How to Write Documents Your AI Chatbot Can Actually Use

A practical guide to writing FAQs and knowledge base docs an AI chatbot can retrieve accurately — chunk-sized sections, customer phrasing, and specific numbers over vague claims.

Every business already has a knowledge base. It is scattered across a pricing page, an old FAQ doc, a WhatsApp thread where you answered the same question for the fifth time, and whatever is in your head. The problem is never that the information does not exist. The problem is that it is not written in a shape a chatbot can retrieve.

Most wrong chatbot answers trace back to the documents, not the model. So before you upload anything, it helps to know what “chatbot-ready” actually looks like. Here is how to write it.

Write in the language of the question, not the answer

Your policy document says “subscription termination and reimbursement eligibility.” Your customer types “can I get a refund?” Those mean the same thing to a person, but retrieval works on similarity between the query and the text — and formal, legal phrasing sits far from plain, everyday phrasing in that similarity space.

Before you write a section, write down the actual question a customer would type. Then answer that exact sentence, using words a customer would use. If people ask “do you deliver on Sundays,” your hours document should contain the words “deliver” and “Sunday” — not just a table labeled “Operating Schedule.”

One heading, one question, one complete answer

Minyut splits documents into 512-token chunks with a 50-token overlap, cutting at paragraph and sentence boundaries so meaning stays intact. That means each chunk your chatbot retrieves is roughly the size of a short paragraph. If the answer to a question is spread across three paragraphs that all depend on each other, the chunk that gets retrieved might be missing the part that actually matters.

Structure documents so each section can stand alone. Give every section its own heading, phrased as the question it answers, and make sure the paragraph beneath it is a complete answer by itself — not a continuation of the point above it. Never write “as mentioned above” or “see the previous section.” Your chatbot will never see “above.” It only sees the chunk that got retrieved.

A useful rule of thumb: if you covered the heading and handed a stranger only the paragraph underneath it, could they still answer the question correctly? If not, rewrite it so they could. That stranger is exactly what your chatbot sees at query time.

Numbers and specifics beat adjectives

“Fast shipping” is not an answer. “Orders ship within 2 business days, and delivery takes 3–5 days within India” is. A chatbot that only has adjectives to work with will either hedge (“shipping is generally fast”) or invent a number to sound helpful — and an invented number is worse than no answer.

Go through your existing docs and replace vague language with the actual figures: prices, dates, durations, quantities, thresholds. If your return window is 7 days, write “7 days,” not “a short window.” Specific documents produce specific, checkable answers — which is exactly what makes a chatbot trustworthy instead of merely confident.

Keep documents focused, not comprehensive

A tight 5–10 page FAQ that covers your actual top questions will outperform an 80-page manual every time. Every chunk in your knowledge base competes for the same handful of retrieval slots at query time — dense boilerplate and filler sections crowd out the passages that actually matter, even when the useful information is technically in there somewhere.

Start with the questions you already answer most: pricing, hours, policies, how something works. Write those first, focused and complete. Add more documents later, once you have confirmed the core ones are answering correctly. This is also why text-based PDFs and plain Markdown files tend to work better than exported slide decks or scanned manuals — the useful text is not buried behind formatting or images.

Treat your documents like a changelog, not a one-time upload

A knowledge base that was accurate on day one degrades quietly. Prices change, policies update, and a document nobody touches for six months starts giving customers information that used to be true. Whenever you would update your website copy or your pricing page, re-upload the matching document to your chatbot in the same sitting. It takes a few minutes and it is the difference between a chatbot customers trust and one that occasionally embarrasses you.

A quick checklist before you upload

Read each document once, pretending you are a customer who knows nothing about your business. Does every section answer a real question, in the words a customer would use, with specific numbers instead of vague claims, without depending on anything written above or below it? If yes, it is ready. If you are correcting yourself as you read, fix it first — the chatbot will not correct it for you at query time.

Good documents are the entire product here. The retrieval, the chunking, and the model are already built — what determines whether your chatbot is genuinely useful is whether the thing you handed it was written to be found. Start a free Minyut account, upload one focused document written this way, and test it against the exact questions your customers ask. The free plan covers 100 queries a month, which is enough to see the difference immediately.