Nearly a year after releasing ReRank 3.5, Coher launched the latest version of its search model, now with a larger context window to help agents find the information they need to complete their tasks.
Kohir said A blog post This rerank 4 has a 32K context window, representing a fourfold increase over 3.5.
According to the blog post, “This enables the model to handle long documents, evaluate multiple parts simultaneously, and capture relationships among parts that are lacking in smaller Windows objects.” “This extension capability, therefore, improves classification accuracy for realistic document types and increases confidence in the relevance of retrieved results.”
ReRank 4 comes in two flavors: Fast and Pro. As a smaller model, Fast is best suited for use cases that require both speed and accuracy, such as e-commerce, programming, and customer service. Pro is optimized for tasks that require deep reasoning, precision and analysis, such as developing risk models and analyzing data.
Enterprise search gained more importance this year, especially as AI agents must access more information and context about the organization they work for. Rankers “significantly increase the accuracy of enterprise AI searches by improving initial retrieval results,” Koher said. ReRank 4 addresses the neighborhood differences created by some two-encoder embeddings—models that help simplify retrieval augmentation generation (RAG) tasks—by using a cross-encoder architecture “that jointly processes queries and candidates, processes subtle term relationships, and reorders the level of most relevant items.”
Performance and benchmarks
The models were benchmarked against other reranking models, such as Kevin Rinker 8B, Gina ReRank V3, and MungoDB’s 2.5 travel series, in tasks from the finance, healthcare, and manufacturing domains. ReRank 4 outperformed its competitors strongly, if not its rivals.
ReRank 3.5 stood out because of its ability to support multiple languages, and Koher said ReRank 4 continues that trend. It understands more than 100 languages, including advanced retrieval in 10 major business languages.
Agents and iterative models
ReRank 4 aims to help agent tasks understand which data is best suited for their tasks and provide more context.
Coheer noted that this model is a key component of its agent AI platform, Northern, as it integrates seamlessly with existing AI search solutions, including hybrid, vector and keyword-based systems, with minimal code changes. “
As more enterprises look to use agents for research and insights, as evidenced by the rise of deep research features, models that help filter out irrelevant content, such as rerankers, become more necessary.
“This is particularly effective for agent AI, where complex, multi-factor interaction model calls can run quickly and saturate the context,” Kuher said.
The company argues that ReRank4 helps reduce token usage and by preventing low-quality information from reaching the LLM, an agent’s effort is needed to get it right.
Self learning
ReRank 4 stands out not only for its robustness capabilities, but also for offering a first-of-its-kind self-learning model, Kuhir said.
Users can customize ReRank 4 for use cases they encounter more frequently without additional annotated data. Similar to foundation models like GPT-5.2, where people can specify preferences and the model remembers them, ReRank 4 users can tell the model their preferred content types and document corpora.
If used with ReRank 4, for example, the model becomes more competitive with larger models because it is more precise and taps the specific data users want.
“Exploring further, we also explored how 4’s self-learning ability performs on entirely new search domains,” Kuher said. “Using healthcare-based datasets that simulate the specific information retrieval needs of a physician—one who doesn’t just specialize in a given medical discipline—we’ve found that self-learning enables consistent, substantial gains. The result: a clear and significant improvement in retrieval quality for ReRank 4 across the board.”