Embeddings vectormath

Development build. This page describes main, not a released package. The latest published Lodestar.Embeddings is 0.4.0 — read its documentation.

VectorMath

The two SIMD primitives dense vector search is built on.

public static class VectorMath

Example — a dot product and a norm, both exact on these inputs.

using Lodestar.Embeddings.Search;

float dot = VectorMath.Dot(new float[] { 1f, 2f, 3f }, new float[] { 4f, 5f, 6f });  // => 32
float norm = VectorMath.L2Norm(new float[] { 3f, 4f });  // => 5

Remarks — these are public because they are useful on their own, not only because EmbeddingIndex needs them. Anything comparing dense float vectors wants the same two operations.

This type carries the package's one deliberate behavioural split between target frameworks. On net10.0, Dot accumulates through System.Numerics.Vector<float>; on netstandard2.0 it is a scalar loop, because the span-based Vector<T> constructor is not available there. Both compute the same dot product, and they add the products up in a different order — floating-point addition is not associative, so the two targets can disagree in the last bits of a long vector. Neither is wrong; they are different roundings of the same sum.

That matters in exactly one place: a score computed on one target and compared for equality against a score computed on the other. Compare with a tolerance, or compare rankings rather than scores.

Applies to — net10.0, netstandard2.0.

See also — EmbeddingIndex, the search index.

Members

Member	What it does
VectorMath.Dot	The dot product of two equal-length vectors.
VectorMath.L2Norm	The Euclidean length of a vector.