Understanding the relationships between urban crime and street environment is essential for effective crime prevention. However, due to difficulty in crime location extraction and data representation, traditional statistical and machine learning approaches are hard to model complex, nonlinear interactions between street features and crime distribution. For these issues, we present a novel method, which combine advantages of large language model and machine learning to assess crime risk on urban street. Specifically, the proposed method uses a knowledge-enhanced large language model to accurately extract criminal locations from legal documents. After that, it employs high-dimensional features from street view images, replacing traditional manual features. Various experiments conducted on public datasets demonstrate that our method outperforms other state-of-the-art methods. Building on these findings, we applied our method in Wuhan, China and results show its well capability for assessing urban crime risk, further validating its superiority over existing approaches.