A Retrieval-Augmented Generation method based on Intent Analysis and Ranking Fusion

Retrieval-Augmented Generation (RAG) can provide reliable and up-to-date external knowledge to drive the solution of the illusion problem of Large Language Models (LLMs) in intelligent question-and-answer tasks. However, existing RAG still suffer from the challenges of query disambiguation and noisy knowledge retrieval. In this regard, this paper proposes a retrieval-augmented generation method based on intent analysis and ranking fusion. In this paper, we first introduce question rewriting on the basis of pre-retrieval augmented generation, generate query expanders by analyzing the original question and the theme of the retrieval results, and generate diversified sub-questions to alleviate query ambiguity by combining with the analysis templates; second, a semantic threshold optimization cleaning mechanism based on ranking fusion is proposed, which combines the text embedding model and similarity calculation method to set semantic thresholds and filter the noise knowledge existing in the retrieval results; Finally, the combination of question rewriting and semantic threshold optimization enhances the ability to capture the user's intent and reduces the filtered noise information in the retrieval results, achieving a deep fusion of pre-retrieval and post-retrieval augmented. The experiment results show that compared to the existing retrieval augmented generation methods, the method proposed in this paper achieves 94.36% and 78.80% in retrieval result accuracy and precision rate, which are 3.47% and 4.19% higher than the baseline model, respectively, and the experiment results validate the effectiveness of the method.