The structured output kind a VisionLLMFeatureStep extracts — the model
emits exactly one named value of this shape per row, which becomes the feature
column's value:
category → a single categorical label (e.g. condition='damaged'); the
produced column is a categorical raw value (typically one-hot encoded by a
downstream preprocessing step).
scalar → a single numeric value the model reads off the image (e.g. an
estimated count); the produced column is numeric.
The structured output kind a VisionLLMFeatureStep extracts — the model emits exactly one named value of this shape per row, which becomes the feature column's value:
category→ a single categorical label (e.g.condition='damaged'); the produced column is a categorical raw value (typically one-hot encoded by a downstream preprocessing step).scalar→ a single numeric value the model reads off the image (e.g. an estimated count); the produced column is numeric.