In the realm of artificial intelligence large language model applications, Retrieval-Augmented Generation (RAG) technology was developed to enhance the precision of responses to complex inquiries. Nevertheless, this technology exhibits certain limitations when it comes to handling the vast diversity inherent in human language. To overcome these constraints, Lexical Diversity-aware RAG (DRAG) technology was introduced.

Addressing the drawbacks of RAG, DRAG incorporates a 'Diversity-Aware Relevance Analyzer' (DRA) during the retrieval stage. This analyzer dissects the components of a question, establishes evaluation criteria, and filters out documents that are more pertinent. In the generation phase, DRAG implements a 'Risk-Guided Sparse Calibration Strategy' (RSC) to evaluate and adjust the risk associated with words, thus minimizing the influence of irrelevant information.

Practical tests have demonstrated that models employing DRAG technology experience a 45.5% boost in accuracy compared to traditional RAG models. It is anticipated that DRAG will furnish more accurate and dependable responses across a broader spectrum of scenarios in the future.