With the development of science and technology, recommender systems continue to attract the attention of countless fields, which provides personalized recommendations for users. In order to obtain more information in the review text and further improve the recommendation accuracy, this paper proposes a search recommendation model based on attention compressed interaction network. It combines the compressed interaction network and attention mechanism to capture the feature information of users and items from the review text, and then learns to predict the interaction between users and items, and based on the learned item representation. Mapping is performed in the natural language space to enable retrieval. Experimental results on the Amazon public dataset show that compared with other baseline models, the MAE and RMSE indicators can be reduced by up to 2.63%and 2.26 % respectively, which demonstrates the feasibility and accuracy of the proposed model.