The Client
A car dealership whose buyers do their research at night, on their phones, long before they ever set foot on the lot. The questions are always the same — do you have this model, what does it cost, what financing do I qualify for, can I book a test drive — and they arrive at all hours, in far greater volume than a sales floor can answer.
The Problem
Generic chatbots fail at dealership work for one reason: the answers live in data that changes daily. Inventory moves, promotions rotate, financing terms shift. A bot that answers from a static script is confidently wrong within a week — and a wrong price quote is worse than no answer at all. The assistant had to ground every response in the dealership's actual, current knowledge base.
The Solution
We built a retrieval-augmented generation (RAG) pipeline: every answer is composed from documents retrieved out of the dealership's own knowledge base — company info, services, current promotions, financing programs, policies, FAQs, and a live inventory catalog — rather than from the model's imagination.
→ Embedding search (sentence-transformers over ChromaDB vector store)
→ Relevant knowledge retrieved (inventory, promos, financing, policies)
→ LLM composes grounded answer (Llama 3 via Groq)
→ Reply in the chat UI — typically under a second
- Structured knowledge base split by domain: inventory, promotions, financing, services, policies, FAQ
- Inventory enrichment scripts keep the vehicle catalog and summaries current
- Prompt-injection testing — the bot stays on script even against adversarial inputs
- Runs 24/7 at effectively zero monthly cost: free-tier LLM inference (Groq), free embeddings, in-memory vector store
- Handles inquiries, vehicle search, financing and promotion questions, and test-drive scheduling
The Stack
FastAPI serves both the chat API and the web UI. ChromaDB provides in-memory vector search over sentence-transformer embeddings, and Groq-hosted Llama 3 generates the answers. The whole system ships as one Docker container.
The Outcome
Every after-hours inquiry gets an immediate, accurate, on-brand answer grounded in the dealership's real data — and the sales team starts their morning with warm leads instead of an unanswered inbox. Because the stack runs on free-tier infrastructure, the assistant costs nothing to keep online. Try it live below.