Relation extraction (RE), a fundamental natural language processing task, extracts semantic relationships between entities from unstructured text. While conventional approaches rely on annotated data and predefined schemas—limiting their effectiveness in low-resource and emerging domains—zero-shot relation extraction (ZS-RE) addresses these constraints by employing large language models (LLMs) to generalize to unseen relations without domain-specific training. However, LLMs often exhibit semantic mismatches in specialized domains, yielding incomplete or unreliable relational triples. To address this, we present GRAG-ZRE, a novel co-evolutionary framework integrating LLMs with knowledge graphs (KGs). The framework incorporates: (1) nucleus entity generation for domain entity grounding, (2) KG-RAG for dynamic triple retrieval and refinement, and (3) bidirectional knowledge enhancement through feedback mechanisms. Experimental results on Conll04, BioRel, and STIXnet demonstrate that GRAG-ZRE outperforms baseline models, achieving state-of-the-art performance in precision, recall, and F1-score while effectively resolving domain-specific semantic discrepancies and reducing annotation dependency.