An Adaptive RAG Knowledge Retrieval Method for Large Models Based on Domain Knowledge Enhancement

Knowledge Q&A, a key application of LLMs, draws much attention. However, LLMs struggle with Chinese abbreviated technical terms due to insufficient domain knowledge, and existing power Q&A methods often neglect text semantic context, limiting performance. This paper proposes an LLM-adaptive RAG method with domain knowledge enhancement for Chinese power Q&A. It first presents an info extraction framework with domain term recognition enhancement, using dual-pointer and generative extraction. Second, a two-way recall RAG method is proposed, predicting entity/relationship links via vectors and keywords, generating canonical forms, and using LLMs to select answers. Experiments show it outperforms mainstream methods in power field and works well generally.