Chunwei Liu (Purdue)
Integrating foundation models into relational database systems allows users to express complex semantic reasoning declaratively over unstructured text and domain concepts. In this talk, I will share our research journey across two systems addressing this problem: Palimpzest and JevDB.
First, I will discuss Palimpzest, introduced in 2024 as an early system exploring declarative AI-powered data analytics. Palimpzest demonstrated how cost-based physical optimization and model cascades could optimize AI queries using relational principles. Over the past two years, as the database community has broadly explored semantic query processing, scaling these workloads to large relational tables and complex joins has emerged as a central challenge.
Next, I will present our recent work on JevDB, which focuses on scaling semantic query processing to large candidate spaces. JevDB observes that many relational SQL operations do not require open-ended text generation, but rather discrete, typed determinations. By combining non-generative, calibrated decision models with Semantic Bloom Filters and Yannakakis-style candidate pruning, JevDB reduces model invocations and achieves substantial speedups on SemBench and large-scale semantic join workloads.
I will conclude with reflections on the transition from generative cascades to decision-first query processing, and discuss open research challenges in building scalable, AI-integrated database engines.
Chunwei Liu is an Assistant Professor of Computer Science at Purdue University, where he directs the Purdue Data & AI System (PDAIS) Lab. His research spans database systems, compound AI architectures, and scalable semantic query processing. He collaborates broadly with industry leaders, having spearheaded major cloud database benchmarking and workload evaluation initiatives alongside partners such as Microsoft, Amazon, Intel, Meta, and Google. Prior to joining Purdue, he was a Postdoctoral Associate in the Data Systems Group (DSG) at MIT CSAIL and received his Ph.D. from the University of Chicago. His research has been recognized with a Best Paper Runner-Up award at VLDB 2023.