How it works Rerank V3.5?
Rerank v3.5 is a reranking model for search engines: it reorders found fragments so that the most relevant ones appear at the top. It supports multi-aspect comparison and works with semi-structured data, while the initial output can be either semantic or keyword search. Via BotHub, the model connects via a unified OpenAI-compatible API alongside embeddings and generative models. Payment is pay-as-you-go in rubles without VPN or foreign cards, and data is protected by AES-GCM encryption. Typical scenarios: second-stage ranking in RAG, hybrid documentation search, sorting product cards with fields and attributes, and output in knowledge bases and support.Frequently asked questions about Rerank V3.5
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-v3.5 is a reranker, not a chat model. It receives a query and a set of documents and orders them by semantic proximity so that truly relevant fragments rise to the top. Use it in knowledge base search, RAG systems, and for refining output after initial candidate selection.
It does not write finished text: the model does not generate answers, but returns a relevance score for each document. The input context is up to 4096 tokens, meaning you can evaluate a 'query plus document' pair within these limits per call; long materials should be divided into fragments.
The reranker does not have explicit reasoning: this is not noted in our data. It does not conduct a dialogue or explain its thought process, but immediately provides a numerical relevance score and the order of documents. You build the decision logic yourself on top of the received ranking.
In our data, function calling and structured JSON output are not noted. In fact, the model returns an ordered list of documents with relevance scores — this is its result. You implement integration with external services on your side if needed.
No, according to our data, the acceptance of images, audio, and video is not noted. The model works with text: a query and text fragments are provided as input, and their order by relevance is provided as output. For multimodal scenarios, other models will be needed.
We do not have specific data on the quality of work with the Russian language, so we will not make promises. The reranker compares texts pairwise, and in practice, it is worth testing on your own dataset: run real queries and documents and see how well the output order matches your expectations.