvement in retrieval accuracy over Cohere Rerank v3.5 across 93 datasets. The provider also states that it outperformed Cohere Rerank v3.5, which is the comparison point Voyage AI chose to publish.

Those capabilities point to a fairly specific set of uses. The obvious one is search: take the top 50 or 100 hits from a first stage retriever, pass them through this model with the user query, and keep only the highest scoring few for the answer or the next step. The same pattern applies to retrieval augmented generation, where a reranker reduces the amount of irrelevant context sent to a generator model. It also fits document ranking tasks, such as ordering internal knowledge base articles or support tickets against a natural language query, and any pipeline where precision at the top of a list matters more than raw recall.

You call it through api14's OpenAI-compatible API using the model id "voyageai/rerank-2.5-lite". Because the endpoint follows the OpenAI schema, you can point an existing OpenAI client at api14 and send a request in the usual shape, with the model id set to that string. The model's supported parameters list is empty, so there are no extra tuning knobs to set beyond what the rerank request itself requires. Plan for the request to carry the query and the documents to be scored, and expect a reranking result rather than generated prose.

One thing to weigh is that this is a reranker, not a chat or completion model. If you need an assistant that writes answers, summarizes, or calls tools, a generative model is the right choice, and this model would only serve as a supporting stage in front of it. It is also text only, so multimodal retrieval is out of scope. Within those bounds, rerank-2.5-lite is aimed at teams that want a low latency second stage reranker with the accuracy claims Voyage AI publishes.