AI-Powered Query Processing

AI can extract useful information from text, images, and other forms of unstructured data, but applying AI models to large datasets remains complex and computationally expensive. This project will develop new components that extend existing data systems, allowing users to express AI-powered queries through high-level, SQL-like interfaces while automatically executing them efficiently on modern hardware.

A first effort in this area is our recent work on Kalypso, an efficient LLM serving system for relational queries with semantic operators. The work introduces a new query execution architecture designed to maximize KV-cache reuse across dependent operators. At its core is a memory-aware scheduler that keeps cached state available until downstream processing completes, avoiding premature cache evictions while preserving parallelism. The architecture supports arbitrary semantic operators without compromising accuracy and delivers speedups of up to 4.5× over existing semantic query processing systems.