Definition
Hybrid search combines traditional keyword-based search (like BM25) with modern semantic vector search to find relevant documents. By fusing these two approaches, it captures both exact keyword matches and conceptual similarity, providing more robust retrieval than either method alone.
Why it matters
Hybrid search addresses the limitations of pure approaches:
- Best of both worlds — catches exact terms AND conceptual matches
- Failure mode coverage — when one method misses, the other often succeeds
- Domain flexibility — works across technical and natural language queries
- Production reliability — more consistent results across query types
- RAG quality — improves document retrieval for generation pipelines
Pure vector search can miss exact terms; pure keyword search misses synonyms—hybrid catches both.
How it works
┌────────────────────────────────────────────────────────────┐
│ HYBRID SEARCH │
├────────────────────────────────────────────────────────────┤
│ │
│ User Query │
│ "VAT rules article 15bis" │
│ │ │
│ ┌───────────┴───────────┐ │
│ ▼ ▼ │
│ ┌───────────────────┐ ┌───────────────────┐ │
│ │ KEYWORD SEARCH │ │ VECTOR SEARCH │ │
│ │ (BM25/TF-IDF) │ │ (Embeddings) │ │
│ │ │ │ │ │
│ │ Exact matches: │ │ Semantic matches:│ │
│ │ - "article 15bis"│ │ - VAT regulations│ │
│ │ - "VAT rules" │ │ - Tax exemptions │ │
│ │ │ │ - Related tax law│ │
│ └─────────┬─────────┘ └─────────┬─────────┘ │
│ │ │ │
│ └───────────┬───────────┘ │
│ ▼ │
│ ┌───────────────────────┐ │
│ │ SCORE FUSION │ │
│ │ • Reciprocal Rank │ │
│ │ • Weighted Sum │ │
│ │ • Convex Combination│ │
│ └───────────┬───────────┘ │
│ ▼ │
│ Combined Results │
│ (Best of both methods) │
└────────────────────────────────────────────────────────────┘Fusion methods:
- Reciprocal Rank Fusion (RRF) — ranks based on position in each result list
- Weighted sum — combines normalized scores with configurable weights
- Convex combination — α × keyword_score + (1-α) × vector_score
Common questions
Q: What ratio of keyword to vector search works best?
A: Start with 50/50, then tune based on your queries. Technical domains (legal, medical) often benefit from higher keyword weight (60-70%) for precise terminology. Conversational queries favor vector search (60-70%).
Q: When does hybrid search help most?
A: When queries mix specific terms with conceptual questions. For example, “Article 15bis VAT exemptions for digital services”—needs exact article match AND semantic understanding of exemptions.
Q: Does hybrid search add latency?
A: Slightly—you’re running two searches. But both can execute in parallel, so overhead is minimal (~10-50ms). The quality improvement usually justifies this cost.
Q: What’s Reciprocal Rank Fusion (RRF)?
A: RRF combines rankings without needing comparable scores. For each document, it computes 1/(k + rank) for both methods and sums them. It’s robust because it only uses position, not score magnitude.
Related terms
- Semantic Search — vector-only retrieval component
- BM25 — classic keyword search algorithm
- Embeddings — vectors for semantic search
- Reranking — often follows hybrid search
References
Robertson & Zaragoza (2009), “The Probabilistic Relevance Framework: BM25 and Beyond”, Foundations and Trends in Information Retrieval. [3,000+ citations]
Karpukhin et al. (2020), “Dense Passage Retrieval for Open-Domain Question Answering”, EMNLP. [3,500+ citations]
Cormack et al. (2009), “Reciprocal Rank Fusion outperforms Condorcet and individual Rank Learning Methods”, SIGIR. [500+ citations]
Ma et al. (2021), “A Replication Study of Dense Passage Retriever”, arXiv. [200+ citations]