What to put in a chatbot knowledge base

A support bot is only as useful as what you give it to read. Feed it one long document and it will miss simple questions. Feed it text written the way you talk to colleagues and it will miss the way customers talk to you. This guide covers what belongs in a chatbot knowledge base, how to split it up, how to write it so real questions find it, and what to leave out.

How the bot uses what you give it

You do not need the technical detail, but one idea helps. The bot does not read everything before each reply. It looks for the few passages that seem closest to the visitor's question and answers from those. Botsmith, for example, retrieves the relevant text from your pages and your uploaded files before replying.

Two things follow. If the right passage is buried inside a huge block about everything, it is harder to find. And if the passage uses different words from the question, it may not be found at all. Almost every tip below comes back to those two points.

What goes in

Start with the facts customers ask about before they buy or book. Give each topic its own entry.

Much of this never made it onto your website. It lives in a price list, a welcome email or your head. That is fine. With Botsmith, anything that is not on the site can be uploaded as a PDF, Word or Excel file, or pasted as text.

Split it by page, not one big lump

The quickest way to set up a knowledge base is to paste everything into one box. It is also the most common mistake. We made it ourselves when building a demo bot: the whole sample business went in as a single entry, and a short question about Saturday opening found nothing. The hours were in there. They were one line among hundreds, and nothing marked them out.

The fix was to split the material the way a website is split: one entry for hours, one for prices, one for each policy. The same question then worked first time.

• One topic per entry. If you would give it its own page on a website, give it its own entry.
• A plain title. "Opening hours" and "Cancellation policy", not "Info 3".
• Short entries beat long ones. A few clear paragraphs under a clear heading are enough.
• Repeat the subject. Write "Our cancellation policy is…" and not "It is…". Each passage may be read on its own, without the heading above it.

Write it in your customers' words

This is the part most owners skip, and it causes the most missed answers. You know your business by its official terms. Your customers do not.

Two examples we have seen go wrong:

• A tutoring centre described its lessons as "one-to-one". A parent asked about "one-on-one tutoring" and the bot found nothing. Same service, different phrase.
• A business had a properly written cancellation policy in formal language. A customer typed "what if I cancel on the day" and got no answer. The policy covered it, but no sentence in it sounded like the question.

The cure is simple. Under each fact, add the ways people really say it.

1. List the other names. "One-to-one lessons, also called one-on-one tutoring, private tuition or individual lessons."
2. Answer the spoken question directly. "If you cancel on the day of your appointment, the full fee is charged." Put that sentence next to the formal policy.
3. Cover the short questions. "Are you open on Saturday? Yes, from 9am to 1pm." A question-and-answer pair is easy for the bot to match.
4. Borrow from real messages. Look through old emails, texts and reviews. Copy the phrases customers used.

Once the bot is live, it will hand you more phrases. In Botsmith, questions it could not answer are collected for you automatically. Each one is a wording you had not thought of. Our guide on how to tell if your support bot is working shows how to go through them each week.

What to leave out

• Anything out of date. Old price lists, last year's timetable, a closed branch. The bot cannot tell old from new. If two entries disagree, it may pick either.
• Internal notes. Supplier costs, margins, staff rosters, remarks about customers. If the bot can read it, assume a visitor can be told it.
• Personal information. Customer records, staff home numbers, anything you would not print on a poster.
• Behaviour rules. "Never give medical advice" is an instruction, not a fact. It belongs in the bot's instructions. See how to write instructions for your support bot.
• Marketing copy with no facts in it. "We go the extra mile" answers nobody's question.
• Promises you cannot keep. If a human would add "it depends", write the "it depends" down.

Check it before customers do

Write down ten questions the way a customer would type them: short, casual, with a typing slip or two. Ask the bot each one. For every miss, find out which of the two causes it was. Either the fact is not in the knowledge base, or it is there under different words. Fix the entry and ask again.

Then ask a couple of questions you know are not covered. A bot that answers only from your content should say it is not sure. If it produces a confident answer anyway, something in your knowledge base is being stretched too far, or the tool is guessing.

The demo bots on our tutoring centre and physiotherapy clinic pages are built this way, with one entry per topic. Try a few casual questions on them to see how it behaves. The list of topics in what a website support bot should answer also makes a handy checklist for your first entries.