This paper presents an information retrieval model based on a kind of probabilistic network. The dependence relationships between terms and documents are represented by the topology of the network. The network contains two parts: the term layer and the document layer. The term relationships are represented as an undirected probabilistic network, with the directed arcs toward the nodes indexed in the document layer. Using a learning algorithm, the topology and the probabilities encoding the strength of the relationships can be learned from the document collection automatically. We also provide a two-part inference process to obtain the relevance of the documents to a given query.