Partitioning Prompts for Higher Efficacy in Network Design with Large Language Model

In this paper, we propose deliverable partitioning in prompt design to assist Large Language Models (LLMs) in improving response correctness for network design and configuration. While recent research has explored the use of LLMs to enhance network management efficiency, their responses often remain inconsistent, incomplete, or inaccurate. Often, LLM-generated configurations contain missing or erroneous configuration commands, which can lead to operational failures. Our proposed partitioning methodology aims to mitigate these issues by decomposing complex network configuration tasks into simplified and focused tasks. To evaluate the effectiveness of this approach, we introduce a scoring policy and conduct extensive experiments across three levels of network complexity and varying degrees of design choice ambiguity. We also compare the performance of leading LLMs, including ChatGPT, Copilot, and DeepSeek. Our findings indicate that partitioning the inquiry process leads to more accurate and consistent responses than non-partitioned approaches, especially in scenarios where design parameters are explicitly defined and leave some but small room, as ambiguity, for inference.