Client system · 2026 · AI & RAG, n8n automation, Integrations
Derb3Omar AI shopping assistant
A multilingual product and policy assistant that turns WooCommerce data and documents into grounded, inventory-aware answers.
Problem
Customers need useful product guidance across French, English, and Arabic without the assistant inventing prices, stock, or policy details.
My role
RAG architecture, n8n workflows, retrieval design, and web/WordPress integration
Stack
How the work moves
- Step 01Catalog + policiesWooCommerce and Drive sources
- Step 02IngestionChunking and multilingual embeddings
- Step 03RetrievalCosine similarity in pgvector
- Step 04AnswerGuardrailed LLM with product cards
The business problem
Product questions are rarely keyword searches. A customer may describe a use case, mix French with Arabic, or ask a policy question alongside an inventory question. The assistant needed to retrieve the right context before it answered and keep the result useful inside the store.
Architecture decisions
Two n8n workflows do the work: ingestion turns WooCommerce catalog data and Drive policy documents into multilingual vectors; retrieval embeds each question, searches Supabase pgvector, and sends grounded context to the language model. The delivery layer is a server-side Next.js proxy and a native WordPress plugin that both call the protected webhook without exposing its URL.
Structured sources are part of the response contract. Product title, URL, price, image, and stock state can render as a useful card instead of disappearing into prose.
Engineering focus
- Validate and limit both the new message and recent conversation history.
- Keep the webhook secret on the server and normalize upstream failures.
- Render assistant text as plain text and linkify URLs safely.
- Support bidirectional French/Arabic content and keyboard focus in the widget.