What to Put in Your Chatbot's Knowledge Base (and What to Leave Out)

A practical guide · Torensa

The knowledge base is the single biggest factor in how good a grounded chatbot turns out. The model is fixed; the content is yours to shape. Get the knowledge base right and even a simple bot feels sharp and reliable. Get it wrong — too much, too vague, or contradictory — and no amount of tuning will save it. Here's how to decide what goes in.

Include the content that answers real questions

The best knowledge base is built from the questions people actually ask. Good candidates almost always include:

• Frequently asked questions, written as clear question-and-answer pairs.
• Policies with exact details: pricing, refunds, shipping, cancellation, and timelines.
• Step-by-step instructions for common tasks ("how do I reset my password").
• Definitions of your product's terms, plans, and limits.
• The small operational facts people ask about — support hours, contact methods, turnaround times.

If you're not sure whether something belongs, ask: has a real person ever asked this? If yes, it earns a place.

Leave out the content that adds noise

More text is not better. Anything that doesn't help answer a question makes it harder for the bot to find the part that does. Leave out:

• Marketing copy and taglines — they read well but answer nothing.
• Outdated or deprecated information that contradicts current policy.
• Long legal documents in full when a plain-language summary is what users need.
• Internal notes, changelogs, or anything you wouldn't want quoted to a customer.
• Duplicate passages that say the same thing two different ways.

Structure it so answers are easy to find

You don't need a special format, but a little structure goes a long way. Short, self-contained chunks work better than sprawling paragraphs. Lead with the question or the topic, then give the answer directly. Spell out names and terms instead of assuming context, because each answer should make sense on its own. When two pieces of information are related — a policy and its exception, say — keep them close together so the bot presents them as one coherent answer.

Resolve contradictions before they reach the bot

Conflicting text is the most common cause of weak answers. If one section says refunds take 14 days and another says 30, the bot has no way to know which is right, and its answer will hedge or pick wrong. Read through your knowledge base specifically looking for two statements that can't both be true, and fix them. This single pass often improves answer quality more than anything else.

Plan for the questions you can't cover

No knowledge base is complete, so write the fallback into it. Include the way you want users routed when the bot doesn't know — an email, a contact form, a support link — so the assistant can offer a real next step instead of a dead end. A bot that says "I don't have that, but you can reach the team at…" is far more useful than one that guesses.

Keep it current

A knowledge base is a living document. The moment a price, policy, or process changes, update the text — every future answer depends on it. A useful habit is to review the questions the bot couldn't answer on a regular schedule; each one is either a gap to fill or a sign that a topic needs clearer wording. Maintained this way, the bot gets steadily better with very little effort.