RAG

Retrieval-augmented generation grounds the model in your data: load documents, embed them, pull back the similar ones at prompt time, and feed them in as context. Three stages (process, retrieve, augment) and this plugin owns the seams between them.

The pipeline

flowchart LR
    load[Load documents] --> embed[Embed vectors]
    embed --> store[Store and index]
    store --> retrieve[Retrieve similar]
    retrieve --> augment[Augment prompt]

Process (load + embed) is Loading and Embedding; storage is Stores; retrieve is Retrieval; augment is just prompting with the winners attached.

The loop, offline

Embed three passages and a query, rank by cosine similarity, take the best: the whole retrieval idea with no provider involved beyond vectors:

php

use Crustum\Ai\Embeddings;

function cosine(array $a, array $b): float
{
    $dot = 0.0;
    $na = 0.0;
    $nb = 0.0;
    foreach ($a as $i => $value) {
        $dot += $value * $b[$i];
        $na += $value * $value;
        $nb += $b[$i] * $b[$i];
    }

    return $dot / (sqrt($na) * sqrt($nb));
}

$passages = [
    'Napa Valley has great wine.',
    'CakePHP is a PHP framework.',
    'The Eiffel Tower is in Paris.',
];

$vectors = Embeddings::for($passages)->generate()->embeddings;
$query = Embeddings::for(['CakePHP framework'])->generate()->first();

$best = 0;
$bestScore = -2.0;
foreach ($vectors as $index => $vector) {
    $score = cosine($vector, $query);
    if ($score > $bestScore) {
        $best = $index;
        $bestScore = $score;
    }
}

echo $passages[$best] . PHP_EOL;

text

CakePHP is a PHP framework.

Three ways to retrieve

Application tables: vectors in your own columns, similarTo finder, SimilaritySearch tool. Your SQL, your data, no uploads.
Provider stores: files uploaded once, searched provider-side with FileSearch. No vectors in your schema.
Search engine: Explorator with semantic/hybrid search over your models. Full-text and vectors fused.

The comparison lives under Retrieval; the storage side under Stores.

RAG ​

The pipeline ​

The loop, offline ​

Three ways to retrieve ​

See also ​

RAG

The pipeline

The loop, offline

Three ways to retrieve

See also