Automated Dialogue Order Processing using Small Large Language Models

This paper addresses the underexplored intersection of prompt engineering techniques and system design in the context of small Large Language Models (LLMs) aimed at generating structured outputs. While prior studies have largely focused on larger LLMs, the potential of smaller models, which operate with significantly fewer parameters, remains largely untapped, particularly concerning entity extraction in real world task-oriented dialogues. We propose a framework for deploying small LLMs in private, resource-efficient environments to enhance task-oriented workflows. Utilizing the TaskMaster-1 (TM-1-2019) dataset, our research demonstrates how structured outputs can be generated effectively in the absence of any fine-tuning. We evaluate various small LLMs and analyze critical metrics such as validity to identify the key factors and components necessary for transforming dialogue into actionable entities within programming environments.