Download the PHP package yu-dev/module-ai-search-engine without Composer

Magento 2 Search Engine - Keyword & Semantic (Vector) Product Search

🔗 Demo: shop.yu.net.ua

An infrastructure module: structured product search that talks directly to the store's search engine (Elasticsearch or OpenSearch - whichever is configured in Magento), bypassing Magento's standard search-query-building layer, with optional semantic (meaning-based) matching blended in. It doesn't show anything on the storefront by itself and needs no admin configuration for keyword search - Yu_AiChat (product search inside the chat assistant) and Yu_AiCatalogSearch (AI-powered catalog search) are built on top of it.

Why it exists

Magento's own search layer (SearchInterface) behaves unreliably in some scenarios - when called from a context other than the search results page (for example, from an AI assistant's tool) - and doesn't always honor every filter (in particular, price range). Yu_AiSearchEngine works around this limitation by talking to the search index directly, while still reading its connection settings (engine address, index name) from Magento's standard configuration - so switching the search engine in the admin panel doesn't require reconfiguring this module separately.

Another reason search goes through a dedicated module instead of being assembled on the fly wherever it's needed: security. The call only accepts plain data (a query string, attribute-value pairs), never a ready-made search query as a whole. This rules out injecting arbitrary code into the search engine, even when the search parameters come from AI-generated text.

Features

• Keyword search - free text across every searchable attribute at once.
• Search by specific attributes - e.g. "color: red" and "material: cotton" at the same time, including the ability to exclude a value ("not red").
• Combined mode - keywords and attribute filters in a single query, with facet counts computed at the same time (how many products of which other color/size exist in the current result set) - this is what powers the refine-your-search suggestions in AI catalog search.
• Price, category and stock filters - built into every search mode.
• Automatic attribute discovery - the module figures out on its own which product attributes the merchant has actually configured as searchable or usable in layered navigation, and which of those are actually populated in the catalog - nothing needs to be listed manually.
• Semantic (vector) search, blended with keyword search. Free-text queries are matched both on literal words and on meaning - a search for "warm winter jacket" can find a product described only as "insulated parka", something keyword matching alone would miss. Off by default; when enabled, the two rankings are blended with configurable weights, and any embedding failure degrades gracefully back to today's keyword-only behavior rather than breaking search.

Semantic (Vector) Search

Vectors come from Yu_AiLlm's OpenAI embeddings provider (see that module's README) and live in a separate, per-store Elasticsearch index this module owns and keeps in sync via a native Magento indexer, AI Search Vector - visible in Index Management, supporting both "Update on Save" and "Update by Schedule" like any other Magento indexer.

In "Update on Save" mode, a product's vector is refreshed right after the product is saved or deleted in the admin panel, and after the mass "Update Attributes" / "Update Websites" actions. Computing a vector means a call to the embedding API, so a product save takes roughly a second longer; on large catalogs or frequent bulk edits "Update by Schedule" (done in the background by cron) is the better choice. Two safety rules apply to "Update on Save":

• Nothing happens until the index has been fully built at least once (the indexer shows "Ready") - so a store that hasn't set up semantic search yet doesn't try to call the embedding API on every save.
• If refreshing the vector fails (the API is unreachable, the search engine is down), the error is written to the log and the product is still saved - the vector index never blocks catalog editing. The product's vector simply stays as it was until the next save or full reindex.

How the two result lists are combined: keyword search finds the products that literally match the query and every attribute requirement (e.g. "color: green"); vector search finds the products closest in meaning. Vector matches are held to the same requirements - a product that is semantically close but has a different color is dropped, not blended in. A product that doesn't have the attribute at all is not treated as contradicting it, exactly as on the keyword side.

Turning it on:

1. Configure and enable an embedding-capable LLM provider under Stores → Configuration → AI → LLM AI (currently OpenAI).
2. Run a full reindex of AI Search Vector (bin/magento indexer:reindex yu_vector_search, or via Index Management in the admin panel).
3. Enable semantic search under Stores → Configuration → AI → Search Engine AI → Semantic Search (see Configuration below).

Until step 2 has completed at least once, the vector index doesn't exist and every hybrid search request quietly falls back to keyword only - turning on Semantic Search before reindexing doesn't break anything, it just doesn't do anything yet either.

What a product's vector is built from

A vector (an "embedding") is a list of numbers an AI model computes from a piece of text; two texts that mean similar things get similar numbers, which is what lets "jacket for rainy weather" find a product whose description never says "rainy". So the vector can only know what's in the text it was computed from. For each product, that text is assembled from, in this order:

1. Name
2. Category paths - the product's categories as breadcrumbs ("Parent > Child"), in the store view's language. Inactive categories are left out, and a parent category isn't repeated on its own when a deeper path already contains it.
3. Attribute values - one "Label: value, value" line per attribute, using the store view's labels (the raw option IDs stored in the database mean nothing to an AI model).
4. Short description and description - with HTML tags removed, so markup from the WYSIWYG editor or Page Builder doesn't pollute the vector.

The order matters: the embedding model accepts a limited amount of text, so a very long product text is cut from the end - the description loses its last paragraphs, while the name, categories and attributes always make it in.

Which attributes are included is configurable (see Configuration). The default, Automatic, takes the attributes you've marked Use in Search or Use in Layered Navigation that have dropdown, multiple-select or short text values. Yes/No attributes are left out by default - a line like "Label: No" carries little meaning - but you can pick them explicitly, in which case the product gets just the label, and only when the value is Yes. Text values made only of digits or codes are skipped, since they add noise rather than meaning.

A vector is a soft signal, not a filter: including a color attribute makes "red dress" lean towards red products, but it doesn't guarantee them. Exact requirements are enforced by the keyword side of the search and its filters, exactly as without semantic search.

Changing these settings, or renaming categories, only affects vectors after the next full reindex of AI Search Vector.

Requirements

• PHP >= 8.1
• Magento 2.4.x with Elasticsearch/OpenSearch configured (Magento's standard search engine); Elasticsearch 7.x's lack of native approximate kNN is fine at typical catalog sizes - vector scoring runs as a brute-force script_score query
• For semantic search on OpenSearch: the k-NN plugin (opensearch-knn), which ships with the standard OpenSearch distribution and Docker image. Elasticsearch and OpenSearch store vectors differently (Elasticsearch calls the field type dense_vector, OpenSearch calls it knn_vector and provides it through that plugin); the module picks the right one from the search engine configured in Magento. After switching the store from one engine to the other, run a full reindex of yu_vector_search so the vector index is rebuilt in the new engine
• The Yu_AiLlm module (installed automatically as a dependency), needed only for semantic search - keyword search has no LLM dependency at all

Installation

Keyword search needs no admin configuration and works immediately. The module doesn't do anything on the storefront by itself - it's used by other modules (Yu_AiChat, Yu_AiCatalogSearch), which are installed separately. Semantic search is opt-in - see the Semantic (Vector) Search section above.

Configuration

Stores → Configuration → AI → Search Engine AI → Semantic Search

• Enabled - the kill switch for hybrid search. Off (default) falls back to keyword-only everywhere.
• Keyword Score Weight / Vector Score Weight - blend weights, default 0.4 / 0.6. Each score is normalized to [0,1] independently before blending, so the two weights aren't required to sum to 1.

Stores → Configuration → AI → Search Engine AI → Vector Index Content

See "What a product's vector is built from" above.

• Include Category Paths - Yes (default) adds category breadcrumbs.
• Include Attributes - Automatic (default), Selected attributes only (pick them in the Attributes list that appears), or None (name, categories and descriptions only).

Author

Yuriy Akishin:

• 📧 Email: [email protected]
• 💼 LinkedIn: https://www.linkedin.com/in/yuriyakishin/
• 💻 GitHub: https://github.com/yuriyakishin

License

MIT