How it works Rerank 4 Fast?
Rerank 4 Fast is a Cohere model for reranking search and RAG results: it prioritizes documents that truly answer the query. The 32K context is suitable for long snippets, support for over 100 languages works for multilingual catalogs, and no data preprocessing is required. In BotHub, the model is available in a single window alongside generative models, connects via an OpenAI-compatible API, is accessible without a VPN or foreign card with payment in rubles upon usage, and ensures privacy with AES-GCM encryption and no data retention. Use it to refine internal search results, select context for RAG assistants, sort catalogs, and filter results in agentic scenarios.Frequently asked questions about Rerank 4 Fast
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-4-fast is a ranker from Cohere: it receives a query and a set of documents and sorts them by semantic proximity to the query. It is useful for knowledge base search, RAG systems, product recommendations, and filtering results when you need to quickly select truly relevant snippets before passing them to another model.
rerank-4-fast does not write text as such: its response is an ordered list of documents with relevance scores. Our data indicates a context of up to 32,768 tokens for input and an output limit of 29,491 tokens, so the volume of the processed query and document set should be kept within these limits.
There is no indication of reasoning before answering in our data. rerank-4-fast belongs to the class of rankers: its task is to evaluate how well a document matches a query and provide an order, not to conduct a dialogue or explain its thought process. If step-by-step analysis is needed, it is better to assign it to a separate text model.
Function calling and structured JSON output are not indicated in our data. However, the ranking result itself is simple: it is a list of documents with relevance scores, which is easy to parse in code and pass along your pipeline.
The acceptance of images, audio, and video is not indicated in our data — the model is designed to work with text. It takes a text query and text documents or snippets as input, so images and recordings should be converted to text beforehand and then sent for ranking.
There are no separate quality measurements for Russian in our data. The ranker compares the meaning of the query and the document, so it can work with Russian text, but how confidently depends on the training data. Test this on your own sample: run typical queries and check the output order.