How it works Rerank 2.5 Lite?
Rerank 2.5 Lite is a lightweight Cohere reranker that re-ranks retrieved chunks by relevance to the query. According to the vendor, it provides a 7.16% increase in search accuracy compared to Cohere Rerank v3.5 across 93 datasets, outperforming the previous version while maintaining fast response times. In BotHub, the model is accessible from Russia without a VPN or foreign card, with payment in rubles. Over 250 models and a unified OpenAI-compatible API are available in one window, making it easy to integrate the reranker into existing workflows. Use cases: enriching RAG pipelines with a second ranking stage, searching knowledge bases and documentation, searching support tickets, and refining output in chatbots and agents.Frequently asked questions about Rerank 2.5 Lite
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.
It is a reranker: it receives a query and a set of retrieved chunks, assigning a relevance score to each so the most suitable ones appear first. Useful for knowledge base search, RAG systems, and document retrieval where precise sorting of retrieved results is important. Context window: up to 32,000 tokens.
The model does not write its own text: its output consists of relevance scores for the provided chunks, not coherent prose. Therefore, the question 'how much text per response' does not apply. It accepts a query and documents totaling up to 32,000 tokens as input.
Our data does not indicate reasoning capabilities before answering. By design, a reranker does not do this: it immediately compares the query with the chunks and provides a score. If your task requires step-by-step reasoning, use a separate chat model via the same interface.
We have no confirmation of this: there are no markers for function calling or structured JSON output for this model. In practice, the reranker returns a list of chunks with scores, and you define the wrapper format and integration on your side via the API.
Our data does not indicate support for images, audio, or video: the model works with text—queries and document chunks. If you need to process images or audio, check out the multimodal models in the BotHub catalog; switching takes just a couple of clicks.
There is no specific note on Russian language performance in our data, so we cannot make promises. Such models are typically trained on multilingual data, but it is best to test with your own queries and documents. Access to the model is available without a VPN or foreign card.