Recommendation systems have traditionally employed one of two paradigms to harness review information: the document-level approach, which amalgamates all reviews of a user or item into a single document, potentially overlooking the distinct significance of each review; and the text-level approach, which analyzes reviews individually to extract features pertinent to users or items. Recognizing the complementary nature of these paradigms, we propose the Multi-Grained Attention mechanism Recommendation system (MGAR), a novel framework that synergistically learns both document-level and text-level representations. Our model features a document encoder for assimilating document-level features and a text encoder that distills text-level attributes from individual reviews. In MGAR, we introduce an asymmetric cross-attention mechanism that captures the varying relevance of reviews for users and items, acknowledging the unique characteristics of each item and the differential importance of reviews. This mechanism leverages both self-attention and cross-attention to discern these nuances. Additionally, we incorporate user and item ID information to refine our predictive performance. Extensive experimental evaluations across diverse datasets confirm the superior efficacy of MGAR in leveraging review information for enhanced recommendation accuracy.