How it works Rerank 3?
Rerank 3 is a reranking model for the second stage of search: you provide a query and found candidates, and receive an order where relevant items are higher. According to the vendor, it is a drop-in replacement for rerank-2.5 with an average gain of 0.80% NDCG@10 across domain evaluations and 3.35% on long documents, so there is no need to rewrite the integration. In BotHub, the model is connected to a unified OpenAI-compatible API along with over 250 other models, available without VPN or foreign cards, with payment in rubles upon usage. Scenarios: RAG assistants for internal knowledge bases, support and website search, selecting answers from long contracts, and e-commerce search results. For teams, we offer contracts, invoices, electronic document management (EDO), and limits in the admin panel.Frequently asked questions about Rerank 3
You can use generated results for commercial purposes. You own all rights to the content you create. The only restriction: make sure your prompt does not include copyrighted third-party material. You are responsible for respecting the rights to any input data.
rerank-3 is a ranking model from MongoDB: it matches a query with documents and reorders the output by relevance. It is used as a second stage on top of standard search, in RAG systems, and when selecting snippets, so that the truly relevant items appear at the top. Context — up to 32,000 tokens.
It does not write ready-made text: at the output, you receive relevance scores and an ordered list of snippets. The input context is up to 32,000 tokens, and the output limit in our data is specified as up to 28,800 tokens, but it is consumed by the ranking result, not by coherent text.
There is no indication of reasoning before answering in our data, and for its task, the model does not need it: it does not think aloud, but immediately compares the query with documents and sets the order of relevance. If you need step-by-step explanations, it is better to get them from a generative model.
There is no confirmed indication of function calling and structured output in our data, so we will not claim it. Practically, the model returns not arbitrary text, but a ranking result — the order of snippets and relevance scores, which you receive via API. Check the exact response format in the documentation.
There is no indication of accepting anything other than text as input in our data. The model works with text: the query and documents are passed as text, and it evaluates how well they correspond to each other. For images, audio, and video, a separate model of a different class will be needed.
There is no quality assessment for Russian language performance in our data, so we will not make specific promises. The absence of an indication does not mean the language is not supported: run the model on your real 'query-snippet' pairs and compare the output order with the expected one before putting it into production.