Application of Multi-Way Recall Fusion Reranking Based on Tensor and ColBERT in RAG

In recent years, Retrieval-Augmented Generation (RAG), as a framework to assist large model (LLM) retrieval enhancement and generation capabilities, has been widely used in the field of natural language processing. However, the recall algorithm, recall rate, precision and recommendation algorithm (Reciprocal Ranker) of different RAG frameworks largely determine the effectiveness of the final retrieval results of large models. This paper proposes a multi-way recall ranking method based on Tensor + ColBERT (Contextualized Late Interaction over BERT) to improve the retrieval part of the RAG system. This method uses the delayed interaction mechanism of ColBERT to improve the diversity and relevance of retrieval results through a multi-path retrieval strategy on the one hand, and uses Tensor to improve the multi-vector MaxSim similarity calculation of ColBERT on the other hand, achieving the efficiency of dual encoders and accurate fusion reranking. Experimental results show that the improved RAG system can significantly improve the recall and precision of retrieval in complex query and long document scenarios, while taking into account the balance of computing resources and processing speed.