How it works Rerank 2.5?
Rerank 2.5 is a reranking model: it selects and orders retrieved snippets, so search quality depends on more than just vector search. According to the developer, the model improves search accuracy by 7.94% across 93 datasets. In BotHub, it is available without a VPN or foreign card, payment is in rubles on a pay-as-you-go basis, and tokens do not expire. Via a unified OpenAI-compatible API, reranking integrates into the RAG pipeline immediately after Qdrant or pgvector search, with over 250 generation models available alongside it. Use cases: more accurate RAG bots for internal knowledge bases, relevant search results in online stores, document selection for corporate assistants, and reduced hallucinations in support.Frequently asked questions about Rerank 2.5
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-2.5 is a reranker: it takes your query and a list of candidates and reorders them by relevance, providing scores. It is suitable for knowledge base search, selecting snippets for RAG, sorting results, and filtering noise before sending to a generative model. Available via the unified BotHub API, without a VPN or foreign card.
This model does not write articles or essays: its response is an ordered list of candidates with relevance scores. Our data indicates a context of 32,000 tokens per request and an output limit of 28,800 tokens, but these are used for processing the request and ranking, not for coherent text.
There is no indication of reasoning before answering in our data, so we will not promise step-by-step reasoning. By design, the model immediately compares the query with candidates and outputs the order. If a chain of thought is needed, it is better to build it on a generative model and use the reranker to select snippets.
There is no confirmed indication of function calling or structured JSON output in our data. However, the reranker's result is already ordered: a list of candidates with scores, which is easy to parse programmatically. If you need a strict response according to your schema, check the format in the API documentation.
Acceptance of images, audio, and video is not indicated in our data. Assume the model works with text: query and text candidates — document snippets, articles, knowledge base answers. For images and video, choose the appropriate model type in BotHub; they are available in the same window.
There is no data on the quality of Russian language performance in the model description, so we will not make any claims. It is more practical to test it with your own materials: upload real 'question-snippet' pairs and compare the output order. This is easy to do in BotHub in one window and without a VPN.