Embed plugin

Turn text into vectors. Cosine similarity. Compact storage. The plumbing under the RAG plugin, also useful on its own.

Stable Reading time ~ 2 min Edit on GitHub

The Embed plugin is a thin wrapper around Ollama’s /api/embeddings plus three useful helpers: cosine similarity, compact base64 packing, and inverse unpacking. The RAG plugin uses it internally, but it’s also useful on its own — for clustering, deduplication, semantic search, anomaly detection.

API

$embed = (new \Nibiru\Module\Ai\Ai())->embed();

$vec     = $embed->one('controller');                  / float[]
$vectors = $embed->batch(['a', 'b', 'c']);             / float[][]

$score   = \Nibiru\Module\Ai\Plugin\Embed::cosine($a, $b);     / 0..1
$packed  = \Nibiru\Module\Ai\Plugin\Embed::pack($vec);          / base64 string
$vec     = \Nibiru\Module\Ai\Plugin\Embed::unpack($packed);     / back to float[]

Pattern: deduplicate near-duplicate strings

$embed = $ai->embed();
$candidates = ['How do I create a module?',
               'How can I make a new module?',
               'What is MMVC?'];

$vecs = $embed->batch($candidates);
foreach ($vecs as $i => $a) {
    foreach ($vecs as $j => $b) {
        if ($i >= $j) continue;
        $sim = \Nibiru\Module\Ai\Plugin\Embed::cosine($a, $b);
        if ($sim > 0.9) {
            echo "Near-dup: {$candidates[$i]} ≈ {$candidates[$j]} ($sim)\n";
        }
    }
}

Pattern: semantic tagging

$tags = ['authentication', 'forms', 'database', 'modules'];
$tagVecs = array_combine($tags, $embed->batch($tags));

function bestTag(string $text, array $tagVecs, $embed): string {
    $tv = $embed->one($text);
    $best = ['_unknown', -INF];
    foreach ($tagVecs as $tag => $vec) {
        $s = \Nibiru\Module\Ai\Plugin\Embed::cosine($tv, $vec);
        if ($s > $best[1]) $best = [$tag, $s];
    }
    return $best[0];
}

echo bestTag('User::isAuthorized', $tagVecs, $embed);  / → 'authentication'
echo bestTag('Pageination::setTable', $tagVecs, $embed); / → 'database' (probably)

Storage

Embeddings are float arrays — typically 768 floats for nomic-embed-text, 1024 for mxbai-embed-large. That’s 3 KB or 4 KB per vector raw.

Use Embed::pack() to base64-encode them as 4-byte floats:

$compact = Embed::pack($vec);    / ~4 KB → ~5.3 KB base64 string
$vec     = Embed::unpack($compact);

The RAG plugin uses this format internally for its JSON files.

Embedding model choices

Pull on your Ollama instance once:

curl https://your-ollama-host.example/api/pull -d '{"name":"nomic-embed-text"}'   # 768 dim, default
curl https://your-ollama-host.example/api/pull -d '{"name":"mxbai-embed-large"}'  # 1024 dim, higher quality

In ai.ini:

[AI]
embed.model = "nomic-embed-text"
embed.dim   = 768

Common pitfalls

Calling one() in a tight loop. Each call is one HTTP round-trip. For >100 items, prefer batch() (still serial under the hood, but with consistent error handling).
Storing raw float arrays in JSON. Use pack() for ~5x smaller files and faster parse.
Comparing cosine to a fixed threshold. Different embedding models have different “similar” baselines. Don’t hard-code 0.85 — calibrate per model.