Text-image person retrieval aims to match textual descriptions to individuals depicted across multiple images, often from varying angles. This multi-view/multi-description nature leads to mismatched positive pairs, where text describes features absent in a specific image view. Existing methods typically ignore this, incorrectly assuming equal matching probability for all positive pairs, which degrades accuracy, and often neglect training efficiency. To address these limitations, we propose two key contributions. First, the Weighted Similarity Distribution Matching Loss (WSDM) leverages CLIP features, using the higher cosine similarity of well-matched pairs to assign them greater true matching probabilities, thus mitigating the negative impact of mismatched pairs. Second, our Bi-Level Alignment (BLA) framework balances accuracy and efficiency. Unlike methods relying solely on global features, BLA incorporates min-pooled features to capture fine-grained details, performing alignment at both coarse and fine levels. This significantly improves retrieval accuracy with minimal computational overhead, requiring only two additional feedforward layers. Our combined approach, BLWS, is validated through extensive experiments on three public datasets, demonstrating superior performance in both retrieval accuracy and training efficiency.