Vector Databases

Embeddings, indexes, filters, and performance trade-offs.

Core Decisions

Index

HNSW, IVF-PQ, Flat (exact), DiskANN

Filters

Metadata pre/post-filter, hybrid keyword

Rerank

Cross-encoder or lightweight MMR

Freshness

Upserts, TTL, delta indexes

Latency vs Cost

Smaller embeddings → cheaper RAM, faster scans
Quantization/IVF → speed with recall trade-off
Batch queries & cache popular items

Metadata Strategy

doc_id, section, product, audience, region
date, version, security_tag
Use filters to narrow before vector search where possible