Cohere Rerank API for Production RAG

Why Cohere Rerank?

• ✅ 40% accuracy improvement over bi-encoders
• ✅ 100+ languages supported
• ✅ Hosted API (no model hosting)
• ✅ Fast (< 200ms for 100 docs)

Quick Start

DEVELOPERpython
import cohere

co = cohere.Client('YOUR_API_KEY')

def rerank_with_cohere(query, documents):
    results = co.rerank(
        model='rerank-v3.5',  # unified multilingual model
        query=query,
        documents=documents,
        top_n=10
    )
    
    return [doc['text'] for doc in results.results]

# Use it
retrieved_docs = vector_search(query, k=100)
reranked = rerank_with_cohere(query, retrieved_docs)

Models (November 2025)

rerank-v3.5 (data: July 2026)

• Unified multilingual (replaces the separate v3.0 English/multilingual models)
• 100+ languages
• $1 per 1000 searches
• 4k token context

rerank-v4.0-fast / rerank-v4.0-pro

• Newest generation, 100+ languages
• Larger 32k token context
• $2 / $2.50 per 1000 searches

The rerank-english-v3.0 and rerank-multilingual-v3.0 models are still available but are now superseded by the unified rerank-v3.5 multilingual model.

With Metadata

DEVELOPERpython
results = co.rerank(
    query=query,
    documents=[
        {"text": doc, "metadata": {"source": "wiki", "date": "2025"}}
        for doc in documents
    ],
    top_n=10,
    return_documents=True
)

for r in results.results:
    print(f"Score: {r.relevance_score}")
    print(f"Text: {r.document['text']}")
    print(f"Metadata: {r.document['metadata']}")

Cost Optimization

DEVELOPERpython
# Only rerank if initial score is low
def smart_rerank(query, initial_results, threshold=0.7):
    # If top result has high confidence, skip reranking
    if initial_results[0]['score'] > threshold:
        return initial_results[:10]
    
    # Otherwise, rerank
    return rerank_with_cohere(query, [r['text'] for r in initial_results])

Cohere Rerank is the easiest way to dramatically improve RAG accuracy. Just plug it in after retrieval.