Automatic fake news detection in Portuguese is still often treated as a text classification task, without explicitly representing the factual veracity of the claims. In this work, we evaluate LLM-based triple extraction in a knowledge graph (KG)-based fact-checking system. We replace the Open Information Extraction component of a previous approach with triples generated by Sabiá 4, while keeping the remaining pipeline unchanged. The graph is built only from triples extracted from true news articles and is used as factual support to verify new instances. The experiments use true and fake news articles across four evaluation settings. In the closed setting with the complete KG, LLM-based extraction achieves an F1 score of 0.9992, outperforming the OIE-based configuration. However, in more restrictive evaluation settings, expressive triples require entity normalization, relation standardization, and semantic consolidation to provide factual support beyond direct evidence.