The researchers proposed the RegionFed architecture for a robust federated learning framework, aimed at addressing the challenges of personalized query understanding in heterogeneous retail environments. This framework relies on gradient-level operations and utilizes the conflict signal of local and global gradients to diagnose data heterogeneity. It adaptsively controls the intensity of personalized strategies and supports direct deployment of various models such as T5-Small, T5-3B, RoBERTa, and CNN without code modifications. Experiments on three datasets: Amazon ESCI, Amazon Reviews, and LEAF-FEMNIST, showed that RegionFed-Meta achieved an accuracy of 92.27% in all four architectures, approaching the upper limit of privacy violations (92.04%), while providing a differential privacy guarantee of approximately 0.60 and a convergence speed of \u03a8(1/\u221aT).