VoyageAI by MongoDB: rerank-3

VoyageAI by MongoDB: rerank-3 is a reranking model from Voyage AI, a provider now part of MongoDB. It sits in the rerank category, and Voyage AI positions it as an optimized reranker and a drop-in upgrade to rerank-2.5. This is not a chat or text generation model: it exists to take a set of candidate documents and reorder them by relevance to a query.

The model works with text input and produces reranking output. It has a context length of 32,000 tokens, which is large enough to accommodate a query alongside a substantial batch of candidate passages in a single call. The maximum output is 28,800. Because it is a reranker, the shape of a response is a ranked ordering of the documents you supply, not free-form prose. There are no supported parameters listed for this model, so callers should not expect the sampling controls (temperature, top_p and similar) that apply to generative models.

The practical use cases follow directly from that design. A common pattern is a two-stage retrieval pipeline: a fast vector or keyword search returns a broad set of candidates, then rerank-3 reorders them so the most relevant results appear at the top. This improves the quality of search results, retrieval-augmented generation context, question answering over a document collection, and recommendation or matching features where relevance ranking matters. Because the provider describes gains on long-document evaluations, it is also a reasonable fit when the passages being ranked are lengthy rather than short snippets.

To call it through api14, use the model id "voyageai/rerank-3" against the OpenAI-compatible endpoint. Because api14 exposes an OpenAI-compatible API, you point your existing client at the gateway and pass that id as the model. Since the model lists no supported parameters, you should send a plain request with your query and the documents to rank, and rely on the model's own ranking behaviour rather than tuning decoding settings.

One limit worth noting is scope. rerank-3 only reranks text; it does not generate answers or summaries, so if you need a model that produces text you would pair it with a generative model. Also, if your pipeline has no separate retrieval step and you only need a first-pass embedding search, a dedicated embedding model may be the simpler starting point, with rerank-3 added later to refine the ordering.