Document-level Relation Extraction (DocRE) aims to identify relations between entities dispersed throughout a document, which requires reasoning beyond the boundaries of individual sentences. Most existing approaches depend on contextual attention mechanisms in pre-trained language models (PLMs), without explicitly incorporating the semantics of relation labels. In fact, we observe that the attention of an entity pair is often misdirected towards tokens unrelated to the actual relation. We propose a relation-aware DocRE method, CRAG, that aligns the contextual representations of entity pairs with the linguistic expressions of relation labels via a prompt graph. For each relation label, we construct multiple prototype representations composed of evidence phrases extracted from training documents. Interior prototypes encode the core semantics of a relation, and border prototypes capture expressions located near the semantic boundaries shared by relevant relations. By connecting an entity pair to relevant prototype nodes in a prompt graph, CRAG effectively aligns contextual semantics of the entity pair with rich relational semantics, such as diverse linguistic patterns in the evidence sentences of labels and inter-label correlation captured in a prompt graph. Experiments on benchmark datasets show that CRAG outperforms baselines, highlighting the effectiveness of a prompt graph.