Spectra
Taking LLM-Based Semantic SQL Operators to the Limit With Our Own Multi-Modal Data Engine
Project Lead: Andi Zimmerer
Description
Spectra is a research project by the Data Systems Lab at the University of Technology Nuremberg (UTN) that explores how multi-modal large language models (LLMs) can be utilized for flexible query processing inside traditional database systems as semantic operators. Our goal is to push the limits of these semantic operators - making them faster, more robust, cost-efficient, and scalable, while preserving result accuracy guarantees.
By integrating the powerful and flexible data processing capabilities of multi-modal LLMs into query pipelines, Spectra enables entirely new ways of interacting with data and generating insights in analytical query processing.
Spectra is implemented as a DuckDB extension and enriches DuckDB’s declarative SQL interface with additional semantic operators.
Collaborators
- Dhia Bouassida: Master’s thesis on “Optimizing Semantic Operators in Analytical Query Engines”