Improving the precision of information retrieval has been a challenging issue on Chinese Web. As exemplified by Chinese recipes on the Web, it is not easy/natural for people to use keywords (e.g. recipe names) to search recipes, since the names can be literally so abstract that they do not bear much, if any, information on the underlying ingredients or cooking methods. In this paper, we investigate the underlying features of Chinese recipes, and based on workflow-like cooking procedures, we model recipes as graphs. We further propose a novel similarity measurement based on the frequent patterns, and devise an effective filtering algorithm to prune unrelated data so as to support efficient on-line searching. Benefiting from the characteristics of graphs, frequent common patterns can be mined from a cooking graph database. So in our prototype system called RecipeView, we extend the subgraph mining algorithm FSG, to cooking graphs and combine it with our proposed similarity measurement, resulting in an approach that well caters for specific users' needs. Our initial experimental studies show that the filtering algorithm can efficiently prune unrelated cooking graphs without affecting the retrieval performance and the similarity measurement gets a relatively higher precision/recall against its counterparts.