natuleadan Hybrid & Multivector Search
Overview
Advanced search patterns combining vector and text search, multi-embedding late interaction models (ColBERT), and metadata filtering.
Quick Reference
Hybrid Search
vector search + FTS → RRF fusion → reranked results
const results = await table.search("query")
.vector(queryVector)
.text("search terms")
.reranker(new RRFReranker(100))
.toArray()
Multivector (ColBERT)
- Late interaction: query tokens ↔ document tokens
- MaxSim scoring for relevance
- IVF_PQ index with cosine metric for multivector columns
Filtering
| Approach | When | Performance |
|---|---|---|
| Pre-filter | Before ANN | Fast with scalar indexes |
| Post-filter | After ANN | May lose results |
| Push-down | SQL pushdown | Most efficient |
References
- Hybrid Search — Vector + FTS combination, RRF reranking
- Multivector Search — ColBERT, late interaction, MaxSim scoring
- Filtering — Pre/post filtering, SQL expressions, scalar indexes