We propose Graph–Beam Retriever (GBR), a lightweight, training-free framework for extracting concise and semantically faithful evidence in fact verification tasks. Motivated by the challenges of applying large language models (LLMs) in low-resource languages such as Vietnamese, GBR constructs a linguistically informed text graph from raw news context and performs beam search traversal from the claim node to candidate evidence sentences. These paths are subsequently refined using semantic similarity and logical entailment filters. The resulting evidence is strictly bounded in length, ensuring compatibility with LLM token limits while preserving factual relevance. We evaluate GBR on the ViFactCheck benchmark and compare it against strong sparse, dense, and neural retrievers. Despite using less than 40% of the original context, GBR achieves comparable retrieval quality to fine-tuned dense methods and improves zero-shot verification accuracy of GPT-4 by +2.2% over using the full context. Ablation studies highlight the contribution of each module, and case analyses demonstrate GBR’s ability to preserve critical information with minimal overhead. Our approach offers a practical and interpretable alternative to dense retrievers and generative methods, especially in settings where training data or computational resources are limited.