Continual weighted retriever for rag-based systems

Abstract

A retrieval augmented generation (RAG)-based system is disclosed that is configured to perform weighted retrieval, weighted augmentation, and weighted generation. A continual learning based predictor model is trained to infer weights that correspond to sources in a knowledge base based on an input query. When a new source is added, synthetic questions and answers are generated such that the predictor model can be retrained to learn the new source. When documents are retrieved from a knowledge based, original rankings can be re-ranked based on the inferred weights. This weighted retrieval ensures that more relevant documents are provided as context to a prompt, which prompt is input to a large model. The predictor model allows the RAG-based system to adapt to changes in the knowledge base and maintain performance.