Lodestar.Embeddings 0.6.0. This page is frozen at that release. Read the current documentation for what main says now. A link to a decision or a migration page follows main, and leaves the archive.

Selects up to count candidates.

public static int[] Select(ReadOnlySpan<float> query, IReadOnlyList<float[]> candidates, int count, double lambda = 0.5)

Parameters — query is what relevance is measured against. candidates are the candidate vectors, all of query's length. count is how many to select; more than there are selects them all. lambda is 1 for pure relevance, 0 for pure diversity.

Returns — int[], the chosen indices in selection order — not re-sorted by score afterwards.

Exceptions — ArgumentNullException when candidates is null. ArgumentOutOfRangeException when count is negative, or lambda is outside [0, 1]. ArgumentException when a candidate is null, of a different length than query, or has a zero or non-finite norm — cosine is undefined in either case, and the same check applies to query itself.

Example — asking for more candidates than exist returns them all, once each.

using Lodestar.Embeddings.Search;

float[] query = [1f, 0f, 0f];
float[][] candidates =
[
    [1.00f, 0.00f, 0.00f],
    [0.80f, 0.60f, 0.00f],
    [0.60f, 0.00f, 0.80f],
    [0.00f, 1.00f, 0.00f],
];

int[] chosen = Mmr.Select(query, candidates, count: 99);
int returned = chosen.Length;  // => 4

Remarks — the default lambda, 0.5, weighs relevance and diversity equally. A zero-vector, NaN or infinite norm is refused rather than treated as a degenerate cosine of zero, on either query or any candidate — a silent zero would rank that candidate as neither similar nor dissimilar to anything, which is not what an undefined value means.

VectorMath.Dot sums in a different order on net10.0 (SIMD) than on netstandard2.0 (scalar), so a genuine near-tie between two candidates can select a different index on the two targets — accepted, not a defect, and the same divergence VectorMath already documents for the dot product itself.

Applies to — net10.0, netstandard2.0.

See also — Mmr, VectorMath.Dot, the search index, the Python equivalence table.