Rerankers

@tags: reranker, rerank, ranking, relevance, rrf, mmr, hybrid, diversify

What are rerankers?

• Combine results from multiple search strategies (hybrid search)
• Reduce redundancy by diversifying results
• Optimize the relevance-diversity trade-off
• Improve the overall user experience of your search application

When to use rerankers?

• Merge multiple search methods: Combine vector search with BM25 keyword search, or merge results from multiple vector searches
• Diversify results: Eliminate near-duplicate content and show varied perspectives
• Optimize ranking: Fine-tune the balance between relevance and variety based on your use case
• Improve search quality: Leverage sophisticated ranking algorithms without changing your underlying search infrastructure

Best practices for reranking

• Retrieve more results initially: Fetch 100-200 candidates to give rerankers sufficient options to work with
• Apply rerankers before RANGE: Rerank first, then limit the number of results returned
• Choose the right reranker: Use RRF for combining searches, MMR for diversification
• Test with your data: Experiment with different parameters to find what works best for your use case

Reciprocal Rank Fusion using RerankRRF

• Combine multiple ranked lists without requiring score calibration.
• Good for merging results from different search methods.

Syntax

::RerankRRF             // Uses default k=60
::RerankRRF(k: 30.0)    // Custom k parameter

Example: Hybrid search fusion

• Schema:

V::Document {
    content: String,
    created_at: Date
}

• Query:

QUERY SearchDocuments(query_vec: [F64]) =>
    results <- SearchV<Document>(query_vec, 100)
        ::RerankRRF             // Apply reranking
        ::RANGE(0, 10)          // Get top 10 results
    RETURN results

Maximal Marginal Relevance using RerankMMR

• Balances relevance with diversity to reduce redundancy.
• Good for showing varied results instead of similar or duplicate content.

Syntax

::RerankMMR(lambda: 0.7)

Example: Simple diversification

• Schema:

V::Document {
    content: String,
    created_at: Date
}

• Query:

QUERY SearchDocuments(query_vec: [F64]) =>
    results <- SearchV<Document>(query_vec, 100)
        ::RerankMMR(lambda: 0.7)  // Apply reranking
        ::RANGE(0, 10)            // Get top 10 results
    RETURN results

Chaining Rerankers

Example: Fusion and diversification

• Schema:

V::Document {
    content: String,
    created_at: Date
}

• Query:

QUERY AdvancedSearch(query_vec: [F64]) =>
    results <- SearchV<Document>(query_vec, 150)
        ::RerankRRF(k: 60)       // First: combine multiple rankings
        ::RerankMMR(lambda: 0.6) // Then: diversify results
        ::RANGE(0, 10)
    RETURN results