How to train a chatbot on your own data: what "training" really means, how to prepare your documents and website, and how to test it
This is for owners who want a chatbot that answers from their own prices, policies and services instead of the internet's guesses, and who do not write code. You will learn what actually happens when a tool says it is "trained on your data", how to get your website, PDFs and spreadsheets ready, and how to test the result before a customer sees it.
Quick answer
Most business chatbots are not retrained on your data. They store your documents, look up the most relevant passages when a customer asks, and write the answer from those passages. This is called retrieval. To "train" one without code, upload clean, current content (website, PDFs, price lists, Q&A pairs), rewrite anything vague, then test it with 20 to 30 real customer questions before going live.
What does "train a chatbot on your own data" actually mean?
In almost every no-code tool, it means "give the chatbot your documents to look things up in". The AI model itself does not change. What changes is the library it is allowed to read before it answers.
Here is the plain-English version. A customer asks, "Do you do root canals on Saturdays?" The system searches your uploaded content for the passages most likely to answer that (your opening hours, your treatment list), hands those passages to the AI along with the question, and asks it to answer using only what it was given. The research name for this is retrieval-augmented generation, introduced in a 2020 paper by Lewis and colleagues, whose authors pointed out that knowledge stored inside a model is hard to update and hard to trace back to a source. Retrieval fixes both: change the document and the answer changes the same minute, and every answer can be traced to a passage.
We have a longer explainer on how retrieval works if you want the mechanics. For setting one up, you need just one idea: the chatbot can only be as good as the content it finds. Training, in practice, is content work.
Retrieval, fine-tuning or training from scratch: which one do you need?
For a small business answering customers, retrieval is the right choice nearly every time. Fine-tuning and training from scratch solve different problems, cost far more, and do not keep prices or policies current.
| Retrieval (looking things up) | Fine-tuning | Training a model from scratch | |
|---|---|---|---|
| What changes | The documents the AI reads before answering | The model's habits, style and format | Everything; you build the model |
| Good for | Facts that change: prices, hours, policies, stock lists | A fixed writing style or a narrow classification task | Research labs and very large companies |
| Update a price | Edit the document; live in minutes | Prepare new examples and fine-tune again | Not practical |
| Can show where an answer came from | Yes | No | No |
| Skills needed | None in a no-code tool | A developer and a labelled dataset | A research team |
| Right for a clinic, salon or agency? | Yes | Rarely | No |
A useful test: if the thing you want the bot to "learn" could change next month, it belongs in retrieval. If you want the bot to sound a certain way, that belongs in its instructions, not in a training run. Our guide to writing agent instructions covers that half.
How do you prepare your data?
Gather what customers actually ask about, remove what is out of date, and make each fact appear once, clearly, in plain text. Each source type has its own traps.
Your website
Good for services, about, locations and policies. Traps: pages that say "contact us for pricing", old promotions still live, and important facts hidden in images or sliders. Crawl the pages that answer questions, not your blog archive from 2019.
PDFs and brochures
Good for menus, treatment lists, course outlines. Traps: scanned images with no real text, two-column layouts that read out of order, and three versions of the same brochure. Upload only the current one.
Price lists and catalogs
Use a table or CSV with one row per item: name, price, what is included, any condition. Traps: "from $80" with no explanation of what changes the price, and prices that differ between the PDF and the website.
Q&A pairs
The highest-value source. Write the question the way a customer types it and the answer you want given. Traps: answers that say "it depends" without saying on what.
Past chats are a goldmine for finding the questions, but do not upload a raw chat export: it is full of personal data and one-off answers. Mine it instead, as described in using past WhatsApp conversations. For what belongs in the library at all, see what to put in your knowledge base.
Rewriting for lookup: before and after
Content written for a brochure often fails in a chatbot, because the search finds passages by meaning and a passage that hints rather than states gives the AI nothing to quote. Rewrite so each passage answers one question on its own.
| Before (hard to find, hard to quote) | After (one clear fact per passage) |
|---|---|
| "We offer flexible scheduling to suit your lifestyle." | "Opening hours: Monday to Friday 9am to 7pm, Saturday 9am to 1pm. Closed Sunday and public holidays." |
| "Ask us about our competitive packages!" | "Basic cleaning package: $120, covers up to 3 bedrooms, about 3 hours. Each extra bedroom adds $25." |
| "Terms and conditions apply to all bookings." | "Cancellation: free if you cancel 24 hours or more before. Less than 24 hours, the $30 deposit is kept." |
| "See our FAQ for parking info." (on a page with no parking info) | "Parking: free in the basement car park, entrance on Jalan Example. Two accessible bays near the lift." |
Three rules make most of the difference: name the thing in every passage (write "the deposit", not "it"), put numbers in words a customer would use, and keep conditions next to the fact they change. The examples above are hypothetical; swap in your own.
Step by step, without code
This is the order we recommend in any no-code tool. It takes most owners an afternoon for the first pass.
- List the questions. Scroll your last month of WhatsApp, website chat and email, and write down the 25 questions you answer most. These are your test set later.
- Collect the sources. Gather the website pages, PDFs, price list and any written policies that answer those questions. Throw out anything outdated.
- Fix the gaps and duplicates. Where an answer exists in two versions, keep one. Where it exists nowhere, write a short Q&A pair.
- Rewrite vague passages. Turn marketing lines into plain facts with numbers and conditions, one fact per passage.
- Upload and connect. Add the website URL to crawl, upload the PDFs and price list, and paste in the Q&A pairs.
- Write the instructions. Tell the bot who it serves, what it should do (answer, book, collect details), and to say so and hand over when the answer is not in its knowledge.
- Test with real questions. Ask the 25 questions in customers' own words, score each answer, and fix the content behind every failure.
- Keep teaching. After launch, review the questions it could not answer each week and add the answers once.
How do you test a chatbot trained on your data?
Ask it the real questions in the messy way customers type them, then score each answer on three things: correct, complete, and honest when it does not know. Do not test with questions copied from your own FAQ, because they match your wording too neatly.
| Test message (as a customer would type it) | Pass looks like | Fail usually means |
|---|---|---|
| "how much for whitening" | Quotes the exact price from your list, with any condition | Price missing or listed twice with different numbers |
| "open sunday?" | States Sunday hours or that you are closed | Hours only in an image or a footer the crawler skipped |
| "can i bring my dog" | Says it does not have that information and offers a person | It guessed; the instructions do not forbid guessing |
| "what's the cheapest option for 2 people" | Compares the right items and shows the working | Price list lacks "per person" or "per couple" labels |
| A voice note or a photo of a product | Understands it and answers or asks one clarifying question | The tool cannot read media; check before you choose it |
Aim for every factual answer to be right before launch, and for the "I don't know" cases to hand over cleanly rather than invent something. Our 20-message pre-launch test goes deeper, and why AI makes things up explains the failures you are guarding against.
How ReplyKit does it
ReplyKit uses retrieval, the same approach described above, and it is built so the owner never has to think about the word. Plans run from $29 to $99 a month with the AI replies included; the pricing section lists what each one covers.
- Four ways in. Crawl a website URL, upload PDFs and price lists, add text and Q&A pairs, or import a product catalog as CSV.
- Answers only from your knowledge. If the answer is not there, the agent says so and escalates to you instead of guessing. You are notified, and you can take over the chat in one click.
- The teach loop. Every question the agent could not answer is listed, so you write the answer once and it is used from then on. More on that in knowledge gaps and the teach loop.
- Same knowledge, two channels. It answers on WhatsApp and on your website through one line of code, and each website assistant can have its own instructions and knowledge.
- Beyond answers. It can offer real open slots and book them, capture a lead's name, need, budget and timing, and understand voice notes and photos in over 30 languages.
When ReplyKit is not the right choice: if you need answers pulled live from your own database (an order status, a bank balance), a developer integration fits better, because ReplyKit answers from the content you give it and has no webhooks or API connections. If your main goal is an internal knowledge assistant for staff, a general workplace AI tool may suit you more. And if you want to see it before reading further, try the live demo on our homepage, or start a free trial and upload your own price list.
Frequently asked questions
Can I train a chatbot on my own data for free?
You can test it for free on most no-code platforms through a trial or a limited free tier. ReplyKit has a 7-day free trial with 300 AI replies and no credit card. Running it for real customers usually costs a monthly fee because every answer uses AI processing.
Do I need to know how to code?
No. With a retrieval-based tool you upload files, paste a website address and write answers in plain English. Code is only needed if you want answers pulled live from your own database or systems.
Does the chatbot learn from customer conversations automatically?
Not in a well-run setup, and that is deliberate. Letting a bot rewrite its own knowledge from chats would spread one wrong answer to everyone. Instead, unanswered questions are listed and you approve the answer once.
How much data do I need?
Less than you think. The answers to your top 25 questions, a current price list and your key policies cover most small businesses. Quality matters more than volume: one clear page beats ten overlapping ones.
What is the difference between training and fine-tuning a chatbot?
In everyday use, "training on your data" means giving the bot documents to look up, which is retrieval. Fine-tuning changes the model itself with example conversations, needs a developer, and is a poor way to store facts like prices because they change.
Is my data safe if I upload it to a chatbot?
Only upload what you would be comfortable showing a customer, because the bot may quote any of it. Keep out personal records, ID numbers, internal costs and passwords, and check the vendor's privacy terms on how uploaded content is stored.
Upload your price list and ask it a question
Set up an agent on your own content in about 15 minutes. 7-day free trial, 300 AI replies, no credit card.
Sources and further reading
- Lewis et al., Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks (arXiv, 2020) · the paper that named retrieval-augmented generation and noted that updating and tracing a model's built-in knowledge are open problems
- ReplyKit, How retrieval works in AI agents · the mechanics in more depth
- ReplyKit, What to put in your AI agent's knowledge base · what to include and leave out
ReplyKit is an independent product and is not affiliated with or endorsed by Meta or WhatsApp. Third-party information was checked in October 2026 and may change. This article is general guidance, not legal or technical advice for your specific setup.