Your company knowledgeKnowledge base Telegram bot:
Knowledge base Telegram bot:
answers from your own documents
A knowledge base Telegram bot answers from your company materials — prices, service descriptions, policies, FAQ — instead of from the model general knowledge. In CHATBOSS.PRO those materials live as the assistant knowledge base of up to 100 000 characters: the platform splits it into fragments, computes an embedding for each and stores the vectors, so only the relevant fragment travels with the request. This is retrieval, not fine-tuning — the model is never retrained, and a correction in the handbook applies from the very next answer.
Prices, service descriptions, policies and FAQ become the source of the answers: the platform splits them into fragments, indexes them and puts the relevant piece into each request. Edit the handbook and the next answer already follows it. From $9/mo.
Author: Павел Седов, founder of CHATBOSS.PRO · Updated: September 19, 2026
Knowledge base
up to 100,000 chars
Edits apply
next answer
Model fine-tuning
not required
Plans
from $9/mo
Why an AI bot without a knowledge base invents answers
A language model knows the language, not your price list. With no materials to lean on it honestly tries to help — and produces a plausible text that has nothing to do with what you sell.
Answers «at market average»
The bot quotes somebody else terms, somebody else delivery times and services you do not offer. The customer believes it, arrives with that expectation and you get a complaint instead of a sale.
The materials are scattered
Prices in a spreadsheet, terms in a document, the real answers in the support chat history. Neither a bot nor a new hire can lean on a single source.
Updating hurts
A price changes and the correction has to reach everyone who answers customers. When knowledge is wired into dozens of scenario branches, an update becomes a project of its own.
How the bot answers from your documents
No magic in the mechanics: materials are stored separately from the instruction, split and indexed in advance, and only the relevant part is sent to the model.
Handbook separate from instruction
The assistant has two fields: a short instruction of up to 50 000 characters that says who it is and how it answers, and a knowledge base of up to 100 000 characters holding the facts — prices, services, FAQ, policies.
Fragments and embeddings
On save the base is split into fragments, an embedding is computed for each with text-embedding-3-small and the vectors are stored by the platform, so the request carries the matching fragment instead of the whole handbook.
Retrieval, not fine-tuning
The model is not retrained on your data and does not memorise it. The materials are attached to the request as context — which is why a correction in the handbook works from the next answer, with nothing to wait for.
Control over size and cost
Sending the whole handbook every time would hit the context limit and multiply the price of each answer. Fragment selection keeps both in check, and the real cost per call is in the spend log.
Answers from uploaded files
Besides the text handbook the assistant reads files from storage: PDF, DOCX, txt and md up to 20 MB. The platform extracts the text and indexes it through the same pipeline, so a price list can simply be uploaded as a file.
One source for every scenario
The same base serves pre-sales, support and internal questions from your team — a correction is made once and every part of the bot follows it.
Conversations can be encrypted
End-to-end encryption is available per bot: the key is issued to the owner once and stored only by them. The scheme is described on a separate page of this site.
Teams that need a bot over their knowledge base
The more written material a company has accumulated and the more often people ask about it, the faster a bot answering from it pays for itself.
Support and first line
The bot answers from the base of common questions and instructions, and hands disputed cases to an operator with the history attached.
Sales with a complex price list
When there are dozens of services and the terms depend on the tier, the bot pulls the exact line of the price list instead of approximating.
Internal policy assistant
Employees ask the bot about processes, access and rules rather than searching chats and shared drives for the current version.
Onboarding and training
A new hire asks the bot and gets the answer from the current version of the materials, without pulling a mentor into the same question again.
How to connect a knowledge base to a Telegram bot
No separate integration is needed: the knowledge base is a field on the assistant. Collect the materials, save the assistant, and the platform does the rest.
Step 1
Collect the materials into one text
Prices, service descriptions, terms and the questions customers actually ask, with answers. The clearer the structure, the more precisely the bot finds the right fragment.
Step 2
Fill the assistant knowledge base
Paste the text, upload files or let the AI wizard fill it from a description of your business — embeddings are computed the moment the assistant is saved.
Step 3
Set the rules of answering
The instruction tells the bot to answer strictly from the handbook and to call a human when the answer is not there — that is what keeps it from improvising.
Step 4
Test on real questions
Run the phrasings your customers use through the test chat and add whatever the base turned out to be missing.
Knowledge base bot — frequently asked questions
Is this fine-tuning a model on our data?
No. The model is not retrained and does not memorise your materials. The knowledge base is stored on the platform side and the relevant fragment is attached to the request as context — the approach usually called retrieval, or RAG.
How large can the knowledge base be?
The assistant knowledge base holds up to 100 000 characters and the main instruction up to 50 000. If you have more, split it across several assistants by topic or move the bulk into files in storage.
Can the bot answer from PDF or Word files?
Yes. Files available to the assistant are indexed alongside the text base: PDF, DOCX, txt and md up to 20 MB. The one thing that will not work is a scan without a text layer — there is nothing to extract, and the file is honestly marked as failed indexing.
How quickly does a price change reach the answers?
As soon as the assistant is saved: the platform recomputes the fragments and their embeddings, and the next answer already uses the new version.
Why not send the whole base with every request?
Every model has a context limit, and input tokens are paid for in each request. Splitting the base in advance keeps answers inside the limit and keeps the cost per conversation predictable.
What does the bot say when the answer is missing?
That is set by the instruction: usually it says it will check and moves the conversation to an operator. Letting the model fill the gap is the option that costs you a complaint later.
Can the same base serve internal questions?
Yes. Policies, instructions and internal rules go into the same handbook, and the bot answers employees from it exactly as it answers customers from the price list.