kaufKI: Chat with your supermarket brochures

About this project
kaufKI.com : Chat with this week's supermarket offers
Problem. Every supermarket prints its own weekly brochure. To find the cheapest milk or plan a budget-friendly dinner, you have to open each brochure and read every page. Because they are only in German, they are hard to use for expats, and printed paper brochures create unnecessary environmental waste.
Approach. kaufKI.com is an AI chat assistant that answers strictly using real brochure offers. An admin uploads a brochure PDF, a vision model reads every page, and each offer is saved in PostgreSQL after automated validation—while unclear offers are flagged for human review. The chat runs on Next.js and the Vercel AI SDK, calling tools that search offers by keyword and semantic meaning (pgvector), compare prices, plan meals, and build shopping lists. All prices and totals are calculated directly in SQL, never by the LLM. My first version used LangChain and LangGraph, which I later replaced with a single, simple tool loop.
Results. The app is live at kaufki.com in closed beta, supporting English, German, and French. Any response containing a price not verified in the database search results is blocked from being shown. Furthermore, letting the model filter which offer cards to display eliminated almost all irrelevant results in my RAGAS evaluation.
Next steps. Automate PDF ingestion and pursue retail partnerships for direct API access. I am also adding WhatsApp integration, price alerts for favorite products, location-based store filtering, and full admin user management.