Class EmbeddingPooler

Namespace

Qavren.Edge.Embeddings.Onnx

Assembly

Qavren.Edge.Embeddings.Onnx.dll

Allocation-free pooling over a session output span. Public because it is worth testing alone.

public static class EmbeddingPooler

Inheritance

object

EmbeddingPooler

Inherited Members

object.Equals(object)

object.Equals(object, object)

object.GetHashCode()

object.GetType()

object.MemberwiseClone()

object.ReferenceEquals(object, object)

object.ToString()

Remarks

The order is fixed and it matters: pool, then LayerNorm(Span<float>, float) when the preset asks for it, then L2Normalize(Span<float>). Layer-norming a unit vector is not the same operation as unit-normalising a layer-normed one, and only the second matches nomic-embed-text-v1.5's reference pipeline.

Methods

ClsPool(ReadOnlySpan<float>, int, int, Span<float>)

Copies row zero, the [CLS] position. bge-small-en-v1.5 pools this way.

L2Normalize(Span<float>)

Divides by the Euclidean norm, in place. A zero vector is left alone: the all-padding MeanPool(ReadOnlySpan<float>, ReadOnlySpan<long>, int, int, Span<float>) guard produces one, and dividing by a zero norm here would turn a defined zero vector into NaNs one call later.

LayerNorm(Span<float>, float)

Mean/variance normalisation over the pooled vector, matching PyTorch F.layer_norm(x, (dim,)) with no learned weight or bias: subtract the mean, divide by sqrt(variance + epsilon). Required by nomic-embed-text-v1.5 and by nothing else here; see PostPoolLayerNorm.

MeanPool(ReadOnlySpan<float>, ReadOnlySpan<long>, int, int, Span<float>)

The masked mean: sum(h[t] * mask[t]) / max(1, sum(mask)). The denominator's max(1, ...) is what keeps an all-padding row a defined zero vector rather than dimensions NaNs.