SAGE Unifies AI Functions in SQL for Better Performance

A novel framework named SAGE (Self-Adaptive Generative Execution) has been developed to streamline the integration and execution of AI functions within SQL databases. Despite the diverse APIs of AI functions for tasks like classification or summarization, they fundamentally perform only three relational roles: transforming individual rows, aggregating groups, or generating relationships between pairs of rows. SAGE captures these roles with three unified primitives: AI_SCALAR, AI_AGG, and AI_JOIN. These primitives are designed to compose naturally with standard relational operators and share a confidence-gated execution interface. A key innovation lies in SAGE's self-adaptive physical strategies, particularly for AI_JOIN operations. The system intelligently analyzes predicates and decomposes complex conditions, using a "recipe card" and a small label-free probe to select the most efficient execution strategy. Extensive evaluations across various AI operators and workloads demonstrate that SAGE consistently enhances both the quality and efficiency of AI-driven SQL queries. For instance, it can reduce pairwise model calls in factorable joins by over two orders of magnitude, leading to substantial cost reductions and superior overall performance on benchmarks like SemBench.