Dense retrieval, as an embedding-based semantic matching technique, is widely adopted in industrial search systems. However, constrained by the need for real-time query embedding, existing methods primarily rely on encoder-based pretrained language models (e.g., BERT) as embedding models to meet latency requirements. This limitation has hindered the exploration of large language models (LLMs) in real-world dense retrieval systems. Considering the high query cache hit rate in Meituan’s food-related search scenario, we propose to deploy LLMs for dense retrieval by implementing embedding caching for both queries and products. We extensively explore the usage of LLMs, and provide practical insights for deploying LLMs in industrial search systems. Specifically, based on the empirical validation of general improved designs, we further propose a generative learning strategy that leverages auxiliary textual information to assist LLMs in domain adaptation, significantly boosting the retrieval capabilities. Offline experiments and online A/B tests demonstrate the effectiveness of our proposed method.