AI Chat for Modified E-commerce

We connect your Modified Shop to the European provider MISTRAL, which is based in France. Of course, the AI speaks German—or rather, the languages that are set up in your shop. Alternatively, you can simply switch to Claude, the American AI from Anthropic, in the admin configuration. Both models come in several tiers that differ in quality, response speed, and cost.
(Spoiler: we're talking about one to three cents per reply)

What your customers see

A discreet speech bubble appears in the bottom right corner of every shop page. A click opens a chat window in your shop's design. Customers ask their questions in standard German or English—the plugin answers them directly, around the clock, without anyone having to be on the hotline.

The AI knows your product range, understands formal and informal queries, and consistently uses formal or informal address (depending on your preference). Product and contact links appear as clickable buttons, and phone numbers are displayed as call-to-action links so that mobile customers can dial with a single tap.

What the plugin does

Product recommendations based on your actual catalog. The AI searches product names, short descriptions, long descriptions, and product numbers — with linguistic flexibility, including German stemming and plural forms. “Do you have a linen skirt in size 46?” leads to the correct answer, even if the product is called “Miami Balloon Skirt, Linen, Black.”

• Variants with stock. Sizes, colors, and materials from your products_attributes are read. If a customer asks for size 54 and the item is listed in the size group “52/54 – 56/58,” the AI recognizes the coverage and honestly states whether this variant is currently in stock or sold out.
• Price surcharges. If a product has variants with additional costs, the base price is displayed with a “starting at” prefix and linked to the product page, rather than showing a potentially incorrect final price.
• FAQs from your CMS. Any content management system pages can be integrated. The AI answers questions about sizing, returns, cancellations, custom orders, or contact directly using your own content—not generic information.
• Language detection. If the shop session is in English, the AI responds in English; otherwise, in German. Brand loyalty included: no recommendations for third-party shops, no price comparisons with competitors.
• Secure handling of customer data. Questions about specific orders, invoices, or package tracking are immediately forwarded to your team. The AI does not ask for order numbers, addresses, or email addresses and does not process them if a customer sends them anyway.
• Protection mechanisms. IP allowlist for test and rollout phases, rate limiting per session and per IP, debug log can be disabled, all tool calls SELECT-based only.
• Self-service admin panel — no need to contact us. On the admin page „AI Chat – Content" you manage everything yourself: custom free-text instructions for the AI, a knowledge base with business-specific facts (e.g. minimum order value, special conditions), product categories and their mapping, FAQ topics from your CMS, plus the chat window's title and greeting text — every change takes effect immediately, no intervention from us needed (see screenshot). Maintaining product categories is optional: leave them empty and the AI simply searches the entire catalog without restriction — a configured category is only a bit more convenient for “show me all X” requests.
• Faster search by defining the most important categories. Example: you sell pet food and have every category duplicated — once as a tree sorted by animal type, once by food type. We can then look only at the animal-type tree and its subcategories, saving the search through the food-type tree.

Technical Overview

• Compatible with modified eCommerce 3.0.x and 3.3.x, PHP 8.1 or later
• Uses Mistral AI or Anthropic Claude, your choice.
• Clean integration via the modified hook system, no changes to the shop core
• Charset-aware: works in UTF-8 shops as well as in shops using ISO-8859-1/-15
• Operating costs: only the Mistral or Claude API fees for your own account. You can set a monthly limit for these in your Mistral or Claude account. Once the limit is reached, the AI responds that it cannot provide any information at the moment.

Here's how a request works

Between “customer types a question” and “answer appears in the chat,” three things happen—usually within three to four seconds:

• The question is sent to the shop. The plugin receives the message, checks rate limits, detects the session language, and forwards the question along with a brief system context (brand name, category list, behavioral rules, your own instructions) to the AI API.
• The model determines whether it needs data. If the query concerns a product, the model automatically calls the search_products tool; if it concerns returns, sizes, or contact, the get_shop_info tool; if it concerns business-specific knowledge, the search_knowledge tool. The plugin then performs a read-only query in the shop—using a multi-level fallback for the catalog: first the exact phrase, then a keyword search with stemming (plural forms, endings), and finally a similarity search. The best up to eight results, including price, category, delivery status, and variants (sizes, colors, material with stock levels per variant), are returned to the model.
• The model formulates the response. The model thus only ever sees genuine, up-to-date shop data and can summarize it in standard German, sort it, and include product links. For more complex questions, multiple tool rounds are possible (e.g., first the catalog, then the FAQ)—the model handles this itself within a specified upper limit.

Important to note: There is no pre-index, no external training, and no cached product data. Every request queries the store database in real time—ensuring that prices, inventory levels, and availability always match the store's data, without the need for synchronization. However, this means the response takes several seconds.

What the plugin deliberately does not do

• No access to customer accounts, orders, payments, or shipping status
• No integration with logistics or tracking systems
• No image generation and no outfit mockups
• No personalized recommendations based on a customer login
• No booking of appointments or callbacks; for this, the AI links to your contact form

This is a deliberate distinction. A chatbot that reads order data requires GDPR-compliant authentication, audit trails, and separate testing—these are not things that belong in an initial consultation.

A real-world example

Customer: I'm looking for a linen tunic, preferably in black. Do you have anything like that in size 54?

Chatbot: Yes, a nice option is the Samoa Embroidered Tunic Blouse in denim blue-sand in sizes 46–58 — we currently don't carry linen tunics in black, but we do have the Felicia Print Tunic in black (cotton jersey). The Samoa tunic covers size 54 in the 52/54–56/58 range and is in stock; the Felicia is available in 44/46–48/50 and 52/54–56/58, with size 54 in stock. Samoa Tunic Blouse · Felicia Print Tunic

Customer: Where is my order number 21438?

Chatbot: Order status, shipping, and payment can only be handled by our team — the fastest way is via our contact form or by phone at 040 123456789.

Two key features of this dialog: The AI scans your catalog in real time, recognizes what “size 54” means in your sizing chart, and simultaneously sets boundaries for sensitive data.

Implementation Process

• Assessment — we'll review your product range, your category structure, and your CMS pages together.
• You tell us your preferences: informal or formal tone?  Process only products, or also specific CMS pages?
• Configuration by us — category mapping, FAQ integration, brand text, shop color, widget positioning.
• Training and internal testing phase; we verify security and correct responses (e.g., deactivated products, sold-out products, price mentions).
• Test phase with IP allowlist — the widget is only visible to you; you check response quality and tone.
• Go-live — release to all visitors, monitoring via debug log or silent operation without a log file.

In total, it takes us at least 10–14 hours to get everything up and running, which is why the price is so high. By comparison, off-the-shelf chatbot solutions cost 80–300 euros per month—and the bot doesn't even really know your product catalog. This plugin, on the other hand, is a one-time purchase that gives you full control over your data, your own Mistral account, and is specifically tailored to your shop and product range — and fully self-maintainable afterwards.



 

Guide: How to configure the plugin yourself

Once we've set everything up, you can make all content changes yourself at any time — no need to call us. Here's how:

1. Set the provider & API key. In the shop admin, go to Modules → System → „AI Chat". Choose Mistral or Claude as the provider, enter your API key, and set the status to „true" to activate the module.
2. Custom instructions. Under „AI Chat – Content" (its own admin menu item) you can give the AI free-text guidance in the „Custom instructions" field — e.g. tone of voice or special notes. These add to the AI's responses but cannot override the built-in safety rules. Use the „Check instructions" button to have the AI pre-check your text for contradictions.
3. Role & widget texts. In the same section you define how the AI introduces itself (opening sentence), plus the chat window's title and greeting — separately for German and English. Leave a field empty and the plugin falls back to a neutral default text.
4. Product categories (optional). Enter category short names („slugs") and link them to the matching category ID from your shop (found in the admin under Catalog → Categories/Products in the URL, e.g. cPath=24). Without any entries the AI still searches the entire catalog — the mapping just makes “show me all X” requests a bit more elegant.
5. Info pages. Link CMS pages here (e.g. size chart, returns, about us) with a short name and the respective content ID, so the AI can fetch and link to them directly when relevant.
6. Knowledge base. For facts that live neither in the product catalog nor on a CMS page (e.g. minimum order value, specifics of your range), add your own entries here. A „Check entry" button is available for these too. IMPORTANT: The knowledge base is permitted to override the hard-coded guardrails. An example of this would be a scenario where we have a hard-coded rule stating "No advice on shipping costs," but you specifically want to provide advice on shipping costs.
7. Test phase. We initially enable the widget only for your IP address, so you can test in peace before it becomes visible to all visitors. By the way, you can also restrict the chat to specific customer groups, such as regular customers or B2B customers.