Fault detection and diagnosis (FDD) play a crucial role in minimizing energy waste and reducing maintenance costs in HVAC systems. Diagnostic Bayesian networks (DBNs), as probabilistic graphical models, offer a promising solution due to robustness to uncertainties, modeling flexibility, scalability, and interpretability. However, the current DBN construction is either a tedious and time-consuming manual process or relies heavily on training data, posing significant barriers to wide-spread adoption. This study proposes a novel large language model (LLM)-driven framework for automating DBN code generation for HVAC systems by extracting knowledge from process and instrumentation diagrams (P&IDs), extending beyond the reliance on traditional sensor data. The results demonstrate that the proposed framework can generate functional DBN code, reasonable symptoms, and DBN parameters. However, fault diagnosis experiments revealed that only the “supply fan stuck” fault was correctly identified, underscoring the need for further refinement. Future work will focus on enhancing LLM capabilities, such as prompt engineering and fine-tuning, and optimizing DBN parameters using limited data to improve diagnostic accuracy.