Cohere: Rerank 4 Fast

Cohere Rerank 4 Fast is the lighter Rerank 4 edition positioned for lower-latency and higher-throughput retrieval tasks. It takes a query and candidate documents, then returns their relevance ordering and scores. It is useful for interactive search, repeated evidence selection and applications that need to reorder an existing result list. Fast remains a distinct edition from Rerank 4 Pro, which is positioned for more complex cases where retrieval quality takes priority.

The producer code is rerank-v4.0-fast. Like Pro, it has a 32,768-token context and multilingual coverage across more than 100 languages. The query and one document share the context; it is not a vector dimension or a budget for an entire indexed corpus. Supply the candidate texts that should be compared rather than expecting the model to retrieve documents from a source it has not been given.

The documented result includes the original document index and a relevance_score. top_n selects how many results are returned, while omitting it returns the full ranked list. Keep source identifiers beside the inputs so the new ordering can be joined to the actual passages. For a product search, describe the required product attributes in the query and rerank the candidate descriptions; a higher rank should remain connected to the product record that produced it.

The model can work with English and non-English documents and semi-structured records. In the V2 workflow, document inputs are strings, and Cohere recommends representing structured data as YAML. Put important fields first and preserve their order so a truncation does not discard the content needed to judge relevance. A support record can carry title, product and procedure text; a message can carry its subject and body. These explicit relationships give the ranker information beyond a list of matching keywords.

The model-specific query limit in Cohere’s preparation guide is 16,384 tokens, with the query and document consuming the shared context. The guide describes longer-document chunk processing, while the V2 reference separately defaults document truncation to 4096 tokens through max_tokens_per_doc. Check that effective document setting rather than assuming the whole 32K model window is always used. Prepare relevant passages or a suitable document budget for the actual retrieval task.

For recommended performance, Cohere advises against more than 1,000 documents in one request. Split larger candidate sets into manageable groups while retaining their identifiers. The output scores are normalised from 0 to 1 and depend on the query and passages. Their ordering is the key result; a score is not a linear multiple of relevance or proof that a passage is correct. Establish any filtering threshold with representative queries and borderline relevant passages from the domain.

A useful Fast result is a ranked list that can immediately drive a search display or select evidence for an answer-writing stage. Preserve the source text with the score and review whether the top results actually address the requested conditions. Fast’s role is retrieval ranking rather than generating a new explanation or an embedding vector. Keep the named edition and input preparation consistent when refining the relevance behaviour.