Custom AI Chatbot on Your Own Business Knowledge: How It Finds the Right Paragraph, and How to Write Knowledge It Can Find (Free Chunk Finder)

Custom AI chatbot usually means one thing in practice: a general AI model that answers from your business knowledge instead of the whole internet. Nobody trains a new brain for your shop. What makes the bot “yours” is the text you give it and how that text is cut up, searched and handed to the model before every reply. When answers go wrong, the cause is almost always in that step, not in the AI. This guide explains the step in plain words and gives you a free tool that shows which paragraph of your knowledge a question would land on.

3steps behind every answer
1topic per chunk
50–80words per chunk works well
0new models trained

⚡ In short

  • A custom bot is a general AI model plus your knowledge. The knowledge is what you customise.
  • Before each reply, the system picks a few chunks of your text and gives only those to the model.
  • If the right chunk isn’t picked, even a brilliant model gives a poor answer.
  • Write one topic per chunk, name the topic in the first words, and use your customers’ words.
Custom ai chatbot in LiveChatBot: AI Agent knowledge sources and a sample chat answered from them
How it looks in LiveChatBot: AI Agent knowledge sources and a sample chat answered from them.

What custom means

You’re not building an AI, you’re feeding one

When a business owner hears “custom AI”, they often picture something expensive: engineers training a model on company data for weeks. That does exist, at large companies with large budgets. It’s almost never what a small business needs, and it’s not what most tools sell.

What most products mean by a custom bot is simpler. You take a capable general model, the kind that can already read and write well, and you give it your own material: your FAQ, delivery rules, product notes, opening hours, return policy. When a customer asks something, the system finds the parts of your material that look relevant and asks the model to answer using only those parts.

That approach has a name, retrieval-augmented generation, but you don’t need the jargon. Think of it as an open-book exam. The model is a smart student. Your knowledge is the book. Before each question, a librarian runs to the shelf, pulls out two or three pages and puts them on the desk. The student answers from those pages.

The student is rarely the problem. The usual problems are that the right page isn’t in the book, the librarian brings the wrong page, or the page makes no sense without the one before it. All three are about how your knowledge is written and divided, and all three are things you can fix yourself without any code.

If you’re still deciding which pages and documents to feed your bot, start with our knowledge audit. This guide picks up after that: how the chosen text gets found.

Three steps

What happens inside a custom AI chatbot before it replies

Every reply from a knowledge-based bot goes through these steps, usually in under two seconds.

1

Chunk

Your knowledge is cut into small pieces, usually by paragraph or heading. Each piece is stored separately.

2

Find

The customer’s question is compared with every chunk. The few that look closest are picked.

3

Answer

The model gets the question plus the picked chunks and writes a reply from them. If they don’t contain the answer, a well-set-up bot says so.

Where it goes wrong

Four ways a good model gives a bad answer

Once you see the three steps, most strange answers make sense. Here are the four failures you’ll meet most often.

The fact isn’t there

A customer asks if you sell fertiliser. Your knowledge never mentions it. The finding step still picks the closest chunks, perhaps your plant care notes, and the model either says it doesn’t know or, worse, guesses. The fix is to add the fact, or to accept a polite “I’m not sure, let me get someone” for rare questions.

The wrong chunk wins

A customer asks “how long will it take?” meaning delivery, but your workshop paragraph says “about two hours” and wins the match. The model answers with the workshop length. The fix is clearer chunks: the delivery chunk should contain the words people use about delivery time (“how long”, “days”, “when will it arrive”).

The chunk needs its neighbour

Your policy page says “Delivery is free above ₹999.” Next paragraph: “It comes with a 7-day guarantee.” Split into chunks, the second one has lost its subject. What is “it”? The plant? The delivery? The model can’t tell. The fix is to make every chunk stand on its own: “Every plant comes with a 7-day guarantee.”

One chunk holds five topics

A long “About us” paragraph mentions your nursery hours, workshops, monsoon offers and parking. It matches lots of questions weakly and none of them well, and it pushes better chunks out. The fix is to split it: one topic per paragraph.

None of these needs a better model. They need better-written knowledge. That’s good news, because writing is something you already know how to do.

Free tool

See which chunk each question lands on

Paste your knowledge with blank lines between paragraphs, add real customer questions, and see which paragraph would be found for each one.

Paste your knowledge (FAQ, policy page, product notes). Leave a blank line between paragraphs; each paragraph is treated as one chunk.

Customer questions, one per line

Runs only in your browser; nothing is sent or saved. It matches on shared words, which is stricter than real AI search, so treat a weak match as a warning, not a verdict.

Reading the results

What the example shows

The example uses a made-up plant shop’s knowledge with seven common questions. It shows each failure from the previous section.

“Do you deliver to Mumbai” lands on the delivery chunk with a strong match. Good. The chunk says orders outside Pune go by courier in 4 to 6 days, so the model can answer correctly even though Mumbai isn’t named.

“What if the plant is dead when it arrives” should land on the guarantee chunk, but it lands on the delivery chunk instead. Look at the guarantee chunk: “It comes with a 7-day guarantee.” It never says “plant”, so the delivery chunk, which does, wins the tie. The tool also flags it for starting with “it”. Rewrite it as “Guarantee: every plant comes with a 7-day guarantee. If a plant arrives damaged or dead…” and run the tool again: now the guarantee chunk wins.

“Do you sell fertiliser” matches nothing. That’s the honest outcome: the knowledge doesn’t say. Your bot should hand over or collect a contact, and you should add a line if people ask.

“Are you open on Monday” finds the nursery paragraph, but that paragraph is also about workshops and the monsoon, and it’s long. The tool flags it. Split it into an opening hours chunk and a workshops chunk.

A word of honesty about the tool: it compares shared words, with a small list of synonyms. Real AI search compares meaning, so it would connect “office gifts” with “corporate gifting” even without the synonym. That makes this tool stricter than the real thing. If a question gets a strong match here, it’ll almost certainly be found by a real system. If it gets a weak match, it’s worth a second look. Use the tool to find problems in your writing, not to predict exact behaviour.

To check what the model then says with the chunk it found, use our grounding checker, which compares an answer against your source text.

Writing findable knowledge

Six rules for knowledge a custom AI chatbot can find

These rules work for any tool that answers from your knowledge, whether it’s built into a chat product or something a developer set up.

  1. One topic per paragraph. Delivery in one, returns in another, payment in a third. If a paragraph needs a heading with “and” in it, split it.
  2. Name the topic in the first few words. “Delivery: we deliver…” or “Returns: you can return…”. It helps both word matching and meaning matching, and it helps a person skimming too.
  3. Never start with it, this or they. Every chunk may be read alone. Repeat the noun.
  4. Use the customers’ words, not only yours. If you call it “corporate gifting” and customers say “office gifts”, put both in. If customers write “COD”, write “cash on delivery (COD)”.
  5. Keep chunks to 50–80 words. Long enough to hold the rule and its exception; short enough to stay on one topic.
  6. Put the number in the same chunk as the rule. “Free delivery above ₹999” must not be split from “otherwise ₹79”.

There’s a seventh habit that matters more than all six: write down the things only you know. Most bots fail on questions the website never answered. Our guide to the 11 facts a site must state is a good list to check against.

When you’ve rewritten your knowledge, run the tool again with the same questions. You should see more strong matches and fewer chunk notes. Keep the question list. Every time you change your knowledge, re-run it, the same way you’d re-check a form after editing a website.

Before and after

The same knowledge, rewritten for finding

Lost subject
🚩 Before“It comes with a 7-day guarantee.”
✅ After“Guarantee: every plant comes with a 7-day guarantee. If it arrives damaged, send a photo within 48 hours and we’ll replace it.” (Illustrative.)
Mixed topics
🚩 BeforeOne 90-word paragraph about hours, workshops and monsoon offers.
✅ AfterThree paragraphs: “Opening hours: …”, “Workshops: …”, “Monsoon offers: …”.
Owner’s words only
🚩 Before“Corporate gifting orders of 20+ get 10 percent off.”
✅ After“Corporate and office gifts: orders of 20 plants or more get 10 percent off. Write to us for a quote.”
Rule without its number
🚩 Before“Delivery is free on bigger orders.”
✅ After“Delivery is free above ₹999; below that it costs ₹79.”

Beyond text

PDFs, tables and price lists

Plain paragraphs are the easiest knowledge to find. Other formats need a little care.

PDFs. Brochures and menus often turn into jumbled text when a tool reads them: columns run together, headings vanish, prices drift away from the product they belong to. Before you upload a PDF, copy its text into a document and check it reads in order. Often it’s faster to rewrite the key facts as short paragraphs.

Tables. A price table with twenty rows becomes twenty disconnected fragments once chunked, and the column headings may end up in a different chunk from the numbers. Rewrite important rows as sentences: “The 1 kg chocolate cake costs ₹850.” Keep a full price list on your website for people, and give the bot the sentences.

Prices that change often. If prices change weekly, don’t put them in the bot’s knowledge at all. Have the bot point to the price page or hand over. A wrong price quoted confidently is the most damaging mistake a bot can make. Our number check guide shows how to test every price, date and limit your bot quotes.

Website pages. If your tool reads your website directly, the same rules apply to your pages. Clear headings, one topic per section and no “click here for details” in place of the actual detail. Writing for the bot tends to improve the page for people as well.

Instructions vs knowledge

Which customisation goes where

You want the bot to…Put it inExample
State a factKnowledgeDelivery areas, hours, guarantee rules
Sound a certain wayInstructionsFriendly, short replies, no emojis
Never discuss somethingInstructionsDon’t give medical advice about plant toxicity to pets; suggest a vet
Hand over in certain casesInstructions and handover settingsOrder problems, complaints, bulk quotes
Answer in the visitor’s languageInstructions (and widget language settings)Reply in Hindi if the visitor writes in Hindi

Setting it up

A custom AI chatbot on LiveChatBot

The instructions side, meaning tone, limits and what the bot must never say, is covered in our guide to writing AI chatbot instructions. The knowledge side is what this guide is about, and in LiveChatBot it works like this.

The AI Agent answers from the knowledge you give it. Paste your rewritten paragraphs, one topic each. Then test with your question list from the tool: ask each one in the chat widget on your own site and check the answer. Our step-by-step guide to training your AI chatbot covers where each setting lives.

Two safety nets matter for a small business. First, when the AI can’t find an answer, it should hand over rather than guess; visitors can leave an email or phone number through the away form when nobody’s online, and offline messages can become tickets. Second, when AI replies aren’t available or the monthly allowance is used, the free keyword bot answers from your saved replies, so routine questions still get answered. That’s why it’s worth keeping your top ten facts as saved replies too.

Every website on the free plan gets a small monthly AI allowance, with AI packs if you need more, plus 100 visitors per website each month. See pricing for plans and packs.

If you’re just starting and want something simpler first, our guide to creating a starter bot from a fact sheet builds keyword answers in an evening. The same fact sheet becomes your AI knowledge later.

Checklist

Knowledge your bot can find

Knowledge progress

Frequently asked questions

Is a custom AI chatbot trained on my data?

Usually not in the technical sense. Most tools keep a general model unchanged and give it the relevant parts of your knowledge with each question. That’s why updating your text updates the answers straight away.

How much knowledge does a small business bot need?

Less than you think. Fifteen to forty short paragraphs covering your real questions beat a hundred pages of brochures. Quality and clarity matter more than volume.

Why does my bot answer with the wrong fact?

Usually the wrong chunk was found. Check whether the right paragraph uses the customer’s words and covers one topic, and whether another paragraph mixes several topics and wins by accident.

Should I upload my whole website?

You can, but check the pages first. Clear headings and one topic per section help. Pages with vague copy, old prices or tables usually need rewriting before the bot uses them well.

Does the chunk finder send my text anywhere?

No. It runs entirely in your browser and nothing is sent or saved. It uses word matching, which is stricter than real AI search, so treat weak matches as warnings.

Keep reading

Related guides

📘

Create Your Own Chatbot Free From a One-Page Fact Sheet: Write the Facts, Generate the Answers, Test Them Tonight (Free Starter Bot Writer)

Read the guide →

Your knowledge, answering customers

Give the AI Agent clear, findable knowledge, keep saved replies as a fallback and hand anything unclear to your team. Start free on livechatbot.in.