eloquage/vector

Framework-agnostic, in-process vector math for PHP 8.3+: dot product, square-rooted Euclidean distance, cosine similarity, L2 normalization, and exact top-k or batch similarity. Pure PHP is the required implementation; TypePHP extension acceleration is optional.

Installation

composer require eloquage/vector

How-to

Instantiate Eloquage\Vector\Vector, then pass zero-based contiguous numeric lists to the public methods. Integers are accepted and canonicalized to floats.

use Eloquage\Vector\Vector;

$vector = new Vector();

$cosine = $vector->measure([3, 4], [0, 4]); // 0.8
$distance = $vector->measure([3, 4], [0, 4], 'l2'); // 3.0
$unit = $vector->normalize([3, 4]); // [0.6, 0.8]

Find the nearest corpus vectors with an exact scan. Results retain the source indices and are ordered by descending score, then ascending index for ties.

$ranked = $vector->similarity(
    [1, 0],
    [[0.8, 0.2], [1, 0], [0, 1]],
    topK: 2,
);

// [
//     ['index' => 1, 'score' => 1.0],
//     ['index' => 0, 'score' => 0.9701425001...],
// ]

Pass a batch of queries to receive an m × n cosine-score matrix. Corpus row order is preserved in the matrix.

$scores = $vector->similarity(
    [[1, 0], [0, 1]],
    [[1, 0], [1, 1]],
);

// [
//     [1.0, 0.7071067811...],
//     [0.0, 0.7071067811...],
// ]

Interface reference

measure(array $left, array $right, string $metric = 'cosine'): float
normalize(array $vector): array
similarity(array $queries, array $corpus, ?int $topK = null): array

• measure() accepts dot, l2, or cosine. L2 is square-rooted Euclidean distance; cosine is higher-is-better and is computed from L2-normalized vectors.
• normalize() returns a unit-length vector using stable scaling, including for very large finite components.
• similarity() treats one vector query as a ranked result list and a matrix query as a score matrix. topK applies to ranked results, must be positive, and is clamped to the corpus size.
• All vectors and rows must be non-empty, contiguous list<float> values of equal dimension. Matrices must be contiguous lists of such rows. Empty corpora return an empty result list or one empty row per query; an empty query batch returns an empty matrix.
• Ranking is deterministic: score descending, original corpus index ascending on ties. Exact scanning costs O(n·d) for one query and O(m·n·d) for a batch; there is no ANN index, persistence, HTTP, or database layer.

Errors

Malformed, associative, nested, non-finite, empty, or dimension-mismatched inputs, invalid metrics, and invalid topK throw InvalidArgumentException. Zero-norm vectors throw DomainException. Arithmetic that cannot be represented as a finite result throws OverflowException.

Testing

composer test
vendor/bin/pest --coverage --min=90

The optional extension is built by maintainers in the shared Docker workflow; it is not a Composer dependency and does not replace src/Vector.php. See AGENTS.md and TYPEPHP.md for repository workflows.