Dynamic Prompting for Efficient Knowledge Graph Triplet Extraction

Large language models often falter in knowledge graph construction as prompt complexity exceeds their context limits. To overcome this bottleneck, we introduce KGD-LLM, a framework that shifts the focus from managing context length to maximizing its information density. At its core, KGD-LLM employs a two-tier engine that leverages Sentence-Transformers (SBERT) for semantic selection and Jaccard similarity for diversity control to dynamically select a small, yet highly relevant and diverse set of examples from a larger repository. Our experiments on the REBEL dataset reveal that this approach not only surpasses a strong fine-tuned baseline (GENIE) but also dramatically boosts performance over static prompting, achieving a 73.2% relative F1-score improvement within the same context budget and without any model retraining. These findings demonstrate that intelligent, adaptive prompting is a powerful and efficient alternative to traditional fine-tuning, despite the modest increase in processing time.