Governing modern cloud infrastructures is increasingly difficult because complexity grows faster than documentation and diagrams. The result is a fragmented and quickly outdated understanding of how the infrastructure actually works.
In this talk, we will show how to use CNCF Cartography to build a knowledge graph of a computing infrastructure using deterministic data collection from cloud and infrastructure APIs. The graph is not inferred by an LLM: it is generated from infrastructure sources, without consuming LLM tokens during the graph construction phase. This creates a continuously refreshed catalog of infrastructure components and, most importantly, the relationships between them. Once built, the graph becomes a structured context on which LLMs can be applied.
We will show two practical use cases. First, using the graph to automatically generate infrastructure diagrams and documentation that reflect the current state of the system. Second, using the graph as context for chatbot-style questions about the infrastructure, allowing engineers to ask how components are connected, what depends on what, and where specific resources are located.
We will share what worked, where the graph model was useful, and where LLM-based reasoning still needed guardrails.
