Rag hallucination inhibition method and system based on multi-agent collaborative verification and storage medium

Abstract

The application discloses a RAG hallucination inhibition method and system based on multi-agent collaborative verification and a storage medium, relates to the technical field of unstructured text data, and comprises the following steps: acquiring an input question and performing hybrid retrieval to obtain a retrieval document set; performing question decomposition on the input question to obtain a plurality of sub-questions and corresponding sub-question answers; generating a plurality of candidate answers based on the input question, the retrieval document set and the sub-question answer set; selecting a target candidate answer through a preference screening agent based on preference learning training, and performing consistency determination through a consistency verification agent based on supervised fine-tuning training; and triggering a closed-loop retry when the determination is inconsistent. Through the technical scheme of the application, a closed-loop inhibition link of retrieval, decomposition, generation, screening, verification and retry can be constructed, hallucination output in the retrieval enhancement generation process is reduced, and the factual accuracy, evidence consistency and overall reliability of the final answer are improved.

Images