A customer support platform using RAG — Retrieval-Augmented Generation. The client is a battery supplier for vehicles, caravans, boats, and yachts. They had a problem: over 5 years, about 20,000 support tickets had piled up in Zendesk, and the questions kept getting more complex.
The problem
Imagine a customer asking: "Can I charge two batteries with different amperage together using the company's original charger?" That's not a FAQ question. That requires knowing product specs, compatibility matrices, charging protocols. Previously, a support agent would spend 15-20 minutes digging through documentation and past tickets to give an accurate answer.
What I built
I imported the entire Zendesk database as question-answer pairs. Every resolved ticket became training data. The system uses Qdrant as a vector database — when a new question comes in, it finds semantically similar past tickets and their answers, then uses a language model to generate a precise response based on that context.
The RAG approach
RAG is the right tool here because the knowledge is very specific and constantly evolving. You can't just fine-tune a model once and forget about it — new products launch, specs change, edge cases appear. With RAG, the knowledge base stays fresh. New resolved tickets automatically become available for future answers.
The result
Response time dropped dramatically. What used to take a support agent 15 minutes now takes seconds. The support team went from 3 people to 1 — the remaining person handles edge cases and validates the AI's answers on tricky questions.
