Updated: 2026-07-24
An agent knowledge base is its own set of documents (a price list, FAQ, policy) that it answers from instead of making things up. You upload data three ways — pasted text, a file or a URL; the service splits the text into fragments and indexes them. For every question the agent gets only the pieces relevant to it and answers from them. The base turns on by itself from the first document and works in chat and in Telegram.
Open the "Knowledge" tab. In the agent settings go to the "Knowledge" section — this is where all its documents are kept.
Add a document. Pick a way: paste text, upload a file (txt, md, pdf, docx, html, csv) or provide a page URL.
Wait for indexing. Indexing runs right away; the document row shows a "Ready" status, the fragment count and the character count.
Check answers in the preview. Ask the agent a question about the document. Also ask something the documents do not cover — make sure it honestly says the answer is missing instead of inventing one.
Connect Telegram (optional). On the "Telegram" tab paste a bot token from @BotFather — the same agent will answer from the same documents in DMs and groups.
The knowledge base lives on the "Knowledge" tab in the agent settings. You can add data in any of three ways, and they merge into one base:
The agent form has a separate "Knowledge" field for short rules ("we are open 9 to 6", "be formal"). These reach every reply in full, because they are short and always relevant.
The knowledge base is different: a large document is indexed once, and only the fragments relevant to the current question are mixed into the reply. A 40-item price list or a 20-page policy will not fit into every reply — so large data goes into the base, not the "Knowledge" field.
Finding the right fragment relies on search by meaning, not exact word match: the question "how much is delivery" finds the "shipping rates" section. Exact terms, codes and SKUs are still taken into account, so questions can be asked in plain words.
Along with the found fragments the agent receives an instruction: rely on them, and if the context has no answer — say so honestly instead of inventing one. For a support bot this is the key property: "let me check with a manager" beats a made-up price.
The base turns on automatically as soon as the first document is uploaded and indexed — no separate activation needed. From there it works in both channels: the on-platform chat and the connected Telegram bot. The base belongs to the agent, not to a channel, so there is no per-channel setup.
Pasted text, files in txt, md, pdf, docx, html, csv, and fetching by URL. Only text is indexed: images, layout and scanned PDFs without a text layer are ignored.
The "Knowledge" field holds short rules that go into every reply in full. The knowledge base holds large documents, and only the fragment relevant to the question is mixed into the reply. Keep rules in the field, large data in the base.
No. The base turns on automatically as soon as the first document is uploaded and indexed. Until there are documents, the agent replies as usual, without consulting a base.
Yes. The base belongs to the agent, not to a channel: the connected Telegram bot answers from the same documents as the on-platform chat. No separate Telegram setup is needed.