DocGraphLM is a document graph language model for information extraction and question answering. It employs advanced vision-rich document understanding techniques, combining pre-trained language models and graph semantics. Its uniqueness lies in proposing a joint encoder architecture to represent documents and a novel link prediction method to reconstruct the document graph. DocGraphLM predicts the direction and distance between nodes through a convergent joint loss function, prioritizing neighborhood restoration and minimizing the weight of remote node detection. Experiments on three SotA datasets demonstrate that incorporating graph features consistently improves performance in information extraction and question-answering tasks. Furthermore, we report that employing graph features accelerates convergence during training, despite these features being constructed solely through link prediction.