Keyword-Enhanced Semantic Retrieval and Multi-Dimensional Relevance Ranking in RAG

In the task of generating medical hot topics, efficiently and accurately retrieving relevant information based on hot keywords and generating credible topics is a challenging problem. Addressing the limitations of existing retrieval and ranking methods in terms of accuracy and relevance, this paper proposes a RAG-based keyword-enhanced semantic retrieval and multi-dimensional relevance ranking approach. First, we construct keyword vectors using the BGE model and perform fast similarity searches to achieve preliminary retrieval. Then, we optimize the relevance of the retrieval results by setting a similarity threshold and applying a cluster and de-duplication strategy. Building on this, multi-dimensional similarity measurement is introduced, combining a cross-encoder model and feature extraction to implement reordering, further improving the accuracy of retrieval results. Experimental results show that this method significantly enhances RAG's performance in generating hot topics from medical news and information, providing an effective solution for complex information retrieval tasks.