Hybrid Search (BM25 + Vector)

Overview

Hybrid search combines the strengths of traditional keyword-based search (typically BM25) with modern semantic vector search. This approach captures both exact keyword matches and semantic similarities, providing more robust retrieval than either method alone.

How It Works

Dual Retrieval

1. BM25 (Keyword Search): Finds documents with exact or partial keyword matches
2. Vector Search: Finds semantically similar documents using embeddings
3. Fusion: Combines results using algorithms like RRF (Reciprocal Rank Fusion)

Key Advantages

• Best of Both Worlds: Captures exact matches AND semantic meaning
• Handles Terminology: BM25 catches specific terms, jargon, IDs
• Semantic Understanding: Vector search finds conceptually similar content
• Improved Recall: Less likely to miss relevant documents
• Robustness: Works well across diverse query types

Fusion Methods

Reciprocal Rank Fusion (RRF)

Most common approach:

• Score = 1/(rank + k) for each result
• Combines scores from both methods
• Parameter k typically set to 60

Linear Combination

• Weighted sum of BM25 and vector scores
• Requires score normalization
• Weights can be tuned per application

Use Cases

• Enterprise Search: Mix of keywords and concepts
• E-commerce: Product names + semantic features
• RAG Systems: Better document retrieval for LLMs
• Legal/Medical: Specific terms + conceptual queries
• Customer Support: Ticket routing and KB search

Implementation Examples

Major platforms supporting hybrid search:

• Weaviate (native hybrid search with alpha parameter)
• Elasticsearch (hybrid scoring)
• OpenSearch (with RRF)
• Qdrant (hybrid search mode)
• Pinecone (sparse-dense hybrid)

Best Practices

• Start with equal weights, then tune based on data
• Use RRF for simplicity and robustness
• Consider query type (keyword-heavy vs conceptual)
• Test with representative queries
• Monitor both BM25 and vector contributions

Pricing

Implementation varies by vector database platform.