Cohere: Rerank 4 Pro

Cohere Rerank 4 Pro is the quality-focused model in Cohere’s Rerank 4 family. It evaluates candidate documents against a query and returns their relevance ordering and scores. The model is intended for complex retrieval tasks, including multilingual documents and semi-structured information. It fits a search or retrieval-augmented workflow after an existing system has collected candidates: the result helps choose which passages should be shown or supplied as evidence to the next stage.

The producer identifier is rerank-v4.0-pro, separate from rerank-v4.0-fast and rerank-v3.5. Pro is positioned for complex cases where retrieval quality is the priority, while Fast targets lower latency and higher throughput. Both Rerank 4 editions use a 32,768-token context. The query and one candidate document share that budget; it is not a separate full window for every input field or a limit on the entire stored search corpus.

Supply a clear query and the list of candidate text strings. Each returned result includes the original document index and a relevance score, and top_n can select how many ranked results are returned; without it, the producer returns all results. Preserve a source identifier beside each input so ranking can be connected back to the exact passage. For a support question, ask about the required conditions and supply the candidate policy sections rather than expecting the reranker to locate policies that were not in the list.

The model supports more than 100 languages and can rerank English and non-English material. For a query with several requirements, state the conditions that make a passage useful, such as the product, situation and needed procedure. For semi-structured records, Cohere recommends YAML strings rather than passing raw objects in the V2 document list. Keep meaningful fields such as title and content in a deliberate order, because truncation can remove later fields. This preparation is useful for product records, messages and document metadata as well as ordinary prose.

The model-specific preparation guide limits the Rerank 4 query to the first 16,384 tokens. It describes longer-document processing through chunks that share the 32,768-token context with the query and reserved tokens. The current V2 reference separately defaults max_tokens_per_doc to 4096 and says longer document strings are truncated to that setting. Treat the model window and the producer request’s document truncation as different controls. Prepare concise relevant passages or choose an appropriate document budget instead of assuming every long source is always read in full.

For recommended request performance, Cohere advises against sending more than 1,000 documents at once. Preserve candidate identifiers when splitting a larger list into manageable groups. Relevance scores are normalised between 0 and 1, but they depend on the query and candidates; they are not a linear measure of factual correctness. Use the ordering as the primary retrieval result. If filtering by a score threshold, establish it with representative queries and borderline relevant passages from the actual domain rather than assuming a universal cutoff.

A useful final artifact is a ranked candidate list carrying source identifiers and scores, ready for a search result or an evidence-selection step. A document with a high relevance score can still contain outdated or incorrect information, so retain its source and content for the stage that uses it. Pro’s purpose is relevance judgement, not writing a free-form answer or generating embeddings. Keep the exact Pro edition attached to the result when comparing or refining a retrieval workflow.