LLM-generated code for data processing

Traditionally, database management systems are built over multiple years by large teams of developers. Since 2022, we have been exploring the idea of generating code for data processing via large language models (LLMs) instead. This can take the form of generating highly customizable code for specific queries in general-purpose programming languages. Alternatively, we can generate code implementing database system components that is reused across various queries.

Initially, our focus had been on generating code in high-level languages such as Python that can be customized according to user instructions. Due to limitations of LLMs available before 2025, it was not possible to generate code rivaling the performance of classical database management system. This has recently changed with the emergence of a new class of powerful LLMs, together with powerful frameworks for agentic code generation. In our most recent projects, including GenDB and Jailbreak, we use LLMs to generate code in low-level programming languages such as C that is highly specialized to queries, data, and hardware.

Using a multi-agent system, powered by the newest generation of LLMs, we were able to generate code that outperforms a variety of existing systems for many queries.