Try maximum filtering for normalizing the template spectra.

[gmbrandt]

The phoenix templates are poorly normalized at wavelengths < 4500 Angstrom.

The phoenix templates are normalized by running a median filter. Because the model spectra are extremely high signal to noise (effectively noiseless), one should in theory do better by running a maximum filter (e.g., a 99% quartile filter) over the spectrum. See if this helps and yields better results.