Rerankers

Rerankers enhance the relevance of search results by refining and reordering them after initial retrieval. The Vectara Python SDK enables you to apply various reranker types in queries to optimize result quality for different use cases—from improving precision with neural models to adding diversity or custom business logic.

This section shows how to integrate different reranker types into your queries and configure them for optimal performance in various scenarios. For more information about the available rerankers, see Reranking.

Prerequisites

This guide assumes you have a corpus called my-docs. If you haven't created a corpus yet, follow the Quick Start guide to set up your first corpus.

Basic query with reranker

Improve result ordering in a query by specifying a reranker configuration. This ensures the most relevant, business-critical results appear at the top according to your chosen ranking logic.

QUERY WITH MULTILINGUAL RERANKERpython

1

This example demonstrates using Vectara's multilingual reranker, which provides advanced neural ranking capabilities for both English and multilingual content, making it ideal for improving result quality in RAG applications.

Key Benefits:

• Enhanced relevance scoring using neural models
• Multilingual support for global applications
• Optimized for RAG and question-answering scenarios

---

MMR reranker for diversity

Use Maximal Marginal Relevance (MMR) reranking to balance relevance with diversity, reducing redundancy in search results while maintaining high quality.

QUERY WITH MMR RERANKERpython

1

MMR Configuration Parameters:

• diversity_bias: Float between 0-1 controlling relevance vs diversity balance
• limit: Maximum number of results to rerank
• cutoff: Minimum relevance score threshold for results

Use Cases:

• Product search where you want variety in results
• Research queries requiring diverse perspectives
• Content discovery applications
• Recommendation systems

---

User Defined Function (UDF) reranker

Implement custom scoring logic using User Defined Functions to boost or filter results based on metadata, business rules, or dynamic conditions.

E-COMMERCE UDF RERANKER WITH INVENTORY BOOSTpython

1

UDF Function Examples:

RECENCY BOOSTjavascript

1

RATING BOOSTjavascript

1

PRICE RANGE FILTERjavascript

1

Use Cases:

• E-commerce inventory management
• Content freshness prioritization
• User preference personalization
• Business rule enforcement

---

Chain reranker for complex ranking

Combine multiple rerankers sequentially to implement sophisticated ranking strategies that incorporate relevance, diversity, and custom business logic.

MULTI-STAGE CHAIN RERANKERpython

1

Chain Reranker Strategy:

1. Neural Ranking: Start with multilingual reranker for semantic precision
2. Diversity: Apply MMR with low bias to reduce redundancy while preserving relevance
3. Business Logic: Boost results with high customer ratings for business optimization

Advanced Chain Examples

ACADEMIC RESEARCHjavascript

1

CONTENT WITH RECENCY BIASjavascript

1

---

Streaming queries with rerankers

Apply rerankers in streaming scenarios for real-time applications while maintaining improved result quality as each chunk is received.

STREAMING QUERY WITH RERANKERpython

1

Streaming with Rerankers Benefits:

• Real-time response generation with improved relevance
• Better user experience for interactive applications
• Optimized result quality for chatbots and live search
• Enhanced performance for long-form content generation

---

List available rerankers

Discover available rerankers in your Vectara instance to identify their names and configurations for use in queries.

LIST AND FILTER RERANKERSpython

1

Listing Parameters:

• filter: Regular expression to match reranker names or descriptions
• limit: Maximum number of rerankers to return per page
• page_key: Pagination token for retrieving additional results

Use the reranker names from this list in your customer_reranker configurations.

---

Best practices and optimization

Reranker Selection Guidelines:

CHOOSING THE RIGHT RERANKERpython

1

Performance Guidelines:

• Use single rerankers for high-volume, low-latency scenarios
• Apply chain rerankers for complex ranking requirements
• Set appropriate limit values to balance quality and performance
• Monitor query latency when adding multiple rerankers
• Use cutoff parameters to filter low-relevance results

Error Handling Best Practices:

• Validate reranker names using the list endpoint
• Implement fallback queries without rerankers for reliability
• Handle reranker failures gracefully in production systems
• Test UDF functions thoroughly before deployment

---

Error handling and troubleshooting

Common Issues and Solutions:

RERANKER ERROR HANDLING PATTERNSpython

1

Common Error Scenarios:

• 400 Bad Request: Invalid reranker name or configuration
• 403 Forbidden: Insufficient permissions for reranker access
• 404 Not Found: Reranker not available in your instance
• 429 Rate Limit: Too many reranking requests

Resolution Strategies:

• Validate reranker availability using client.rerankers.list()
• Implement graceful degradation to basic search
• Use appropriate error handling for production reliability
• Monitor reranker performance and adjust configurations

Next steps

After understanding rerankers:

• Query optimization: Combine rerankers with metadata filtering for precise results
• Performance tuning: Monitor and optimize reranker configurations for your use case
• Custom business logic: Develop sophisticated UDF functions for domain-specific ranking
• A/B testing: Compare different reranker configurations to optimize user experience