Framework for Normalizing Multi-Vendor Network Inputs for LLM-Assisted Network Management

Large Language Models (LLMs) are increasingly used to assist network management tasks such as troubleshooting, intent translation, and automation. Network operations, however, rely on data from many vendors and sources: CLI output, configuration snippets, telemetry, alarms, and vendor-specific APIs. These inputs differ in format, structure, and semantics, which makes it difficult to present a consistent interface to an LLM. This document describes a framework for standardizing such multi-vendor inputs for LLM-assisted network management. Incoming messages from multi-vendor network elements are first handled by an Input Classifier, which determines the nature of each input and assigns it to one of three categories: performance, configuration, or response, using a hybrid approach (rule-based classification first, with escalation to a Small Language Model (SLM) when rules are insufficient). The classified input is then passed to the corresponding Structurer among the Performance Structurer, Configuration Structurer, and Response Structurer, each of which produces a normalized, structured representation (typically with SLM assistance and optional confidence scoring). That structured output is fed to a Prompt Schema Generator, which creates a structured, vendor-agnostic schema and supplies schema-aligned prompts to the central LLM. This document specifies the architecture and component roles for use in design and implementation. It does not define a wire protocol; it is published for informational purposes.