Text-Based Person Search (TBPS) seeks to accurately locate target pedestrian images within extensive image databases by leveraging natural language descriptions, with the primary challenge centered around achieving cross-modal semantic alignment. Textual descriptions in existing datasets frequently lack sufficient semantic detail to fully represent the content of pedestrian images, resulting in semantic imbalance that significantly hampers the performance of TBPS models. To mitigate this issue, we fine-tune the Multimodal Large Language Model (MLLM) specifically for the TBPS task, utilizing its robust generative capabilities to enrich the semantic content of the text. However, MLLM faces the challenges of visual illusions and potential semantic mismatches in applications. To this end, we propose the Constrained Semantic Augmentation Module (CSAM) and the Influential Semantic Mining Module (ISMM). CSAM filters mismatched information through explicit semantic filtering to improve data quality and reliability; ISMM achieves precise local fine-grained semantic alignment by adaptively extracting and integrating influential semantic features across heterogeneous modalities, effectively reducing the impact of irrelevant information. Experimental results on various TBPS benchmark datasets show that our method significantly improves multiple evaluation metrics and effectively addresses the challenges of semantic imbalance.