← AI Terminology
Hybrid Search
Hybrid search combines lexical (e.g. BM25) and semantic (vector) retrieval, usually fusing scores or ranks to get the best of both.
It is the default production pattern for high-quality RAG.
It is the default production pattern for high-quality RAG.
Why It Matters in AI
Pure vectors miss exact tokens; pure keywords miss synonyms. Hybrid fusion reliably improves recall and precision on real corpora — enterprise search teams treat it as standard.
Key Points
| Aspect | Description |
|---|---|
| Parts | BM25/sparse + dense vectors (+ optional learned sparse) |
| Tools | Elastic, OpenSearch, Weaviate, Vespa, LangChain ensembles |
| Fusion | Linear score mix, RRF (reciprocal rank fusion) |
| Tuning | Weights per corpus; always remeasure |
| Related | Semantic search, BM25, ColBERT |
| Next step | Cross-encoder rerank on fused candidates |
Simple Analogy
Looking up a restaurant both by star rating map (semantics) and by exact name spelling (keywords) — then merging the shortlists.
Common Usage Examples
- RRF over BM25 hits + vector hits
- Elasticsearch hybrid queries
EnsembleRetrieverwith weights- A/B: vector-only vs hybrid recall@10
Summary
In short: Hybrid search merges keyword and vector retrieval so systems catch both exact terms and paraphrased meaning.