VectorRAG and GraphRAG reduce LLM hallucinations in SMEs by grounding responses

A new arXiv paper from Md. Samiul Islam and colleagues proposes two RAG-based approaches—VectorRAG and GraphRAG—to address LLM hallucinations in small and medium enterprises (SMEs). As SMEs increasingly rely on LLMs for question-answering and decision support, fake or fabricated outputs can erode trust and lead to costly mistakes. The researchers evaluated both RAG variants across multiple open-weight models, including LLaMA, Mistral, and Qwen, focusing on useful response generation, hallucination risk, contextual relevance, and human interpretability.

The results show that integrating external knowledge via RAG significantly improves response reliability. VectorRAG retrieves relevant text chunks by embedding similarity, while GraphRAG builds a knowledge graph to capture entity relationships and provide more structured, context-aware answers. Both methods outperformed baseline LLMs in reducing misinformation, with GraphRAG showing strengths in handling complex relational queries. The findings offer SMEs a practical path to deploy LLMs with higher confidence, without needing costly fine-tuning or custom models.