Single-step retrosynthesis aims to predict a set of reactions that lead to the creation of a target molecule and is a crucial task in molecular discovery. Although a target molecule can often be synthesized with multiple different reactions, it is not clear how to verify the feasibility of a reaction, because the available datasets cover only a tiny fraction of the possible solutions. Consequently, the existing models are not encouraged to explore the space of possible reactions sufficiently. To resolve these issues, we first propose a Feasibility Thresholded Count (FTC) metric that estimates the reaction feasibility with a machine-learning model. Second, we develop a novel retrosynthesis model, RetroGFN, which can explore outside the limited dataset and return a diverse set of feasible reactions. We show that RetroGFN outperforms existing methods on the FTC metric by a large margin while maintaining competitive results on the widely used top-k accuracy metric.