Optimizing Vertical Domain RAG: Domain-Adaptive Data Processing and RLAIF Synergy

With the rapid advancement of large language models (LLMs) in question answering systems, retrieval-augmented generation (RAG) has significantly improved LLM performance on knowledge-intensive tasks by integrating external knowledge sources. However, due to incomplete data processing pipelines and domain-specific information biases in LLMs, conventional RAG frameworks often perform poorly in vertical domain question answering scenarios. Focusing on the domain of food safety standards, this study proposes an enhanced RAG framework. The approach improves the efficiency of the retrieval module through customized data processing strategies and incorporates reinforcement learning from AI feedback (RLAIF) to infuse the generation module with domain knowledge and high-quality response preferences using a self-constructed dataset. This framework effectively mitigates issues of value misalignment and factual inaccuracy commonly encountered in LLM applications within vertical domains. Experimental results demonstrate that the proposed RAG framework outperforms conventional RAG in terms of response accuracy and domain adaptability, validating its effectiveness. This work offers new insights and methodologies for building verticaldomain QA systems and enhances the applicability of LLMs in specialized fields.