Cohere Reranker

Use Cohere’s specialized reranking models

Overview

Cohere Reranker uses Cohere’s specialized models for high-quality document reranking. It provides state-of-the-art relevance scoring.

Setup

Install Cohere

Cohere is included with Mini RAG:

uv add mini-rag

Get API Key

1. Sign up at Cohere
2. Get your API key from the dashboard
3. Add to your environment:

COHERE_API_KEY=your-cohere-api-key

Configuration

Basic Usage

from mini import AgenticRAG, RerankerConfig
import os

rag = AgenticRAG(
    vector_store=vector_store,
    embedding_model=embedding_model,
    reranker_config=RerankerConfig(
        type="cohere",
        kwargs={
            "api_key": os.getenv("COHERE_API_KEY")
        }
    )
)

Complete Configuration

reranker_config = RerankerConfig(
    type="cohere",
    kwargs={
        "api_key": os.getenv("COHERE_API_KEY"),
        "model": "rerank-english-v3.0",
        "max_chunks_per_doc": None
    }
)

Available Models

rerank-english-v3.0

Best for: English text
Quality: Highest
Speed: Fast
Recommended for: Production English applications

rerank-multilingual-v3.0

Best for: Multiple languages
Quality: High
Speed: Fast
Recommended for: International applications

Supports: English, French, Spanish, German, Italian, Portuguese, Chinese, Japanese, Korean, Arabic, and more

rerank-english-v2.0

Best for: Legacy applications
Quality: Good
Speed: Fast
Recommended for: Existing integrations

Direct Usage

Use the Cohere reranker directly:

from mini.reranker import CohereReranker

# Initialize
reranker = CohereReranker(
    api_key=os.getenv("COHERE_API_KEY"),
    model="rerank-english-v3.0"
)

# Rerank documents
query = "What is machine learning?"
documents = [
    "Machine learning is a subset of AI...",
    "Python is a programming language...",
    "Deep learning uses neural networks..."
]

results = reranker.rerank(query, documents, top_k=2)

for result in results:
    print(f"Score: {result.score:.3f}")
    print(f"Document: {result.document[:100]}...")

Complete Example

import os
from mini import (
    AgenticRAG,
    LLMConfig,
    RetrievalConfig,
    RerankerConfig,
    EmbeddingModel,
    VectorStore
)
from dotenv import load_dotenv

load_dotenv()

# Initialize RAG with Cohere reranking
rag = AgenticRAG(
    vector_store=VectorStore(
        uri=os.getenv("MILVUS_URI"),
        token=os.getenv("MILVUS_TOKEN"),
        collection_name="documents",
        dimension=1536
    ),
    embedding_model=EmbeddingModel(),
    llm_config=LLMConfig(model="gpt-4o-mini"),
    retrieval_config=RetrievalConfig(
        top_k=10,
        rerank_top_k=3,
        use_reranking=True
    ),
    reranker_config=RerankerConfig(
        type="cohere",
        kwargs={
            "api_key": os.getenv("COHERE_API_KEY"),
            "model": "rerank-english-v3.0"
        }
    )
)

# Index and query
rag.index_document("document.pdf")
response = rag.query("What is the main topic?")

print(response.answer)

Pricing

Cohere reranking pricing (as of 2024):

• Search Units: Charged per 1000 search units
• Search Unit: 1 query + 1 document to rerank
• Example: 1 query with 10 documents = 10 search units

Typical costs:

• Free tier: Limited searches
• Paid: ~$1-2 per 1000 queries (10 docs each)

Check Cohere Pricing for current rates.

Best Practices

Choose the Right Model

• English only: Use rerank-english-v3.0
• Multiple languages: Use rerank-multilingual-v3.0
• Best quality: Always use v3.0 models

Optimize top_k

Retrieve more initially, rerank to fewer:

RetrievalConfig(
    top_k=15,        # Cast wide net
    rerank_top_k=3   # Keep only best
)

Handle Errors

try:
    response = rag.query(question)
except Exception as e:
    if "rate_limit" in str(e).lower():
        # Handle rate limit
        time.sleep(1)
        response = rag.query(question)
    else:
        raise

Monitor Usage

Track API usage in Cohere dashboard to manage costs

Advantages

✅ Highest Quality: Best-in-class reranking
✅ Fast: Low latency (~50-100ms)
✅ Easy Setup: Simple API integration
✅ Multilingual: Supports many languages
✅ Maintained: Continuously improved by Cohere

Limitations

❌ Cost: Requires API subscription
❌ Cloud Only: Not available for local deployment
❌ API Dependency: Requires internet connection
❌ Rate Limits: Subject to API rate limits

Troubleshooting

API Key Error

Ensure your API key is set:

import os
print(os.getenv("COHERE_API_KEY"))

Rate Limit

Implement backoff:

from time import sleep

for attempt in range(3):
    try:
        results = reranker.rerank(query, docs)
        break
    except Exception as e:
        if attempt < 2:
            sleep(2 ** attempt)

Model Not Found

Use correct model names:

• rerank-english-v3.0 ✅
• rerank-multilingual-v3.0 ✅
• rerank-v3 ❌ (incorrect)

See Also

Rerankers Overview

Compare rerankers

Reranking Feature

Learn about reranking

Cohere Docs

Official Cohere docs