How it works Rerank 3 Lite?
Rerank 3 Lite is a compact reranker optimized for low latency and quality. Compared to rerank-2.5-lite, it provides an average 0.94% NDCG@10 gain on domain benchmarks and connects as a drop-in replacement without rewriting your pipeline. It acts as the second stage of search: it selects candidates via vectors, then refines their order by relevance to the query. BotHub offers a unified OpenAI-compatible API without VPN or foreign cards, pay-as-you-go in rubles, encryption, and over 250 models in one window. For teams, we offer contracts, invoices, EDI, and limits. Useful for RAG on Qdrant and pgvector, knowledge base search, product and document matching, Q&A on regulations, and filtering irrelevant snippets.Frequently asked questions about Rerank 3 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 evaluates how well each text snippet matches the query and reorders results by relevance. Use it on top of search or RAG to boost accurate answers and remove random matches before sending them to a generative model.
It does not write coherent text; the output is ranking and relevance scores. In one request, the model accepts up to 32,000 tokens of context including documents; the output limit is 28,800 tokens, which are used for the service response, not for composition.
This capability is not noted in our data. The reranker solves a narrow task—comparing a query with a snippet and assigning a relevance score—so long chains of reasoning are not expected. For step-by-step reasoning, use a generative model in BotHub.
Function calling and structured JSON output are not noted in our data. The model's task is to return relevance scores for a list of snippets, and a standard HTTP request from your search pipeline is usually sufficient for integration.
Acceptance of images, audio, and video is not noted in our data; the model works with text: queries and document snippets. You cannot send screenshots or recordings to it, but you can send text via BotHub without a VPN or foreign card.
There is no specific information about Russian language performance in the model data, so I cannot promise quality. The reranker compares the meaning of the query and the snippet, and it is reasonable to test it on your own documents: run a dozen queries and see how the results are ordered.