ChatGosen: A Network Configuration Synthesis Approach with Semantic-Computation Decoupling

Network configuration synthesis aims to translate high-level configuration intent into concrete device-level configurations. Recent research has explored leveraging large language models (LLMs) for intent inference and configuration synthesis. However, configuration synthesis inherently requires deterministic topology analysis and protocol-parameter resolution, such as deriving OSPF link costs, BGP route preferences, or static-route next hops from topology constraints and intended forwarding behavior. These operations require explicit protocol semantics and cannot be reliably guaranteed by the probabilistic generation process of LLMs. We propose ChatGosen, a network configuration synthesis method based on semantic-computation decoupling. This approach separates intent inference from deterministic protocol-parameter resolution through a structured, vendor-independent intermediate representation (IR). The LLM handles intent reasoning, task decomposition, generation of device-level configuration subtasks, and final configuration file generation, while topology parsing and protocol-parameter resolution are processed by a deterministic protocol-parameter resolution module. We implemented a prototype system supporting static routing, OSPF, and BGP protocols. Testing on the Topology Zoo real-world network topology dataset demonstrated ChatGosen achieves fully correct routing forwarding configurations in single-intent scenarios across 64-router topologies, with accuracy exceeding 85% in multi-intent scenarios. The synthesis process completes in mere minutes, proving that semantic-computation decoupling provides a practical and scalable foundation for network-wide configuration synthesis.