SEMA-SQL: Beyond Traditional Relational Querying with Large Language Models (arxiv.org)

Researchers have introduced SEMA-SQL, a groundbreaking system designed to enhance the capabilities of traditional relational databases by integrating large language models (LLMs) for improved querying. While existing systems allow natural language querying, they often fail to utilize the full semantic reasoning power of LLMs or require cumbersome manual pipeline construction. SEMA-SQL aims to address these limitations by automating query generation, optimization, and execution, thereby enabling dynamic responses to complex questions that involve both structured and unstructured data. At the core of SEMA-SQL is the Hybrid Relational Algebra (HRA), which merges traditional relational operations with LLM-powered user-defined functions (UDFs). The system employs innovative techniques, such as in-context learning for query generation, cost-based optimization, and algorithms that significantly reduce the invocation of LLMs by an impressive 93% during semantic joins. This advancement not only streamlines data retrieval and analysis for more sophisticated queries but also demonstrates considerable improvements in query performance during extensive experiments with benchmark datasets. SEMA-SQL signifies a notable step forward in the AI/ML landscape, bridging the gap between relational database capabilities and the emerging need for semantic understanding in data querying.