A grounded AI chatbot answers from a specific, named body of content: a help centre, a policy folder, a set of product guides. It does not pull from the open web or from whatever its underlying model absorbed in training. When a question arrives, the bot retrieves the most relevant passages from that content, writes its reply from them, and shows which document it used. And when the content holds no answer, a properly grounded bot says so rather than improvising.
The word exists because the default is the opposite. A general-purpose model will answer almost anything fluently, whether or not it knows; ask an ungrounded support bot about your returns policy and it may compose one that has never existed. Grounding is the set of mechanics that prevents this: restrict the source material, cite it, and refuse when it runs out. Our earlier note on what “grounded in your docs” really means covers the promise — this article is about the machinery behind it.
How a grounded AI chatbot works
Beneath the widget, most grounded chatbots share the same shape. Your content is split into passages and indexed. Each visitor question is matched against that index and the best few passages are pulled out — Serochat retrieves at most six, never more than two from the same document, so one long PDF cannot crowd out everything else. The language model is then instructed to answer from those passages alone, and the reply carries a citation naming the document it drew on.
That retrieval-then-generation step is what the industry calls RAG, and it matters, but it is not what earns the word “grounded”. Instructions to a model are probabilistic: told to stay inside the passages, it will usually comply and occasionally will not. Grounding you can trust is a property of the whole system, not of the model inside it — it depends on what happens after the model answers.
What happens after the model answers
Serochat treats the model’s citation as a claim to be checked. When a reply names its source, that name is verified against the passages actually retrieved; if it cannot be verified, the answer is discarded and the visitor is offered a human instead, because no reply is worth a footnote the system cannot back. A second check compares wording — an answer sharing less than a fifth of its content words with the retrieved passages is replaced with the handoff message too. For any question that draws on your content, what reaches the customer is either an answer with a checkable source beneath it, or a plain admission and a Talk to a human button.
What grounding doesn’t fix
Be wary of anyone selling grounding as the end of wrong answers. It removes the largest source, a model free-associating from its training data, but three failure modes remain.
Retrieval can miss. A question phrased in words your documents never use may surface the wrong passages, or none. Serochat’s retrieval is keyword full-text search, so we live with this limit knowingly: the bot hands off rather than stretches, and the unanswered question lands in a training queue so you can write the missing answer once.
The documents can be wrong. A grounded bot repeats your out-of-date delivery page faithfully, citation and all. Grounding moves the quality problem from the model, which you cannot edit, to your documents, which you can — an improvement you still have to collect by keeping content current.
And the model can drift. Even with the right passages in front of it, a model sometimes writes beyond them; that is why post-answer verification exists. If a vendor’s grounding story ends at “we use RAG”, that drift lands in front of customers unchecked.
How to tell if a chatbot is actually grounded
“Grounded” appears in a great deal of chatbot marketing; no badge certifies it. You can test it from the outside in a few minutes.
- Ask something your documents clearly answer, and check the reply names its source: a citation you can open is evidence; an answer without one is a claim.
- Ask something your documents do not cover. The right behaviour is a plain admission and a route to a human; a fluent, confident answer here is disqualifying.
- Edit a document, re-sync, and ask again. A grounded bot’s answer should change when its source does.
- Ask the vendor what happens when the model cites a document that was never retrieved. If nothing checks for that, the footnotes are decoration.
Common questions
Does grounding stop hallucinations completely?
No, and it is worth distrusting any vendor who says otherwise. Grounding removes the biggest source of fabrication, and verification after the answer catches much of what remains. The honest claim is fewer, easier-to-catch wrong answers — never zero.
Is “trained on your data” the same as grounded?
Usually not. That phrase often describes fine-tuning, where your content nudges the model’s internal weights. A fine-tuned model still answers from memory: it cannot cite a page, and correcting it means training again. A grounded bot keeps your content outside the model and reads it at answer time, which is why it picks up this morning’s policy edit at the next sync, and why you can check its references.
Do I need a vector database for a grounded chatbot?
No. Vector search can improve retrieval, especially when customers use words your documents avoid, but it is a retrieval upgrade rather than what makes a bot grounded. Serochat ranks passages with plain full-text search, and we say so. What makes it grounded is the constraint, the citation, and the willingness to hand off when retrieval comes up empty — judge any vendor on those three, whatever the search technology underneath.
A grounded AI chatbot, then, is one that shows where its answers came from and admits when it has nowhere to point. The definition is short; the discipline is in the checking. Hold every vendor to it, ours included.