Chatbot

Chatbot Knowledge Base: How Much Content Can It Hold

Your chatbot is only as good as the documents behind it. This guide explains what a token is, how much content fits into a 1 million token knowledge base, and which files belong in it. It also covers what to leave out and why a small clean knowledge base answers better than a large messy one.

A chatbot knowledge base is the set of documents your bot answers from. In the NERO panel it holds up to 1 million tokens — roughly several hundred pages of text. For most businesses that is far more room than they will ever fill, so the useful question is not how much content fits, but which content deserves to be in there.

This guide explains what a token is in plain language, how many real pages fit into one million tokens, what to upload, what to leave out, and why a small clean knowledge base usually answers better than a large messy one. Everything below applies to the chatbot platform at app.aimedia.az, where the knowledge base, the visual flow builder and the unified inbox all sit in the same panel.

What is a token, in plain language

A token is a small piece of text — usually a short word, or part of a longer one. The word "delivery" is about one token. A long or unusual word may be split into two or three pieces. Punctuation and spaces are counted too.

A working rule of thumb: one token is roughly three quarters of an English word, and one ordinary page of text lands somewhere between 400 and 700 tokens depending on the language and the formatting. Languages written in non-Latin alphabets normally use more tokens to say the same thing.

You never count tokens by hand. The panel measures the uploaded content for you. The number only matters when you want to be sure that your full catalogue plus your internal manual will fit — and in practice they always do.

How much is one million tokens in real documents

Here is a realistic picture of what a small or medium business actually loads into an AI chatbot knowledge base, with rough token sizes.

What you upload	Rough size in tokens	Comment
Frequently asked questions (60–80 pairs)	15,000 – 40,000	The single most valuable file
Product catalogue (500 items, short descriptions)	100,000 – 200,000	Depends on description length
Pricing and discount rules	3,000 – 15,000	Rules work better than fixed lists
Service descriptions	5,000 – 20,000	One short section per service
Delivery, returns and warranty policy	3,000 – 10,000	Write it the way you explain it by phone
Internal staff manual and objection handling	10,000 – 50,000	Turns the bot into a trained employee
Branch addresses, working hours, contacts	1,000 – 5,000	Keep one source and update it in one place
Typical total	Far below 1 million	Plenty of room left for growth

Most companies finish their first knowledge base at well under a tenth of the limit. The businesses that get close are usually large retailers with tens of thousands of catalogue lines, and for them the better answer is a live connection to the database rather than a static document — that kind of work sits under custom development on our services page.

What you should upload to the knowledge base

• Frequently asked questions with real answers. Open the last two or three hundred messages in Instagram DM and WhatsApp and write down the questions that keep repeating. This is the highest-value content you will ever add.
• The catalogue. Product or service names, short descriptions, variants, availability rules. Consistent naming matters far more than beautiful copy.
• Pricing rules, not only numbers. How a quote is calculated, what changes it, what is included, and when a human must confirm it. Rules stay correct much longer than fixed lists.
• Policies. Delivery zones and timings, returns, warranty, cancellation, prepayment conditions.
• Internal guidelines. How your best salesperson answers "is it worth it", "why should I choose you", "can you do it faster". Upload that reasoning and the bot starts using it.
• Operational facts. Addresses, working hours, holidays, parking, which branch handles what.

Content can be typed straight into the panel, pasted in, or added as files. Because the bot detects the language a customer writes in and replies in that language, you rarely need a separate translated copy of every document — one clean source is usually enough for a multilingual chatbot.

What you should not upload

• Outdated price lists and finished campaign pages. If two documents disagree, the bot has no way of knowing which one you meant.
• Several versions of the same file. "final", "final_v2" and "final_new" together are the fastest route to inconsistent answers.
• Scans and photographed documents with no text layer. If you cannot select the text with your mouse, the bot cannot read it either.
• Long marketing brochures. Adjective-heavy copy gives the bot very few checkable facts and a lot of noise.
• Internal data customers must never see — supplier terms, margins, staff personal data, partner agreements.
• Legal boilerplate nobody ever asks about. If no customer has raised it in a year, it does not belong in version one.
• Anything you are not willing to keep updated. An unmaintained document turns into a wrong answer within a few months.

Quality beats volume

A knowledge base is not a bookshelf where more books mean more knowledge. When a customer asks something, the system searches your content, picks the passages that look most relevant, and builds the reply from them. Two things break that process.

Contradiction

If delivery time appears as two days in one file and four days in another, both passages look equally relevant. The answer becomes unstable — sometimes right, sometimes wrong, and always hard to debug. One fact should live in exactly one place.

Dilution

If forty pages of company storytelling surround a single line about return conditions, that line is much harder to retrieve. Deleting the storytelling makes the chatbot more accurate, not less capable.

A tight knowledge base of thirty well-written pages will beat three hundred pages of mixed material almost every time. That is why content preparation, not technical setup, is the part of a chatbot project that takes real effort. Once the content is ready, a standard bot is built and handed over in about an hour.

How to write content the bot can actually use

1. One topic per section. A short heading, then the answer. Never bury three subjects in one paragraph.
2. Write the way customers ask. "How long does delivery to the regions take?" works better than "Logistics service level framework".
3. Use full sentences, not fragments. A table cell that says only "2–4" means nothing on its own.
4. Spell out the conditions. "Free above a certain order size" is useless to a bot; state the exact rule.
5. State the exceptions. Holidays, out-of-stock cases, regions you do not serve, items you refuse to handle.
6. Say when a human takes over. Put the handover rule into the content, so the bot passes the chat to a live agent instead of guessing.

Common mistakes, and when you do not need a big knowledge base

The most common mistake is treating upload volume as a milestone. The second is uploading the whole website. A website is written for browsing, not for answering, and it usually repeats the same claim in six different wordings.

The third is loading a catalogue that changes daily. If stock and prices move every morning, a static document goes stale immediately; a database connection is the correct answer there.

You may not need a large knowledge base at all if:

• Your business has one service and one clear pricing rule — a short scenario in the visual flow builder covers it.
• Almost every enquiry ends in a booking — you need appointment logic more than documents.
• Your real problem is order, stock or production tracking rather than answering questions. In that case a management system fits better than a bot; see our products.

Saying this out loud early saves everyone time and money.

Keeping the knowledge base correct over time

A knowledge base is not a one-off upload. A simple monthly routine keeps it healthy:

• Review the conversations the bot could not answer — they are your next content items.
• Update the single source document when a rule changes, never a copy of it.
• Delete retired information instead of adding a note next to it.
• Re-read the pricing rules section after every campaign ends.
• Ask one operator to test five real questions after each update.

Teams that do this reach a point where the bot handles the large majority of first messages without a human. Teams that upload once and forget usually switch the bot off and blame the technology. The difference is maintenance, not model quality.

Next step

If you already have an FAQ list, a catalogue and your pricing rules written down, your knowledge base is mostly done. Send us what you have and we will tell you what is missing, what to cut, and whether a chatbot is even the right tool for your case. The first conversation is free and carries no obligation.

Contact page · Message us on WhatsApp · Platform: app.aimedia.az · Pricing page · AI Media company profile in English