To identify and constrain the target zone for the planned SAFOD Main Hole through the San Andreas Fault (SAF) near Parkfield, California, a 32-level three-component (3C) geophone string was installed in the Pilot Hole (PH) to monitor and improve the locations of nearby earthquakes. The orientation of the 3C geophones is essential for this purpose, because ray directions from sources may be determined directly from the 3D particle motion for both P and S waves. Due to the complex local velocity structure, rays traced from explosions and earthquakes to the PH show strong ray bending. Observed azimuths are obtained from P-wave polarization analysis, and ray tracing provides theoretical estimates of the incoming wave field. The differences between the theoretical and the observed angles define the calibration azimuths. To investigate the process of orientation with respect to the assumed velocity model, we compare calibration azimuths derived from both a homogeneous and 3D velocity model. Uncertainties in the relative orientation between the geophone levels were also estimated for a cluster of 36 earthquakes that was not used in the orientation process. The comparison between the homogeneous and the 3D velocity model shows that there are only minor changes in these relative orientations. In contrast, the absolute orientations, with respect to global North, were significantly improved by application of the 3D model. The average data residual decreased from 13?? to 7??, supporting the importance of an accurate velocity model. We explain the remaining residuals by methodological uncertainties and noise and with errors in the velocity model.