RAG Boosts Precision but Cuts Breadth in Knowledge Base Systems, Study Finds

A June 2025 research paper examined three approaches to extracting structured historical events from narrative text, comparing direct generation, knowledge-graph enhancement, and retrieval-augmented generation using the first ten chapters of Thucydides. The study tested multiple large language models, including GPT-4, Claude, and Llama 3.2, finding that no single architecture outperformed the others across all quality dimensions. RAG improved precision, coordinate accuracy, and metadata completeness, but reduced breadth and coverage compared to direct generation methods. Model behavior also varied significantly, with larger models showing stable incremental gains from RAG while Llama 3.2 produced highly inconsistent results. The research concludes that teams building knowledge bases should first identify which failure mode — imprecision or missing coverage — is most costly for their specific use case before choosing a retrieval strategy.