AI chatbot on your data: answers from your documents and price list

Two layers of knowledge in an AI chatbot

An answer is built from two layers. The first is a knowledge block in the bot settings: short reference facts about the company, rules, tone of voice. For client bots it takes roughly one to eight thousand characters. The second layer is a search over uploaded documents: the AI chatbot finds the relevant fragments and answers on their basis.

Small reference data is convenient to keep in settings, bulky material goes into the knowledge base. This shows in practice: one bot had 25 thousand characters sitting right in its settings, and the review recommended moving them into the knowledge base.

Which files the knowledge base accepts

You can upload PDF files, including scans with text recognition, Word documents and Excel tables. A document is cut into fragments of about 500–800 tokens with an overlap of about 15 %, so a thought is not cut off at a boundary. Each fragment stores its source and the chapter or page, so the AI chatbot can refer to it: "according to such-and-such document, page so-and-so".

Every bot has its own base. A platform rule forbids mixing data of different owners: a bot searches only its own collection. That protects against "pulling out someone else's data" and keeps answers on topic.

Example: an AI chatbot on an expert's book

One of our AI bots runs on the material of a book: roughly 131 thousand words became 748 fragments. The bot answers in its own words in the author's manner, without word-for-word quotes, and uses up to three fragments for an answer. If the search fails for some reason, the AI chatbot answers the way it would without the base, instead of crashing.

Example: an AI chatbot on a price list

Another example is a sales AI bot with an equipment catalog. Its price list has 54,829 rows, and the master catalog has 28,687 items. The bot looks up a price first in the accounting system, then in the catalog, then in the price list. It names a price from the database, not "from its head", and the words "in stock" and "lead time" are forbidden in price list answers: availability and dates are confirmed by a manager.

Parametric selection rests on semantic search over the catalog; an example query is "a pump for 63 liters per minute".

Honest limits of an AI chatbot on your data

No AI chatbot offers a guarantee of "not a single error", because a language model sits underneath, and we do not give one. What we do: limit the bot to the context it is given, forbid prices and dates outside the database, test the bot on a staging setup, and keep a log of conversations that shows where the bot answered itself, where it passed the conversation to a person and where it was wrong. The log exists so scenarios can be improved.

A large upload of documents is a planned operation. While mass indexing runs, search is slower, so we do the first load of a big archive outside the bot's working hours.

Where the data goes

In the standard setup the text of a question and the fragments found are processed by an external language model provider. A closed circuit for an AI chatbot on your own servers is possible by separate agreement, with the terms clarified in a brief. We do not promise the country of data storage for the knowledge base in advance: we fix it in the project agreements.

What next

If you need not just a bot but a separate industry AI service or an API over your catalogs, that is the AI on your data service. Requests from the dialog go to the owner and to accounting systems; see the lead capture chatbot and the integrations pages. How to check that a bot does not give away more than it should is covered in the AI Agent Security Kit.