KathDB

An Explainable Multimodal Database System with Human-AI Collaboration

About KathDB

Modern data-intensive applications increasingly rely on multimodal data such as text, images, and videos, yet traditional database systems are limited to structured tables, while recent AI-powered systems often sacrifice explainability and semantic guarantees.

KathDB is an explainable multimodal database management system that bridges this gap. It combines the relational model and cost-based query optimization of traditional DBMSs with the reasoning capabilities of foundation models, enabling users to query multimodal data using natural language while retaining structured semantics.

KathDB introduces a unified relational semantic layer over text, images, and video, together with a function-as-operator execution model that compiles queries into modular, versioned functions. This design enables fine-grained lineage tracking, cost-based optimization, and rich explanations that trace query results back to their underlying data and transformations.

Unlike black-box LLM-based systems, KathDB keeps users in the loop through interactive clarification, debugging, and explanation channels, allowing users to iteratively refine queries and understand results across modalities.

KathDB: Explainable Multimodal Database Management System with Human-AI Collaboration. Guorui Xiao, Enhao Zhang, Nicole Sullivan, Will Hansen, Magdalena Balazinska. CIDR, 2026. arXiv Preprint