This paper presents a system for automatically generating discourse structures from written text. The system is divided into two levels: sentence-level and text-level. The sentence-level discourse parser uses syntactic information and cue phrases to segment sentences into elementary discourse units and to generate discourse structures of sentences. At the text-level, constraints about textual adjacency and textual organization are integrated in a beam search in order to generate best discourse structures. The experiments were done with documents from the RST Discourse Treebank. It shows promising results in a reasonable search space compared to the discourse trees generated by human analysts.