shopbuilder / app / AI / Embedding / DataEmbedding.php

DataEmbedding.php in ShopBuilder – WooCommerce Builder For Elementor 3.2.0, at app/AI/Embedding/DataEmbedding.php

152 lines 4.8 KB

Raw Download Zip

1	<?php
2	/**
3	* Class DataEmbedding
4	*
5	* Handles the generation, storage, and retrieval of AI-powered text embeddings
6	* for WooCommerce products using supported AI services like OpenAI, Gemini,
7	* or DeepSeek. Also provides semantic search capabilities using cosine similarity.
8	*
9	* @package RadiusTheme\SB\AI\Embedding
10	* @since 1.0.0
11	*/
12
13	namespace RadiusTheme\SB\AI\Embedding;
14
15	use RadiusTheme\SB\AI\DB\AIDB;
16	use RadiusTheme\SB\AI\AIFns;
17	use RadiusTheme\SB\Traits\SingletonTrait;
18
19	if ( ! defined( 'ABSPATH' ) ) {
20	exit( 'This script cannot be accessed directly.' );
21	}
22
23	/**
24	* Class DataEmbedding
25	*
26	* @package RadiusTheme\SB\AI\Embedding
27	* @since 1.0.0
28	*/
29	class DataEmbedding {
30	/**
31	* Singleton Trait.
32	*/
33	use SingletonTrait;
34
35	/**
36	* Generate and store an AI embedding for a product.
37	*
38	* @param int $product_id The product ID.
39	* @param string $title The product title.
40	* @param string $content The product content or description.
41	*
42	* @return bool True on success, false on failure.
43	*/
44	public function generate_and_store( $product_id, $title, $content ) {
45	$text = $title . ' ' . wp_strip_all_tags( $content );
46	$ai_service = AIFns::initializeAIService();
47	$embedding = $ai_service->generateEmbedding( $text );
48	if ( empty( $embedding ) \|\| ! is_array( $embedding ) ) {
49	return false;
50	}
51	$info = [
52	'word_count' => str_word_count( $text ),
53	'source' => 'product',
54	];
55	$result = AIDB::upsert_embeding( $product_id, $title, $embedding, $info );
56	return ! empty( $result );
57	}
58
59	/**
60	* Perform a semantic search based on a given query.
61	*
62	* @param string $query The user search query.
63	* @param int $limit Optional. Number of results to return. Default 0 (all).
64	*
65	* @return array List of matching product titles.
66	*/
67	public function search( $query, $limit = 0 ) {
68	$ai_service = AIFns::initializeAIService();
69	$query_embedding = $ai_service->generateEmbedding( $query );
70	if ( empty( $query_embedding ) \|\| ! is_array( $query_embedding ) ) {
71	return [];
72	}
73	// Log query embedding.
74	$rows = AIDB::get_all();
75
76	if ( empty( $rows ) ) {
77	return [];
78	}
79	$results = $this->find_similar( $query_embedding, $rows, $limit );
80	return wp_list_pluck( $results, 'post_id' );
81	}
82
83	/**
84	* Find the most semantically similar embeddings using cosine similarity.
85	*
86	* Optimized to reduce redundant computations and use a dedicated
87	* score calculation method for better maintainability.
88	*
89	* @param array $embedding The query embedding vector to compare.
90	* @param array $rows The stored embedding records from the database.
91	* @param int $limit Optional. Number of top matches to return. Default 5.
92	*
93	* @return array List of matched items with product ID, title, and similarity score.
94	*/
95	public function find_similar( array $embedding, array $rows, int $limit = 5 ): array {
96	$minimum_match = AIFns::get_embedding_minimum_accuracy();
97	$query_norm = sqrt( array_sum( array_map( static fn( $x ) => $x * $x, $embedding ) ) );
98	$matches = [];
99	foreach ( $rows as $row ) {
100	if ( empty( $row['embedding'] ) ) {
101	continue;
102	}
103	$vector = maybe_unserialize( $row['embedding'] );
104	if ( ! is_array( $vector ) ) {
105	continue;
106	}
107	// Avoid redundant norm computation if similarity calc includes it.
108	$score = $this->calculate_similarity_score( $embedding, $vector, $query_norm );
109	if ( $score >= $minimum_match ) {
110	$matches[] = [
111	'post_id' => isset( $row['product_id'] ) ? (int) $row['product_id'] : 0,
112	'post_title' => isset( $row['title'] ) ? sanitize_text_field( $row['title'] ) : '',
113	'score' => round( $score, 4 ),
114	];
115	}
116	}
117	if ( empty( $matches ) ) {
118	return [];
119	}
120	// Use array_multisort for faster sorting on large datasets.
121	array_multisort( array_column( $matches, 'score' ), SORT_DESC, $matches );
122	return $limit > 0 ? array_slice( $matches, 0, $limit ) : $matches;
123	}
124
125
126	/**
127	* Calculate cosine similarity score between query and stored embedding.
128	*
129	* This version avoids redundant normalization by reusing the precomputed
130	* query norm and computes the dot product and target norm in one pass.
131	*
132	* @param array $query_vec The query embedding vector.
133	* @param array $target_vec The stored embedding vector.
134	* @param float $query_norm Precomputed L2 norm of the query vector.
135	*
136	* @return float Cosine similarity score (0.0–1.0).
137	*/
138	protected function calculate_similarity_score( array $query_vec, array $target_vec, float $query_norm ): float {
139	$dot = 0.0;
140	$normB = 0.0;
141	$len = min( count( $query_vec ), count( $target_vec ) );
142	for ( $i = 0; $i < $len; $i++ ) {
143	$dot += $query_vec[ $i ] * $target_vec[ $i ];
144	$normB += $target_vec[ $i ] ** 2;
145	}
146	if ( $query_norm <= 0.0 \|\| $normB <= 0.0 ) {
147	return 0.0;
148	}
149	return $dot / ( $query_norm * sqrt( $normB ) );
150	}
151	}
152