Reranking Role

The reranking role is specialized for reordering search results, improving retrieval system accuracy and relevance.

Configuration

Configure in config.yaml or ~/.bytebuddy/config.yaml:

models:
  - name: "rerank-model"
    provider: "cohere"
    model: "rerank-english-v3.0"
    apiKey: "${COHERE_API_KEY}"
    roles: ["rerank"]

Core Features

Result Reordering

• Relevance Scoring: Calculate query-document relevance
• Intelligent Sorting: Reorder based on semantic understanding
• Precise Matching: Improve ranking of most relevant results
• Quality Optimization: Enhance search result quality

Application Scenarios

• Semantic Search: Improve search accuracy
• RAG Systems: Optimize retrieval quality
• Recommendation Systems: Improve recommendation ranking
• Q&A Systems: Enhance answer relevance

Reranking Model Configurations

Cohere Reranking

models:
  - name: "cohere-rerank"
    provider: "cohere"
    model: "rerank-english-v3.0"
    apiKey: "${COHERE_API_KEY}"
    roles: ["rerank"]

Multilingual Reranking

models:
  - name: "multilingual-rerank"
    provider: "cohere"
    model: "rerank-multilingual-v3.0"
    apiKey: "${COHERE_API_KEY}"
    roles: ["rerank"]

Specialized Reranking

models:
  - name: "specialized-rerank"
    provider: "cohere"
    model: "rerank-english-v3.0"
    apiKey: "${COHERE_API_KEY}"
    roles: ["rerank"]

Use Case Configurations

Code Search Reranking

models:
  - name: "code-rerank"
    provider: "cohere"
    model: "rerank-english-v3.0"
    apiKey: "${COHERE_API_KEY}"
    roles: ["rerank"]

Document Retrieval Reranking

models:
  - name: "doc-rerank"
    provider: "cohere"
    model: "rerank-english-v3.0"
    apiKey: "${COHERE_API_KEY}"
    roles: ["rerank"]

RAG System Reranking

models:
  - name: "rag-rerank"
    provider: "cohere"
    model: "rerank-english-v3.0"
    apiKey: "${COHERE_API_KEY}"
    roles: ["rerank"]

Complete RAG Configuration Example

models:
  # Embedding model
  - name: "embed-model"
    provider: "openai"
    model: "text-embedding-3-large"
    apiKey: "${OPENAI_API_KEY}"
    roles: ["embed"]

  # Reranking model
  - name: "rerank-model"
    provider: "cohere"
    model: "rerank-english-v3.0"
    apiKey: "${COHERE_API_KEY}"
    roles: ["rerank"]

  # Generation model
  - name: "chat-model"
    provider: "openai"
    model: "gpt-4"
    apiKey: "${OPENAI_API_KEY}"
    roles: ["chat"]
    defaultCompletionOptions:
      temperature: 0.7
      maxTokens: 2000

Workflow

Typical RAG Flow

1. Query Embedding: Convert user query to vector
2. Vector Retrieval: Retrieve candidate documents from vector database
3. Reranking: Use reranking model to sort results
4. Generate Answer: Generate response based on sorted results

Example Configuration

models:
  # Steps 1 & 2: Embedding and retrieval
  - name: "retrieval-embed"
    provider: "openai"
    model: "text-embedding-3-large"
    apiKey: "${OPENAI_API_KEY}"
    roles: ["embed"]

  # Step 3: Reranking
  - name: "result-rerank"
    provider: "cohere"
    model: "rerank-english-v3.0"
    apiKey: "${COHERE_API_KEY}"
    roles: ["rerank"]

  # Step 4: Generation
  - name: "answer-generator"
    provider: "anthropic"
    model: "claude-3-sonnet"
    apiKey: "${ANTHROPIC_API_KEY}"
    roles: ["chat"]
    defaultCompletionOptions:
      temperature: 0.5
      maxTokens: 2048

Best Practices

1. Candidate Document Count

• Initial Retrieval: 20-100 candidate documents
• After Reranking: Select top 3-10 results
• Balance: Quality and performance

2. Reranking Strategy

• Use after vector retrieval
• Handle semantically similar but irrelevant results
• Boost exact match rankings

3. Performance Optimization

• Limit candidate document count
• Use batch processing
• Cache reranking results

4. Quality Control

• Set relevance thresholds
• Monitor reranking effectiveness
• A/B test different strategies

Evaluation Metrics

Relevance Metrics

• MRR (Mean Reciprocal Rank): Average reciprocal rank
• NDCG (Normalized Discounted Cumulative Gain): Normalized discounted cumulative gain
• Precision@K: Precision of top K results
• Recall@K: Recall of top K results

Troubleshooting

Common Issues

1. Poor Ranking Results

   • Increase candidate document count
   • Check query quality
   • Try different reranking models

2. Performance Issues

   • Reduce candidate document count
   • Use batch processing
   • Implement caching strategy

3. High Costs

   • Optimize candidate document count
   • Use only for critical queries
   • Set usage limits

Comparison with Embeddings

Embedding Models

• Convert text to vectors
• Used for fast retrieval
• Lower cost
• Medium precision

Reranking Models

• Evaluate query-document relevance
• Used for fine-grained sorting
• Higher cost
• Higher precision

Combined Usage

models:
  # Stage 1: Fast retrieval (embedding)
  - name: "fast-retrieval"
    provider: "openai"
    model: "text-embedding-3-small"
    apiKey: "${OPENAI_API_KEY}"
    roles: ["embed"]

  # Stage 2: Accurate ranking (reranking)
  - name: "accurate-ranking"
    provider: "cohere"
    model: "rerank-english-v3.0"
    apiKey: "${COHERE_API_KEY}"
    roles: ["rerank"]

Environment Variables

# ~/.bashrc or ~/.zshrc
export COHERE_API_KEY="your-cohere-api-key"
export OPENAI_API_KEY="your-openai-api-key"

Through proper reranking role configuration, you can significantly improve retrieval system accuracy and user experience.