Magento AI Search

An AI provider-agnostic semantic search solution
for Magento, Mage-OS, and Adobe Commerce.

See the User Guide for complete installation and configuration instructions.

See the Cost Analysis to learn how this module can save up to $29,993 per year.

💡 Motivation

What if there was a module using Magento's default search engine to provide actual meaning-based AI search without breaking the default one? This is that module.

Inspired by Google Cloud's Commerce AI Search, the module reads and chunks product descriptions, uses a remote AI server to convert the text chunks into vectors, then saves those vectors to Magento's default search engine, OpenSearch. This enables intelligent search based on meaning rather than keyword or synonym matching.

💰 Cost Analysis

This chart compares estimated first-year direct vendor costs for a catalog of 100,000 products in two languages with 1 million submitted searches per month. The assumptions, calculations, pricing sources, and limitations are documented in the Cost Analysis.

The module bars appear almost flat because the managed services cost hundreds or thousands of times more on the same linear scale.

🛍️ Use Cases

Customer product discovery

Semantic search is most useful when shoppers describe a need, experience, or intended use instead of entering the exact words used in the catalog. It works especially well when product descriptions and dynamic documents contain rich, specific information.

Catalog content quality review

Admin semantic-search testing can reveal weak product descriptions by showing whether representative shopper queries find the expected products and text chunks. This is a content-quality signal, not a measure of marketing effectiveness, which still requires conversion analytics or A/B testing.

The module produces stronger results when indexed product content clearly describes the benefits, attributes, and use cases that shoppers are likely to search for.

🔍 What Semantic / AI Search Is

🏗️ Architecture

• Magento's indexer detects product changes, reads the selected store-scoped content, splits it into chunks, and saves the documents and chunks locally.
• Scheduled workers send pending chunks to the AI server in batches, receive their vectors, and publish them to a versioned index in Magento's OpenSearch service.
• When a shopper searches, the AI server converts the query into a vector and OpenSearch finds products with similar meaning. Magento continues to handle the catalog query and result page.
• If semantic search is disabled or unavailable, the request falls back to Magento's default search.

Reindexing workflow

Only key configuration changes trigger automatic full reindexing. The Admin UI flags these settings with tooltips for administrators.

⚙️ System requirements

Distribution

Magento

OpenSearch must include the k-NN plugin, which is bundled with the standard OpenSearch distribution normally used with Magento.

Storefront

AI server

📦 Install

Complete installation steps are available in the User Guide.

🔧 Settings

The module settings are available in Stores > Settings > Configuration > AI Search.

The module dashboard is available in System > AI Search > Dashboard.

Complete configuration instructions are available in the User Guide.

📊 Performance

Catalog Scaling

The estimates are based on a cron job scheduled to run every 60 seconds, with 100 chunks per embedding batch, up to 3 concurrent embedding requests, and a maximum worker runtime of 600 seconds. Each generated description contained approximately 1,500 estimated tokens and produced around 5 chunks. The chunks averaged about 300 estimated tokens, with a maximum size of 350 tokens and an overlap of 50 tokens.

The measurements show predictable linear scaling under the tested configuration, proving that the worker architecture can support larger, long-running catalog workloads.

These estimates describe an initial full ingestion or a deliberate full rebuild of the product content used for search. In normal operation, only changed content is processed, and product descriptions usually change far less often than prices or inventory. Processing an entire catalog is therefore an occasional workload, such as during initial setup, major content campaigns, or embedding configuration changes, rather than a daily requirement.

Time Distribution

The processing configuration can be tuned further for larger catalogs. However, most of the computational work is performed by the AI server. These measurements used a local development AI server; production servers typically scale further.

📚 Documentation

• User Guide
• Cost Analysis
• Solution Discovery
• Contributing Guide
• Security Policy
• Code of Conduct
• MIT License