Chatbot on your knowledge base. It answers citing the document, or it doesn't answer.

The same fifteen questions, every day, about the same product sheets and the same policies. This piece answers by drawing on your documents, puts the document it comes from next to the answer, and when the answer isn't there it says so.

This is the page for a single piece of the system. The other pieces, and the criterion for choosing which to start from, are on the services page.

The problem is the fifteen questions that come back every day

In every business there is a small set of questions that arrives constantly, always the same, and the answer is already written somewhere. The cost isn't the single answer, it's that a person looks for it every time in a different file and in the meantime has stopped doing something else.

The second cost only shows later: whoever answers goes from memory, and the memory of two different people gives two different answers. The trivial question becomes a contradiction in front of a customer, and it is the kind of error nobody records anywhere.

A question repeated fifteen times a day isn't a problem of patience. It's a document nobody has yet put where it's needed.

What changes in practice

The recurring questions stop reaching a person. Whoever asks gets the answer straight away, with the document it comes from next to it, and whoever used to answer finds their day free of those fifteen interruptions.

The second change is that there is only one answer. Everyone receives what is in the approved documents, not what a person remembers reading, and contradictions in front of the customer disappear because there is a single source.

The third is information nobody has ever had: the list of questions the assistant couldn't answer. It is the most honest way to know which documents are missing, and it arrives without asking anyone.

Retrieval from documents reduces invention, and doesn't eliminate it

It is the point where people buy badly, and there is a measurement. Stanford's RegLab group carried out the first preregistered evaluation of three legal research tools built precisely on document retrieval, and sold as hallucination-free.

The result is stated word for word in the paper: «each hallucinate between 17% and 33% of the time». The same authors state that hallucinations are reduced compared with a general-purpose assistant, which is exactly the point: reduced, not removed.

On a simpler task the numbers improve a lot. The Vectara leaderboard, updated to 11 May 2026 on more than 7,700 documents, measures the best model at 1.8% of answers not faithful to the text provided, with 99.5% of questions answered.

Together the two measurements say the useful thing. Summarising a document you have in front of you is reliable; choosing on your own which document is needed and then answering is much less so. Almost all of the building work lies in that second step.

You decide the boundary, beforehand

During analysis you decide what the assistant can answer, what it must refuse and what it must pass to a person. It is a commercial choice, not a technical one, and it has to be made before writing a line: an assistant without a declared boundary will answer everything anyway.

Type of question	What the assistant does	What you get
Approved informationit is in the documents	It answers straight away, with the document and the point it took the answer from next to it.	An identical answer every time, verifiable in a second by whoever receives it.
Missing informationit isn't anywhere	It says it doesn't know, passes the question to a person and writes it on the list of things to add.	No invented answers, and a list that says which documents are really missing.
Commitment for the businessprices, discounts, confirmations	It never answers on its own. It collects the request with its context and queues it for whoever decides.	No figure promised by a machine, and whoever decides receives the request already prepared.
Off-topicnothing to do with it	It refuses and brings the conversation back within the boundary, without chasing the question.	The assistant stays yours, and doesn't become a toy for those who try to derail it.

What the system does, step by step

Every question goes through the same five steps, and each one leaves a log entry with the question, the documents retrieved and what was answered. That log is why an error gets found, instead of being recounted from memory.

Step	What happens	What you get
The baseapproved documents	Sheets, terms, policies and past frequently asked questions go into the base, each with the date of its last check.	Answers come from documents someone has approved, and you know when.
Retrievalwhich pieces are needed	When the question comes, the system retrieves the relevant passages and puts them in front of the model, which answers only on those.	The answer rests on a text of yours, instead of on what the model remembers about the world.
Citationwhere it comes from	Next to the answer appear the document and the exact point, clickable for anyone who wants to check.	Checking costs a second, and an error shows before it becomes a problem.
Refusalwhen it doesn't know	If the retrieved documents don't contain the answer, the assistant says so and passes the question to a person.	The worst case becomes a delay, instead of wrong information given with confidence.
Verificationfixed questions	A set of real questions with the right answer next to them is rerun at every change to the system.	A deterioration shows immediately, instead of being discovered by a customer three months later.

The knowledge base is the product; the model is a spare part

An assistant is worth as much as the documents underneath it. If the terms of sale exist in three versions and nobody knows which is right, the assistant will repeat the wrong one with the same confidence it would have repeated the right one.

That is why the work starts from the documents and not from the technology: you choose a good version for each topic, put a date next to it, and decide who updates it. It is the same principle as the base of facts that sales copy and materials works on, and it is best to keep just one for both.

The choice between writing the information directly into the instructions and building a retrieval system is made on volume and update frequency. A few stable pages need no infrastructure, and adding it would be a cost that buys nothing.

The assistant states that it is a system

From 2 August 2026 Article 50 of Regulation (EU) 2024/1689 applies, which requires providers to ensure «that AI systems intended to interact directly with natural persons are designed and developed in such a way that the natural persons concerned are informed that they are interacting with an AI system».

The statement comes before the conversation and takes up one line. What it really means for a small business, with the scope and the four things to put in order, is on the page about Article 50 of the AI Act for Italian SMEs.

In practice it works as a sales argument. Whoever discovers on their own that they were talking to a machine stops trusting even what was true, and the opening line removes that risk instead of adding one. The full reasoning is on the page about the principles we build with.

What goes out on its own and what waits for a person

Everything that repeats already approved information goes out on its own: opening hours, availability, the content of a sheet, written terms, internal procedures. They are things decided beforehand by a person, and the system simply reports them, citing where they are.

Whatever commits the business never goes out on its own. A price, a discount, a confirmation and a delivery promise are collected with their context and queued for whoever decides. Requests that arrive from several channels and need sorting are a different piece, handled by Inbox AI, the first reply to every request.

The same thing changes name with the trade

The mechanism is identical; the knowledge base isn't. It is worth looking at your own case, because that is where you see which document should be put in order first.

Sector	What is in the base	Where you see it
Food and agricultureand export	Technical sheets, allergens, formats and pieces per carton, delivery terms. Questions a buyer asks before asking for a price.	The typical day of a small producer who exports
Restaurantsand bars	Updated menu, allergens, opening hours, rules for groups and pets. Questions that arrive in bursts in the hour when nobody can answer.	The service of a dining room where the phone rings unanswered
Hospitalityaccommodation and events	House rules, what is included, hall capacities, arrival and departure times, nearby services.	The pre-season of an accommodation business

What this piece doesn't do

It doesn't promise zero errors, and be wary of anyone who does: the Stanford measurement exists precisely because three large providers had promised it. What gets built is a system in which the error shows, is cited and gets corrected, with a set of fixed questions that keeps it under control over time.

It doesn't replace whoever answers. It removes the repetitive questions and lets those that need judgement reach a person, with more context and sooner.

It doesn't learn on its own from conversations. What goes into the base goes in because someone approved it, and a wrong answer given yesterday doesn't become tomorrow's truth.

Questions and answers