Semantic communication (SemCom) lies in the emphasis on ensuring the correct semantic understanding rather than error-free bit transmission. However, traditional hybrid automatic repeat request (HARQ) mechanism relies on a bit-level check, and it cannot effectively address errors at the semantic level. In this paper, we propose a semantic codebook-based HARQ (SCB-HARQ) mechanism for the reliable and efficient SemCom. To enable semantic-level error detection, SCB-HARQ leverages a shared semantic codebook trained offline at both the transmitter and receiver. This codebook serves as prior information for evaluating the distortion of received semantic features, quantifying the extent to which they deviate from the intended meaning. To reduce the transmission overhead of features in the codebook, a weighted semantic feature index (WSFI) clustering method is introduced to map features into a compact index representation. Then, a masked rate-adaptive joint source-channel coding (JSCC) method is proposed to locate and retransmit the distorted features. The simulation results demonstrate that the proposed SCB-HARQ outperforms the traditional HARQ mechanism, achieving a 46.29% improvement in image reconstruction performance while reducing the transmission data volume by 51.27%.