Researchers have developed EnSI-RAG, a framework for long-document question answering that constructs an entity-centered index to retrieve relevant evidence. This approach separates evidence localization from answer synthesis and preserves source evidence. The framework outperforms existing methods on two benchmark datasets, achieving an average accuracy of 78.24%. The code is available on GitHub. --- Why it matters: This matters because it improves the performance of question answering systems over long documents, which can help in applications such as information retrieval and text summarization. Source: https://arxiv.org/abs/2608.21252