⚠️ EmbeddingGemma's model card explicitly states activations do not support fp16 — outputs may be
inaccurate or contain NaNs. Shown here for exploration only.
The model's ONNX graph already L2-normalizes its output embedding internally (mean-pool → 2 dense
layers → divide by L2 norm), so with the toggle on and metric = cosine, this reduces to a plain
dot product on unit vectors. Turn the toggle off or switch metric to see the effect of normalization
and metric choice directly.