No training phase. Connect sources and she starts learning from real work immediately; there is no dataset to prepare, no model to fine-tune, no taxonomy to design. Corrections along the way sharpen her, but that is teaching, not training.
The machine-learning sense of 'training' does not apply: Heidi uses frontier models as-is and puts your context around them, so there is nothing to train and nothing to retrain when things change.
What she does need is exposure and correction, the same as a sharp new hire. The difference is she reads a year of history in her first week and never forgets what you tell her once.