RA-CFGPT: Chinese Financial Assistant with Retrieval-Augmented Large Language Model

Retrieval-Augmented Generation (RAG) enhances the generative capacity of Large Language Models (LLMs) by appending retrieved documents to the current context. This approach has shown success in reading comprehension and language modeling. RAG assumes the intent is in the input query, which can be expanded with a task description. However, in the financial domain, queries often span multiple sectors, challenging the ability of retrieval phase to adequately inform the generation phase. The complexity of financial texts necessitates large language models to adeptly understand intricate financial terminology and concepts. In this paper, we introduce an integrated system tailored for Chinese financial tasks, including question answering, document analysis, and risk assessment. The system features: (1) the hybrid financial knowledge base to provide comprehensive information, (2) the training of large language model to fit the RAG process, and (3) the system pipeline to ensure the accuracy, compliance and risk warning in output.

Frontiers of Computer Science (FCS 2024)