AI knowledge base for a Telegram bot: feed it right (2026)

TL;DR. A knowledge base answers exactly as far as it has an answer, and not one word further — but it is only half of what the assistant draws on, the other half being your live shop data. AdminHub takes eight kinds of source, from a website crawl to a voice memo, and charges differently for each because the work differs: 5 credits for pasted text, 30 for a PDF, 5 per started minute for audio, nothing for a CSV you already have. Everything an AI pulled out waits in a review queue before going live. And what decides whether the base is any good six months from now is not the upload — it is the gaps list, where every handover is logged automatically.

Almost every guide about training a bot on your own data ends in the same place: upload a file, watch a progress bar, the bot knows things. That is where the easy half ends.

A base has one honest property no upload screen shows you: the questions it cannot cover are invisible. You find them when a customer asks and nothing useful comes back. So this is about the base itself — what you can feed it, what happens to that material inside, and why unanswered questions are worth more than answered ones. Worth knowing before you start: the base is not the assistant’s only source. Order status, tracking, the customer’s own bookings, your catalogue and your opening hours are read live from your shop, so none of those need an entry. For how the assistant behaves in a live conversation, see AI customer support on Telegram.

Eight things you can feed it

Source	What it is for	Cost to ingest
Website	Crawls up to 10 pages of one domain	15 credits
Pasted text	A policy you can copy in one block	5 credits
CSV	Pairs you have already written down	free
PDF	Manuals, terms, spec sheets	30 credits
DOCX	The same, in Word form	30 credits
YouTube video	A walkthrough you recorded once	10 credits
Voice file	Answers that live only in your head	5 credits per started minute
Manual entry	The one answer you keep retyping	free

CSV and manual entries are free because nothing is generated — the pairs exist already. Everything else runs the material through a model that pulls out question-answer pairs, and that pass is what you pay for.

A Notion importer is written but not switched on: it needs a separate integration that is not connected yet, and the tile only appears in the app once it is. Until then, export the page and paste it in as text.

Ceilings apply whatever the source: a crawl follows at most 10 pages, any single ingest reads about 30,000 tokens and stops, PDF and DOCX cap at 10 MB, voice at 25 MB and 30 minutes. A 400-page manual will not go in whole — and should not, since the part a customer needs is maybe six pages of it.

The prices differ because the work differs

Not a price list with tiers — the shape of the underlying job showing through.

Pasted text is 5 credits and is typically a few hundred words. A PDF is 30 — not because documents are a premium feature, but because one can carry the full input ceiling. Same extraction pass, up to sixty times more text going into it.

Voice is the only source priced by duration, at 5 credits per started minute, because it is the only one needing an entire extra step before reading can begin: the recording has to be transcribed, and that cost scales with length and nothing else. Ten minutes of voice is 50 credits; the same content pasted as text is 5. Not a penalty for talking — the cost of the extra step.

YouTube sits at 10, in between: no audio is processed, since the published transcript is fetched as text, but transcripts run long. Website is 15 because it adds an external crawl before anything can be read.

On the Free plan — one source, 30 credits a month — a single PDF is the whole month. If your answers are written down somewhere already, a CSV costs zero. Uploading a 100-page catalogue because it is the file on your desktop is the most expensive route to the same six answers.

Between the upload and the first answer

Extraction. The material is read and turned into question-answer pairs, each tagged as a general question, a product description, a policy, or other. A translation, not a copy — a 40-page document might legitimately yield eleven pairs.

Review. Everything an AI extracted arrives inactive, in a review queue, and answers nobody until you approve it. CSV is the single exception, on the grounds that you wrote those pairs yourself.

Deduplication. Pairs whose question already exists are dropped quietly on the way in. Separately, a duplicate finder surfaces live entries 0.85 or more similar by meaning — the near-twins exact matching misses, like one shipping rule that arrived twice in different words.

Indexing. Approved entries queue for embedding, and a worker drains that queue every 15 seconds in batches. This is what makes search work by meaning rather than keywords, in any language: “when will it get here” finds a shipping entry that never uses that phrase, and a question asked in one language finds an entry written in another.

Retrieval. When a customer writes, the 5 nearest entries go in front of the model — not the whole base. If the person changes topic mid-chat, the assistant can search once more for up to 5 more.

Two things follow. Write entries so each stands alone — an answer that only makes sense beside its neighbour will often arrive without it. And a bigger base is not automatically better: 300 vague entries make it harder for the right 5 to surface than 60 sharp ones do.

What does not belong in the base at all

The most expensive mistake here is not a bad upload — it is writing down what the assistant can already look up.

Order status, payment state, delivery method, the tracking number, a customer’s order history, their upcoming bookings, the free slots for a service, what is currently in your catalogue and what your working hours are: all of that is read live from your shop when the question is asked. An entry saying “order 312 ships on Friday” is stale the moment Friday passes. The assistant can also cancel a booking or take a callback request, both only after the customer taps a confirm button — again, nothing you write into the base.

So the base earns its keep on everything a record cannot hold: the refund window and its exceptions, how shipping actually works, what happens if a parcel is refused, the difference between two similar products, whether the warranty survives a resale.

The gaps list is the part that compounds

Here is the mechanism the upload-and-done guides never reach. When the assistant cannot answer confidently from the base, it does not improvise. It hands the conversation to a human — and writes a line into the gaps list. Every handover, logged automatically, in one place.

What that line contains depends on the surface: from a bot chat it is the assistant’s own short note on why it gave up, capped at 200 characters; from a Telegram Business chat it is the customer’s message itself. Either way you are reading the list for repeats, not for individual wording.

That list is the only honest map of what your base is missing, because real customer questions write it rather than your guesses about them. Nobody predicts that people will ask whether the warranty survives a resale — but the third time it shows up in gaps, you know where it belongs.

You can also ask AI to draft the missing answer, for 2 credits. It reads what the base already contains and proposes a pair grounded in that material, which you approve or edit before it goes live. When the material does not support an answer, it returns an explicit “not enough here” with a reason rather than inventing something.

So the real loop is: ship a partial base, let customers hit its edges, fill the gaps that repeat. A month of that produces a base shaped by your actual customers rather than your idea of them.

What the extractor refuses to write down

Non-answers are rejected outright: “depends on your needs”, “contact us for details”, “negotiated individually”. That is text, not information, and a base full of it produces an assistant that sounds helpful and resolves nothing.

Facts with a short shelf life are skipped on purpose — specific prices, dates, whether something is in stock today. Those drift without the base drifting with them, and a confidently wrong price does more damage than a missing answer.

And nothing that is not literally in the material becomes an entry. Feed it a marketing page with no facts and you get very few pairs, or none — not a malfunction, just a page with nothing a customer could be told.

Where this stops being the right tool

• Lookup, not judgement. The assistant reads your documents and your shop records. Ask something needing a judgement call rather than a lookup and the right outcome is a handover, not an answer.
• No material, no answer. If a fact was never written down and no record holds it, no source type conjures it. Admitting the gap is the honest move, and that is what the gaps list is for.
• Live negotiation and contested cases need a person. Discounts, complaints, anything where the reply depends on who is asking.
• Regulated specifics — legal, medical, financial. Keep those manual by design. An approximate answer there is a liability, not a convenience.
• An unmaintained base decays. Sources do not re-read themselves; a re-sync is a button you press, and it costs the same credits again.

What to do now

• Start with what is already written. A CSV is free and goes live immediately, with no review queue.
• Do not write down what the shop already knows. Order status, tracking, bookings, slots, hours and the catalogue are read live.
• Feed one source, not eight. Extraction quality is easier to judge on a small batch.
• Read the review queue. Rejecting half of it is a normal, healthy result.
• Check gaps weekly. That list is your content plan, written by the people you sell to.

---

How the assistant uses that base in a conversation is the other half, in AI customer support on Telegram. For the product view see AI Support for Telegram; if you are not fielding customer questions yet, start at how to sell on Telegram.