A retrieval-enhanced generative inference method based on large language models

Chat models, such as ChatGPT, have shown outstanding performance in understanding and responding to human instructions, with far-reaching impacts on natural language questioning and answering. However, in vertical domains, harmful false facts can be generated due to the lack of targeted optimization during question answering. To address the generation of harmful false facts and improve the accuracy rate of large language models in answering questions in vertical domains, this paper proposes a retrieval-enhanced generative inference method based on large language models. Firstly, utilizing large language models to extract entity information from question sentences, enhancing database retrieval; secondly, through semantic encoding and embedding vectorization of question sentences, enhancing knowledge base retrieval; finally, we propose a new technique called self-consistency prompt word engineering, and strengthening the retrieval results, generating multiple independent thought chains, and summarizing to obtain the highest consistency answer. Compared with existing benchmark models on a self-built dataset, the experimental results show that the retrieval-enhanced generative inference method based on large language models proposed in this paper can effectively improve the accuracy of large language models in answering questions in the vertical domains.