Group Relation Policy Optimization Algorithm in RAG Accuracy Improvement Implementation

Retrieval-augmented generation (RAG) has become a key technology for large-model knowledge-driven retrieval tasks, but in the customer service field, it still faces problems such as insufficient retrieval recall, serious generation hallucinations, large consumption of computing resources for reinforcement learning optimization, and poor convergence stability. To solve these problems, this paper proposes a RAG training method based on group relative policy optimization (GRPO). This method abandons the practice of relying on value models in traditional RL and optimizes policies through semi-supervised relative reward estimation within the group, which significantly improves retrieval accuracy and text generation quality while reducing computing resource consumption. Experiments show that after GRPO optimization, the top-4 accuracy of knowledge sources increases from 53.14% to 95.50%, and the proportion of useful model summary polishing increases from 48.49% to 80.28%. This study provides a new idea for the efficient training of RAG models in customer service knowledge management systems.