Model-based retrieval is an emerging information retrieval paradigm, which directly returns ranked docids of relevant documents for a given query input, based on a pre-trained generative language model. For model-based retrievers, how to define and represent docid is a vital issue affecting retrieval performance. Existing methods lean solely on mining information from the unstructured corpus itself to generate docids, suffering from a lack of structural information to characterize document relations. To tackle this problem, we explore a novel model-based retrieval framework named KG-ASI, enhanced by an external knowledge graph. Specifically, rich structural knowledge is injected into docids through graph representation learning on a corpus-related KG subgraph as external knowledge enhancement, facilitating the end-to-end training process. Besides, to mitigate the exposure bias in query generation, we extend beyond original supervised fine-tuning by implementing a novel contrastive learning method to fine-tune the pre-trained query generation model, augmenting the quality of pseudo-queries. Experiments on two public benchmarks demonstrate the superiority of KG-ASI over strong baselines across common metrics, verifying the effectiveness of our proposed method.