Research on retrieval enhancement generation methods for cross-knowledge base retrieval

Abstract

Multi-layered knowledge sources and technologies are used for the knowledge management in power firms in terms of policies and rules, internal systems, project experience and breakthroughs. Most existing RAG approaches lack heterogeneous knowledge integration, retrieval and generation objectives, and knowledge tracing capabilities in cross-knowledge base settings. We design a heterogeneous Knowledge Dynamic Alignment Retrieval Enhancement Generation (HDR-RAG) algorithm to build a three-level system of “heterogeneous knowledge dynamic alignment, bidirectional feedback retrieval, and collaborative optimization generation”. HDR-RAG synthesizes structured knowledge, raw documents and business knowledge fragments in a unified way through contrastive learning. It leverages cross-database filtering, original text tracing, and multi-model collaborative question answering. We adopt hybrid retrieval and production feedback to achieve high retrieval performances and we use joint reward and lightweight reinforcement learning to maximize generation quality. Experimental results show HDR-GR outperforms major comparison algorithms (R@10 reaches 89.7%, P@10 reach 82.3%), generation quality (BLEU-4 reaches 65.3%, fidelity reaches 91.5%), and efficiency to support cross-connection question answering, knowledge retrieval, intelligent creation in power knowledge service setting.