Cohere: Rerank v3.5

Cohere Rerank 3.5 is a multilingual model for relevance ranking of candidate documents and semi-structured text. It evaluates each candidate in relation to a query and returns a reordered result list with relevance scores. It can refine candidates obtained through keyword or semantic search and select passages for a retrieval-augmented answer. Its task is to improve the evidence ordering rather than generate a free-form answer or create a new corpus index.

The exact producer identifier is rerank-v3.5. It uses a 4096-token context, distinct from Rerank 4 Pro and Fast’s 32K context. Query and document tokens count together against that per-candidate window. The model-specific preparation guide truncates a query beyond its first 2048 tokens and describes document chunks of 4093 tokens before the query is taken into account. Keep the query focused and prepare candidate passages containing the information needed to judge the requested relationship.

Supply the query and a list of candidate text strings. Each result carries an index pointing to the original document position and a relevance score; top_n chooses how many ranked results to retain. Without it, the documented producer workflow returns all results. Preserve document identifiers independently of the score so the top passage can still be traced to the original policy, product record or message.

Rerank 3.5 supports more than 100 languages, including English and non-English material. A query can contain several aspects of the need, such as the product, user situation and desired procedure. For structured records, the V2 release removed raw document objects and recommends YAML strings. Keep useful fields in a deliberate order, with decisive content before less relevant metadata, because long strings can be truncated. This works for titles and content, product attributes or a message subject and body.

The V2 producer reference uses max_tokens_per_doc for document truncation, with a default of 4096. Cohere’s preparation guide also describes long-document chunking; the effective input depends on the selected workflow and its document settings. For greater control, prepare identifiable passages yourself rather than assuming every section of a long source remains visible. Preserve enough surrounding context to distinguish an answer-bearing passage from one that only shares its vocabulary.

For request performance, the current reference recommends not sending more than 1,000 documents at once. Relevance scores are normalised between 0 and 1, but their values depend on the query and candidates and are not a linear scale of correctness. Use the ranking to identify useful evidence. If a score threshold is needed, choose representative domain queries and borderline relevant passages so the threshold reflects the actual retrieval task.

A practical support workflow can rerank candidate return-policy sections for a question about a damaged item, then pass the selected source text to an answer-writing stage. An email search can rank messages by a specific issue and retain the original message identifier with each result. The useful artifact is the ordered evidence list. Keep Rerank 3.5’s shorter context and V2 document preparation attached to that result instead of substituting later-model limits.