Curating Information for LLMs in Database Tuning

Develop methods to provide curated, task-relevant information to large language models used for automated database tuning, including retrieval-augmented mechanisms and prompt construction that leverage query-level semantics and historical tuning artifacts to enable accurate configuration recommendations.

Background

The paper introduces Booster, a framework that enriches LLM prompts with query-level historical artifacts to improve tuner adaptivity. While prior work demonstrates LLM capabilities for schema and SQL understanding, effectively supplying curated, relevant tuning knowledge to LLMs remains a challenge cited by the authors as unsolved.

This open problem motivates Booster’s design choices around constructing query-configuration contexts (QConfigs), embedding them for retrieval, and augmenting prompts on a per-query basis to improve recommendations and composition into holistic configurations.