In future wireless communication systems, reconfigurable intelligent surfaces (RIS) present a promising technology. However, in RIS-assisted communication systems, feedback of downlink channel state information (CSI) poses significant challenges, particularly with a large number of base station (BS) antennas and RIS unit cells. Traditional CSI feedback methods based on compressed sensing assume channel sparsity and require extensive computation and storage operations by both the user equipment (UE) and the BS, thereby increasing the burden on the UE. This paper proposes a lightweight CSI feedback mechanism based on global context attention network (GCANet). This method uses a lightweight autoencoder at the UE for CSI encoding, reducing the computational and storage burden on the UE, while deploying the decoder at the BS to leverage its powerful computational capabilities for complex decoding tasks. Simulation results demonstrate that this method significantly reduces feedback overhead and enhances system performance, with the UE’s encoder model parameters and computational load being substantially reduced compared to baseline methods.