How an AI assistant answers only from your knowledge base — and what happens when it doesn't know
Short answer: the assistant does not “know” your business; it is handed the three or four passages from your own documents that best match the question, and told to answer from those or admit it cannot. Code rules check the reply before it leaves. Questions it could not answer are collected so you can fill the gap once.
The question every business owner asks first is some version of “will it say something we didn't approve?”. It is the right question. This is how we answer it in the systems we ship, in plain language.
Step one: find, don't recall
When a customer asks “are you open on Saturday?”, the model is not asked to remember your opening hours. The system searches your knowledge base — the documents you gave it — for the passages most similar to the question, and puts those passages in front of the model along with the question. The model's job becomes reading comprehension, not memory. If the search returns nothing relevant, the model is told so.
Two details matter. The search is scoped to your business at the database level, so another customer's documents are never in the pool. And configured facts — working hours, holidays, services you do not offer — are injected directly from settings rather than searched, because a search can miss and a setting cannot.
Step two: the instruction, and why it is not enough
The model is instructed to answer only from the material it was given, to say “I don't have that information” otherwise, and never to invent prices, recovery times or directions. This instruction works most of the time. Most of the time is not good enough for something that talks to your customers unattended.
So every reply passes through rules written in code before it is sent. Some examples from the receptionist we build: a reply that quotes a price when no price exists in the knowledge base is rejected; a reply that offers to describe transit directions is replaced with a map link, because the assistant has no transit data and would either stall or invent; a reply that ends without inviting a response is corrected, because we measured that 38% of unassisted replies did exactly that. In testing, the polite instruction alone failed repeatedly on each of these. The code rule did not.
Step three: what “I don't know” triggers
Admitting ignorance is only useful if something happens next. When the assistant cannot answer, it takes the customer's contact details so a person can follow up, and the question is logged. The business sees a running list of unanswered questions, grouped by topic, with a button to write the answer. Once written, the next customer who asks gets it. This loop is the reason the assistant is better in week four than in week one, and it needs no engineer.
What the knowledge base should contain
- Services you offer, and explicitly the ones you do not.
- Hours, address, how to reach you, what to bring.
- The five questions you answer by phone every day, with the answer you actually give.
- What must go to a person: complaints, refunds, anything medical or legal.
We import your website and structure your notes, then ask only for what is still missing. We do not write answers for you. A fact we invented would reach a real customer in your name, and no amount of engineering makes that acceptable.
If you want to see this on your own documents, send us your site and one page of notes; we will load them and let you ask the assistant anything on a 20-minute call.
Frequently asked questions
What does it mean for an AI to answer from a knowledge base?
Before the model writes anything, the system searches the business's own documents for the passages most relevant to the question and gives the model only those. The model is instructed to answer from that material and to say it does not know when the material does not cover the question. This is usually called retrieval-grounded generation.
Can it still make things up?
Any language model can. That is why the instruction alone is not enough: we add rules in code that check the output — no prices unless a price exists in the knowledge base, no medical assessments, no promises the business has not written down. In our testing, a polite instruction failed repeatedly where a code rule did not.
What happens when a customer asks something that is not in the knowledge base?
The assistant says it does not have that information, takes the customer's details so the team can follow up, and the question is logged. The business sees a list of unanswered questions and can add the answer in one place; the next customer gets it.
Who writes the knowledge base?
The business does, with our help: we import the website, structure the notes, and then ask only for what is still missing. We never write an answer on the business's behalf — a made-up fact would go straight to a real customer.
How is it kept isolated between customers?
Every document and every conversation carries the business's identifier, and the search is scoped to it at the database level. One business's assistant cannot see another's documents, by construction rather than by policy.
— Written by the Bulsu Labs team · Talk to us