AI Shopping Assistant

No. The knowledge base is stored in your CS-Cart database and searched in PHP. The only outside connection is the one to your AI service (or none, if the model runs on your own server).

Any service with an OpenAI-compatible chat completions and embeddings API: OpenAI, Azure OpenAI, Mistral, Groq (chat only, pair it with another service for embeddings), OpenRouter (chat only), and local servers such as Ollama or LM Studio. The add-on was tested end to end with a local Ollama server (llama3.2:3b for chat, nomic-embed-text for embeddings).

The add-on has no usage fees. Your AI provider bills you per token; see What it costs you. The Overview page counts tokens and estimates the cost; a daily token or cost budget stops new answers for the rest of the day (shoppers see your fallback text).

That depends mostly on the chat model and on your content. In our test with a small 3-billion-parameter model running locally and a demo catalogue of 257 products, the assistant answered product, price, stock and policy questions from the right sources, quoted live prices correctly, declined unrelated questions and did not invent products or prices for items the store does not sell. Weaker points of a small model: it sometimes describes a product too loosely (for example calling a game for one console "for" another console) and lists more or fewer products than ideal. A larger model of your provider gives more natural and more careful answers. Check the Conversations page, the ratings and the unanswered questions, and add FAQ entries where answers are weak.

Product cards always show the live price for the shopper (with user group prices and promotions). In the text, every amount with a currency mark is compared with the prices and numbers in the sources and in the question; with Remove sentences with prices that are not in the store data on (default), a sentence with any other amount is removed and the shopper is pointed to the product card. Amounts written without a currency mark are not checked.

Yes, at the moment of the question: stock and prices are read live from the store for every answer, never from the knowledge base.

The shopper sees your fallback text and a Contact us link, and the question appears under Unanswered questions so you can add an FAQ entry.

The model gets no tools and no private data: it cannot read orders, customers or settings, and it cannot place orders. Typical instruction attacks get a fixed answer without reaching the model, the built-in rules tell the model to treat sources and messages as data, and rate limits, message length limits and a daily budget limit abuse. Flagged messages are marked in the history.

It is built for it: the page output is the same for every visitor and the conversation runs over AJAX requests that are never cached. It has not been tested on a store with the full page cache switched on yet.

The search is plain PHP. Measured on PHP 8.3 with 768-dimension vectors: about 0.2 s for 4,000 chunks (exact comparison), 0.2 s for 20,000, 0.4 s for 50,000 and 1 s for 100,000 chunks (fast pre-filter, then exact comparison of the best candidates), using about 6 to 50 MB of memory. A product usually takes 1 or 2 chunks, so 50,000 chunks are roughly 30,000 products. Above that the search still works but gets slower; see Developer notes.

The shopper's message, the last few messages of the conversation and the matching store information are sent to your AI service to write the answer. Order lookups, names, e-mail addresses and IP addresses are never sent. Choose a provider whose data terms suit you, or run the model on your own server.

Not in this version. See the roadmap.

Without a valid license (no key after the 14-day start period, an invalid or revoked key, a domain the license does not cover, or no successful check for 14 days), the chat widget disappears from the storefront and the knowledge base stops syncing; your settings, knowledge base, FAQ entries, documents and conversation history are kept. Once the license is valid again, everything works again.

Open Add-ons → Manage add-ons → AI Shopping Assistant → Settings → License. The status line says what is wrong: no key, a key for another add-on, a revoked license, or a domain that the license does not cover. Fix it and click Check now. See License.

Yes. Subdomains of your licensed domain (for example staging.yourstore.com) and local development hosts (localhost, private IP addresses, .local, .test) are covered by the same license.

No. The check runs once a day in the admin panel; the storefront only reads the stored result.