The mathematical theory underlying classical extreme value methods relies on strong assumptions. Such assumptions are often debated, particularly in the context of rainfall extremes. A possible strategy consists of modeling the ordinary events and making inferences on yearly extremes based on the acquired knowledge of the ordinary events distribution. In this paper, we propose a flexible Bayesian semi-parametric model for the daily rainfall. This method is able to induce spatial dependence and improves previous contributions following similar motivations. Posterior inference is conducted via an efficient Markov Chain Monte Carlo method. Specifically, we show that the proposed algorithm is rejection-free and does not require the calibration of any tuning parameter. We illustrate the method of generating maps of the level return of rainfall extreme in the Veneto and Trentino-Alto Adige regions.