RGFN is a generative model that extends the GFlowNet framework to directly operate in the space of chemical reactions, enabling synthesizability while maintaining high-quality molecule generation. The model uses a curated set of reactions and building blocks to generate a vast search space of molecules, significantly larger than existing screening libraries, with low synthesis costs. It also scales to large fragment libraries, increasing the number of potential molecules. The approach is evaluated across various tasks, including docking score approximation, GPU-accelerated docking, and biological activity estimation. RGFN outperforms existing methods in terms of synthesizability and diversity, and demonstrates the ability to generate molecules with valid synthetic routes. The model uses a novel action embedding mechanism that improves scalability to larger building block spaces. The study highlights the effectiveness of RGFN in generating molecules with high synthesizability and diverse chemical structures, offering a promising alternative to traditional high-throughput screening methods. The model's ability to generate molecules with valid synthetic routes and its scalability make it a valuable tool for drug discovery.RGFN is a generative model that extends the GFlowNet framework to directly operate in the space of chemical reactions, enabling synthesizability while maintaining high-quality molecule generation. The model uses a curated set of reactions and building blocks to generate a vast search space of molecules, significantly larger than existing screening libraries, with low synthesis costs. It also scales to large fragment libraries, increasing the number of potential molecules. The approach is evaluated across various tasks, including docking score approximation, GPU-accelerated docking, and biological activity estimation. RGFN outperforms existing methods in terms of synthesizability and diversity, and demonstrates the ability to generate molecules with valid synthetic routes. The model uses a novel action embedding mechanism that improves scalability to larger building block spaces. The study highlights the effectiveness of RGFN in generating molecules with high synthesizability and diverse chemical structures, offering a promising alternative to traditional high-throughput screening methods. The model's ability to generate molecules with valid synthetic routes and its scalability make it a valuable tool for drug discovery.