Cohere makes Rerank 5 models that cut the text an LLM must read

Cohere made Rerank 5, 2 models that put search results in sequence before an LLM reads them. Rerank 5 Pro gives accurate results. Rerank 5 Fast is for low cost and high speed. Cohere says that Fast is more accurate than its Rerank 4 Pro model. Both models work with the Cohere API. Cohere gave no test numbers.

How it works

A reranker reads the text of each result at query time and gives it a score. The optional Context Focus tool also gives the parts of each document that are about the query. A team can send only those parts to the LLM and use less tokens.

The proof

Cohere gave no test results or numbers. Cohere says that Rerank 5 Pro gives more precision in finance, medicine and science. The models work in more than 100 languages and read 32k tokens at one time. They also read JSON, tables and YAML records.

What is not known

Cohere gave no number for the text that Context Focus cuts. It gave no number for the increase from Rerank 4. To change from Rerank 4, a team changes the model name. re-indexing is not necessary. Rerank 5 is on the Cohere API, Microsoft Foundry and Amazon SageMaker.

Insight: The reranker decides what the LLM reads

Rerank 5 can cut the text that goes to an LLM. vLLM Semantic Router released small decision models. RaReCache uses a small model to prepare a large one. This reads as a fight to control the small model in front of the large one, because it sets the cost and the text of each LLM task. Cohere is on 2 large clouds. A team can test it with small risk.