More and more on-line information systems allow users to assign the tags to such documents as entries in encyclopedia or articles in blogs or topics in social networks. Traditional way to retrieve these documents is the document retrieval method, but fails at using the tag information. This work proposes a tag-language-model, which builds a language model of the tags used to promote searches on documents with tags. The model integrates the document similarity model and tag similarity model. The experiments show that tag-language-model can effectively improve the searches on documents with tags.