natuleadan Vector Search Core
Overview
Core concepts for vector search with LanceDB: distance metrics, ANN indexing, embedding models, and search patterns.
Quick Reference
Distance Metrics
| Metric | Use Case | Behavior |
|---|---|---|
l2 |
General purpose | Smaller = more similar |
cosine |
Text/document similarity | Smaller = more similar |
dot |
Normalized vectors | Larger = more similar |
hamming |
Binary vectors | Smaller = more similar |
Search Modes
- Exact search: Brute force, 100% recall, no index needed
- ANN search: Approximate, fast, requires index, configurable recall
- Hybrid: Vector + FTS combined with reranking
Embedding Functions
- OpenAI:
text-embedding-ada-002,text-embedding-3-small,text-embedding-3-large - Sentence Transformers: local models via
sentence-transformers - Custom: create your own embedding function
References
- Vector Fundamentals — Distance metrics, ANN vs exact, nprobes
- Vector Search Patterns — Prefiltering, binary search, batch, brute-force
- Embeddings — Embedding function registry, dimension selection