2026年AI資料庫的發展現況 - 四種架構截然不同的方案

I wrote a survey of how AI capabilities are migrating into the database layer, and found at least four architecturally distinct categories: vector databases (embedding similarity), ML-in-database (train-then-predict in SQL), LLM-augmented (route to LLM per query), and predictive databases (Bayesian inference at query time, no model lifecycle). The post covers how inference actually works in each, with architecture diagrams and a comparison table. Also discusses what the taxonomy leaves out: fea