A chatbot that makes up delivery times or order status doesn't help: it does harm. And it's a more common problem than it seems, because most language models are designed to sound convincing, not to tell the truth. If they don't have the data in front of them, they fill the gap with something plausible. In an online store, that "plausibility" turns into angry customers and support tickets that could have been avoided.
The technique that solves this is called RAG, short for Retrieval-Augmented Generation, and it's what we use when we build conversational assistants for e-commerce. The underlying idea is simple, even if the name sounds technical: instead of asking the model to "remember" everything it knows about your business, we first retrieve the relevant chunks from a source of truth —your catalogue, your policies, the real status of orders— and only then ask it to write the answer grounded in them. The assistant still speaks naturally, but everything it states is anchored in real, verifiable information. And when the data doesn't exist, it says so rather than improvising.
Where it makes the difference
In a store, this completely changes the experience. When a customer asks where their order is, the assistant looks up the real status instead of guessing a date. When someone describes what they need in their own words, that description becomes a semantic search over the catalogue, so it finds the product even if they don't use the exact name. And when the question is about delivery times, returns or terms, the answer cites the right source instead of repeating a myth floating around online.
What really matters
The key to a good e-commerce assistant isn't that it "speaks well", but that it tells the truth. That depends less on the model you pick and more on the quality of the knowledge base behind it: if it's well-structured and up to date, the answers are solid; if it's messy, not even the best model fixes that. That's why we put so much care into the system's limits —no source, no made-up answer— and into traceability, so we can always know where each piece of data came from.
You don't need to rewrite your whole store to get started. It almost always pays to scope a first concrete case, like order status, confirm it genuinely works, and grow from there. If you want to see how it would fit your business, we'll show you a prototype before committing to anything.