Embedding

Vectors are the retrieval side of RAG: every passage and every query becomes a float array, and similarity decides what the model sees. Embeddings::for() embeds many inputs in one call; the full reference (generation, multimodal inputs, caching) lives under Providers.

Widths for your columns

Size the vectors to the column that will hold them. dimensions() pins the width explicitly; without it the provider default applies:

use Crustum\Ai\Embeddings;

$response = Embeddings::for(['Napa Valley has great wine.'])->dimensions(1536)->generate();

echo count($response->first()) . " dimensions\n";

1536 dimensions

Pick the width once: the vector column length, the provider model, and every query must agree. Changing any of the three means re-embedding. Which providers embed, and their defaults, is a provider capability; local versus hosted is a cost and privacy call, not a code change.

Testing without keys

Fake the vector path and assert on the prompts the code built:

use Crustum\Ai\Embeddings;
use Crustum\Ai\Prompts\EmbeddingsPrompt;

Embeddings::fake();

$response = Embeddings::for([
    'Napa Valley has great wine.',
    'CakePHP is a PHP framework.',
])->dimensions(1536)->generate();

echo count($response->embeddings) . " embeddings\n";

Embeddings::assertGenerated(
    fn (EmbeddingsPrompt $prompt) => $prompt->contains('CakePHP') && $prompt->dimensions === 1536,
);

2 embeddings

assertGenerated() / assertNotGenerated() / assertNothingGenerated() cover the prompt; assertQueued() covers queued generations. Fakes return shape-correct vectors, so ranking code runs offline. For a byte-stable loop see RAG. Queued variants mirror the same assertions with QueuedEmbeddingsPrompt.

See also

• Providers — Embeddings — generation, multimodal, caching, similarity
• Loading — files in
• Stores — where vectors wait