The knowledge base is your documents, from which the bot answers client questions. Instead of inventing answers, AI looks for a matching piece in your texts and answers from it, attaching the source. Knowledge bases belong to a bot: one bot can have several bases — the price list separately, the delivery rules separately, for example.
Documents
You can add plain text, marked-up text, a table and a website page by link. Each document in the list shows its processing state: queued, processing or ready. An unneeded document is deleted — its traces in search go with it, so outdated facts never surface in answers.
Preparing documents
To make documents searchable, the application splits them into small fragments and turns each into a set of numbers — a so-called vector — using the embedding model you choose. That model is fixed to the base when it is created. If you later change the model, the base is marked “needs re-indexing”: until that is done, search quality will be lower. The application warns about it.
Search two ways at once
Search runs simultaneously by meaning and by words, and the results are merged and ranked. That is how both the question “how much is a haircut” and the exact wording from the price list are found. The best fragments found are shown with a score and a link to the document they came from — you can see what the answer is based on.
Answers from documents
The “AI chat” node can answer strictly from the base and attach sources. A mandatory “no answer” rule applies here: if nothing suitable was found in the base, the bot does not invent — it says so honestly and can, for example, call an operator. That is what protects you from wrong promises to a client.
The gaps report
The application collects the questions that found no answer in the base and shows them with a count over the last thirty days. It is a ready list of what is worth adding to the documents. From any question you can create a draft answer in one click and index it right away.
Testing the base
The “Knowledge” section has a testing window: ask any question and see which document fragments were found, with what score, and what the model would have answered. So answer quality is visible before clients ever see it — you can calmly extend the documents and check again.
Refining with scripts
If an export arrives with extra markup or junk, the document’s text can be run through your own Python script before it is split into fragments — it cleans or rewrites the document. The second place for a script is the order of what was found: after the search, a script can refine which fragments matter more and drop the extra ones. Both are optional and configured separately; details in the Scripts section.
Next: how to configure models — in the Models & Prompts section, how to use search in the scenario — in Flow Nodes, and how to put it all into a bot — in Bots.