Hybrid Search

Overview

Hybrid search combines traditional keyword search (BM25) with semantic vector search, leveraging strengths of both approaches.

Components

Keyword Search (BM25)

• Exact term matching
• Fast and efficient
• Handles specific terms well
• No semantic understanding

Vector Search

• Semantic understanding
• Handles paraphrases
• Context-aware
• May miss specific terms

Fusion (RRF)

• Combines result sets
• Balances both approaches
• Improves overall quality

Advantages

• Better Recall: Catches more relevant results
• Improved Precision: Reduces irrelevant matches
• Robust: Works across query types
• Best of Both: Keyword precision + semantic understanding

Implementation

1. Run parallel searches (BM25 + vector)
2. Get top-k from each
3. Apply fusion (RRF recommended)
4. Optional: Rerank fused results
5. Return final ranking

Supported By

• Weaviate
• Qdrant
• Elasticsearch
• OpenSearch
• Milvus (via plugins)
• Vespa

When to Use

• Production RAG systems
• Enterprise search
• When both precision and recall matter
• Diverse query types

Pricing

Technique, costs from vector DB + compute.