Sparse retrieval and dense retrieval are currently the two primary methods in search systems. Although dense retrieval can better understand complex semantic relationships and contextual information, sparse retrieval still demonstrates significant advantages in terms of retrieval efficiency and interpretability, especially when handling large-scale document retrieval. Considering the advantages of both approaches, we propose Embedding-Based Sparse Retrieval (EBSR), a sparse framework that can be integrated with any transformer-based semantic model. EBSR encodes queries and documents into sparse embeddings consisting of ID-weight pairs of variable dimensions. These IDs are used to build the inverted index, while the weights are used to calculate similarity. We compared EBSR against the state-of-the-art sparse retrieval methods on the public MS MARCO datasets and deployed it in an industrial search system for an online A/B test. Both offline and online evaluations demonstrated the effectiveness of EBSR.