LLM ORDER BY: A Semantic Operator for Efficient Sorting

This study presents the LLM ORDER BY semantic operator, serving as a logical framework for organizing data through large language models. The researchers suggest enhancements to current semantic sorting techniques and introduce a semantic-aware external merge sort. Their analysis indicates that no single implementation excels in all scenarios, highlighting a scaling relationship during testing between sorting expenses and the quality of ordering in comparison-based approaches. They developed a budget-conscious optimizer that employs heuristic strategies, LLM-as-Judge evaluations, and consensus aggregation to choose nearly optimal access paths dynamically. This optimizer demonstrates ranking precision comparable to or exceeding that of the best static methods across various benchmarks.