The paper presents and discusses direct and indirect tuning of a knowledge-driven harmonization model for tonal music. Automatic harmonization is a data analysis problem: an algorithm processes a music notation document and generates specific meta-data (harmonic functions). The proposed model could be seen as an Expert System with manually selected weights, based largely on the music theory. It emphasizes universality - a possibility of obtaining varied but controllable harmonies. It is directly tunable by changing the internal parameters of harmonization mechanisms, as well as an importance weight corresponding to each mechanism. The authors propose also indirect model tuning, using supervised learning with a preselected set of examples. Indirect tuning algorithms are evaluated experimentally and discussed. The proposed harmonization model is prone both to direct (expert-based) and indirect (data-driven) modifications, what allows for a mixed learning and relatively easy interpretation of internal knowledge.