In this paper, we present a system to browse a set of graphical documents without prior knowledge of their content. This system relies on a new structural method locating regions potentially embedding a symbol, using an unsupervised reconstruction of the documents from their representation as chain of points. The method is complemented by a querying mechanism that retrieves from these regions those that are the most similar to a given input symbol. This mechanism uses a specific user feedback scheme to improve the accuracy of the retrieval.