Answers are grounded in this paper and cite their sources.
Managing systems and network policies for large-scale organizations is challenging for business process automation. Although Policy-as-code (PAC) platforms can ease the task of policy management by defining and executing systems and network policies in the form of programmable codes, converting existing organizational policies into PAC-compliant code is not straightforward due to the need for complex dependency resolutions across platforms and applications. On the other hand, policymakers/top management of a business prefer natural language (NL)-based policies that are easy to comprehend. This paper explores large language models (LLMs) to facilitate the automated conversion of NL-based policies to PAC-complaint code. We observe that public LLMs like ChatGPT need thorough multi-round prompt engineering to generate PAC policies. This concerns privacy and security as the organizational policies are sensitive business information. Consequently, we explore using a private and personalized setup, like private LLMs. Notably, we observe that existing personalized LLMs like PrivateGPT fail to understand the system-specific policy semantics. Consequently, we develop a framework called AutoPAC, which uses a micro-service architecture coupled with fine-tuned models to generate and validate PAC-complaint policies over a personalized LLM framework. An evaluation with more than 100 test cases indicates that the proposed framework effectively generates and validates PAC policies on the fly.
No takes yet. Share an insight, caveat, or question.
Chowdhary et al. (2025) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: