Rerank-v3.5, Explained Simply

Cohere just released a new AI model designed to sort information. If you have ever used a search engine and wondered how it picks the *best* answer out of millions of web pages, this model does exactly that. It is built for app developers and businesses who need to quickly find the most relevant documents for their users' questions.

What it does well

This model is a "reranker," which means its main job is to take a list of possible search results and put the best ones at the very top. It is especially good at understanding different languages and has strong "reasoning capabilities," meaning it is better at thinking through what a question is actually asking. It performs at "SOTA" (State of the Art, a tech term meaning the best currently available) for general searches, as well as for specific areas like finance, online shopping, hotel info, project management, and email searches.

How it compares

Compared to Cohere’s older models, this version is their most powerful yet for ranking information. The company also updated the API (the set of rules programmers use to talk to the AI). Now, programmers must specifically name the model they want to use. Also, instead of breaking documents into random chunks, the system now looks at "tokens" (small pieces of words) to decide how much text to read at once.

What it costs and what it can handle

The model has a "context length" of 4,096. In plain English, that means it can read and rank about 4,096 tokens at a time (a token is roughly three-quarters of a standard word, so this is roughly 3,000 words). Programmers can adjust this limit using a new setting called max_tokens_per_doc, which automatically cuts off text that is too long.

Worth knowing

There are a few catches for developers using this. The new API no longer lets programmers send complicated, structured data (like lists or dictionaries) directly to the model. If your data is structured, the creators recommend turning it into "YAML strings"—which is just a simple, plain-text way of formatting information so the AI can read it easily.