How to Train a Chatbot on Your Own Data in 6 No-Code Steps

Summary

• Training a chatbot on your data means connecting your help center, docs, and files so it looks up each answer there. No model gets retrained.

• Pick your route first. Your data, your comfort with code, and where the chatbot will answer decide the tool.

• Clean before you connect. The failure Wonderchat’s support team sees most is an old policy page answered correctly.

• Test with your most common questions and one your docs can’t answer, then open every source.

Your website, help center, PDFs, and Google Drive are the data. The chatbot looks up the answer there every time someone asks.

With a no-code tool, the setup takes six steps. Connect your site, add your files, set the role, set handover rules, test, and go live.

The hard part is the confident wrong answer a customer sees before you do. The version worth building is an AI customer service agent that shows its source and says when it doesn’t know.

What training AI on your own data means (RAG vs. fine-tuning)

No-code tools of this kind use retrieval-augmented generation. When a question arrives, the tool searches your content, pulls the passages that match, and the model writes the answer from them.

Fine-tuning is the other meaning of “training,” and it changes the model itself. OpenAI’s own guide to optimizing accuracy describes retrieval as giving the model “domain-specific context” and fine-tuning as teaching it a task so it performs “consistently and reliably.”

A grounded answer stays inside what you connected, and a source link lets you check it in one click.

Wonderchat answers only from what you connect and shows the source on every reply, with web and PDF-page citations. So “chatbot training data” here means your own pages and files.

The agent answers from the docs and links every page it used, so the answer can be checked.

No-code tool or Python script? Three questions to pick your route

Three answers decide the tool before you compare a single feature.

If your data sits in a database but rarely changes, export the table to CSV and upload it as a file, no code needed.

If a developer is part of the plan, you can build a RAG chatbot yourself and control every part of the retrieval.

As for where it answers, Wonderchat’s website widget, search bar, and live chat are on every plan, Free included. Slack and WhatsApp open from the Starter plan, Microsoft Teams from the Basic plan, and the same answer, with its source, shows up in each one.

How to prepare your docs before you connect them

Run two checks before you connect anything, one for pages that are no longer true and one for pages that disagree or that the chatbot can’t read.

Remove outdated pages and old policies first

The failure Wonderchat’s support team sees most is retrieval pulling an archived policy and answering from it correctly.

“Common things that are answered wrongly are old policies, old pricing, or outdated documentation,” says Luke Tan, CTO and co-founder of Wonderchat.

The agent did its job. It found a page, quoted the text faithfully, and cited the source, but the page was no longer true.

Say a refund policy from two years ago is still live at an old URL next to the current one. A customer asks about refunds and gets the old terms, with a source link that looks official.

The agent quoted the old page correctly and cited it, so the answer looked right and was wrong. An illustrative example; the policy and dates are made up.

Before you connect anything, find and remove:

• retired prices and plan names

• expired policies and old terms of service

• release notes that read like current rules

• pages for products or features you no longer offer

If an old page has to stay live for your records, add it under “Ignore URLs” when indexing in the Data Source instead. An asterisk at the end, like yourdomain.com/archive/*, covers a whole folder.

Wonderchat’s docs note that a deleted page can come back after a re-crawl, while an ignored URL is never crawled again.

Pages on the ignore list are never crawled, so a retired page can’t come back into an answer.

Then merge conflicting pages and move text out of images

More pages aren’t always better. Zendesk’s own help center warns that “having lots of sources can in some cases lead to reduced accuracy and increased latency.”

Two pages that answer the same question differently force the chatbot to pick one. Merge them, or decide which one is right and retire the other.

On the pages that make an agent’s answers worse, Luke Tan, Wonderchat’s CTO and co-founder, names “Conflicting or vague pages - Pages with information that is vague or conflicting with other pages.”

Then check the format.

Wonderchat’s documentation guide says it “reads text-based information best,” so text locked inside screenshots can lose its context. Put the text on the page itself.

How to train an AI agent on your own data in six steps

Before you start, have your help center address, the files customers ask about, and a list of your ten most common customer questions ready.

1. Connect your website and help center.

2. Add your files and Google Drive.

3. Set the role, tone, and fallback message.

4. Set when the agent hands over to a person.

5. Test the agent in the preview.

6. Add the agent to your website and channels.

1. Connect your website and help center

Paste the address of your site or help center, and the tool reads the pages. In Wonderchat, open your agent and go to Data Source.

Your site goes in as a root URL, and any page on the ignore list stays out of answers.

An asterisk at the end, like yourdomain.com/*, decides how much gets added.

Start with the help center. It’s already written to answer customer questions.

Each self-serve plan caps the pages per agent, at 100 on Free, 200 on Starter, 1,000 on Basic, and 2,000 on Scale. If your whole site runs past the cap, connect the help center and the pages customers ask about, and leave the blog out.

Website and help center pages are open on every Wonderchat plan, Free included.

If your help center sits behind a firewall or bot protection, allow requests that carry the X-Wonderchat-Identifier header. Wonderchat’s technical FAQ recommends it over allowlisting IP addresses.

`X-Wonderchat-Identifier: Wonderchat-Prod`

2. Add your files and Google Drive

Upload the files that hold answers your site doesn’t, like PDFs, Word, PowerPoint, Excel, CSV, and text files.

Files and Google Drive sit next to your website as sources, so one agent answers from all of them.

Match each source to the answers it holds.

To connect Google Drive in Wonderchat:

1. Open Settings, then Integrations, find the Google Drive card, and click Connect.

2. Sign in with the Google account that owns your Wonderchat subscription. A teammate’s own Drive can’t be added to the plan.

3. Open your agent, open More in Data Source, choose Google Drive, select the files, and click Update.

Only connect what you’d let a customer read. An internal pricing sheet sitting in the same folder can end up in an answer.

Before you upload, check each file for three things:

• text you can select and copy

• one current version, with older copies out of the folder

• no internal notes, draft prices, or customer data

If the files are sensitive, check where they’re stored. Wonderchat is EU-hosted in Frankfurt, with SOC 2 Type II, GDPR, and a HIPAA BAA available.

3. Set the role, tone, and fallback message

Tell the agent who it is and how to sound. Wonderchat’s controls cover role and tone, creativity (precise, balanced, or creative), preset messages, and a consent checkbox before chatting.

In Wonderchat, the role and its instructions sit under Model & Behavior in the agent editor. Pick the Customer support role, or write your own, like this one.

`You are the support assistant for [Company].`

`Answer only from the connected help center and docs, and link the page you used.`

`Keep answers short and plain.`

`If no page covers the question, say you can't find it and offer to connect the customer with the team.`

`Never state a price, refund term, or date that isn't on a connected page.`

The AI model is a setting too, from OpenAI, Anthropic, Google, or DeepSeek, and answers still come only from your content. The Free plan runs fast models, and Basic and Scale add frontier models.

Free runs fast models, and frontier models, which can use more than one AI reply per answer, open from Basic.

Set the greeting and the fallback message as well.

The fallback message matters most, because it’s what a customer reads when the docs run out.

The fallback message covers questions your docs don’t, and source links keep the other answers checkable.

4. Set when the agent hands over to a person

Decide when a person takes over. Wonderchat hands over with the full conversation attached, on any of three rules.

The agent can also collect an email so your team can follow up when nobody is online. Handovers land in Wonderchat’s help desk as tickets with the full chat, or in live chat, both on every plan.

In Wonderchat, the rules sit under Handover and Live Chat in the agent editor, with handover email addresses under Handover and your team’s online hours under Live Chat.

Any trigger you turn on opens the handover form, so a customer doesn’t have to hunt for a person.

Teams on Zendesk or Freshdesk can keep them, with ticket sync from the Basic plan (up to 50 tickets until Scale), so use Wonderchat’s help desk or keep yours.

5. Test the agent in the preview before you publish

Run the four-question test before any customer sees the agent, while you can still fix what it gets wrong.

In Wonderchat, open the agent and ask your questions in Real time preview of your chat widget on the Profile page, or click Preview Widget in the top bar.

Asked about a status page no page covers, the agent says it can’t find one and offers a live agent.

6. Add the agent to your website and channels

Share a hosted link first, then add the widget with one snippet, pasted where your analytics code goes.

The widget is on every plan, Free included, while Slack and WhatsApp open from the Starter plan and Microsoft Teams from the Basic plan.

Install gives you one snippet for your site, and the widget switch is already on.

In Wonderchat, it takes three moves:

1. Click Share Link in the agent editor’s top bar, and send the hosted page to your team to try.

2. Click Install, copy the snippet, and paste it into your site’s HTML just before the closing </body> tag.

3. Open your site on a desktop, a tablet, and a phone, and ask one question on each.

In a typical page, the snippet sits here:

`<!-- your analytics code -->`

`<!-- paste the Wonderchat snippet from Install here -->`

`</body>`

`</html>`

Wonderchat’s docs have install guides for WordPress, Wix, Webflow, Shopify, and Drupal. On WordPress, the snippet goes in the Scripts in Footer box of the Header and Footer Scripts plugin.

Set your expectations from real numbers. 80% of Wonderchat agents that go live do so the same day (384 of 478), while knowledge bases over 500 pages take a median of 15 days, so give a large help center more time.

When you’re ready to grow past the free plan, see plans for what each one includes.

Test your AI agent the way your customers will

Whatever a tool calls training, the test that matters is the question your docs can’t answer. Ask four questions, in this order:

1. The question customers ask most.

2. The second most common one.

3. A pricing, refund, or policy question, where a wrong answer costs you.

4. One plausible question your docs don’t cover, asked on purpose.

For the first three, open every source link and check that the page says what the answer says. For the fourth, a pass is a clear “I can’t find that” with a person offered.

An honest “I can’t find that” is the fallback message you set in step 3. When no page covers a question, Wonderchat sends that message and offers a person.

Sourced answers to your common and policy questions, plus an honest “can’t find that” for the fourth, mean the agent is ready.

How to keep your AI agent’s answers up to date

Every question the agent can’t answer points to a missing or outdated page. In Wonderchat, those questions land in one review inbox, sorted into user-reported answers, knowledge gaps, quick fixes, conflicts, and model issues.

A question the docs miss goes to a person first, and the review inbox turns it into an answer for next time.

Start with the refund question from the cleanup step. If a customer still gets the two-year-old terms, paste the current policy’s URL in Review and the agent answers from it from then on.

Check Review every week, because a stale answer looks right to everyone except the customer.

Your team fixes each one once. Open the item in Review, go to Where does the answer live?, and paste the URL, pick a file, or write the answer so the agent uses it from then on.

For answers that don’t belong on a page, add a correction. A correction is an approved question and answer the agent uses next time for questions with the same meaning or an exact match.

Wonderchat retrains on your website every month on Starter and Basic and every week on Scale. On Free, run a manual recrawl after you change a page.

Stale answers are a real complaint. Across 1,309 complaints in reviews of 18 vendors, which Wonderchat read in September 2026, stale knowledge accounted for 38.

How to train an internal AI agent on Google Drive or SharePoint

The same steps work when the people asking are your team. Wonderchat’s Internal AI Agent, the third agent on the same engine, answers them in Slack from Starter or Microsoft Teams from Basic.

Cleanup matters even more inside a company. Internal folders keep drafts, old versions, and private files side by side.

Your team asks in Slack or Teams and gets the answer with the file it came from.

Connect only the folders that hold current, approved answers, so the file linked in Slack or Teams is the right one.

When Chatbase or Botpress is the better choice

On usage per dollar, Chatbase comes out ahead on paper. On 30 October 2026, its paid plans listed more message credits for the price than Wonderchat’s plans list AI replies, though the two meters don’t count the same work.

Chatbase’s gap tool, Backstage, comes with a draft of the fix for each gap. Wonderchat doesn’t draft pages, so your team writes the answer once.

Botpress is the better pick for a team with a developer who wants to build its own logic. The flexibility is worth the extra setup only when someone on the team will own the build.

How Wonderchat answers from your own docs

Training a chatbot on your own data comes down to two habits, connecting clean pages and testing the question your docs can’t answer.

A customer asks about refunds, and Wonderchat answers from the current policy page with the link. A question no page covers gets an honest “I can’t find that,” a person, and the whole conversation handed over.

That question lands in the review inbox, where your team answers it once. The free plan refills 50 AI replies every month with no credit card, enough to see the source and the honest miss on your own docs.

Start for free, paste your help center’s URL, and ask your hardest question, then one your docs can’t answer, or book a demo to see it on your docs.

Frequently asked questions

Can I train ChatGPT with my own data?

A custom GPT can answer from your files. OpenAI says Knowledge “lets your GPT use information from files you upload,” up to 20 files of 512 MB each. For a custom GPT for business on your website, you’ll also want source links and a handover to a person.

How do I train AI on my own data for free?

Start with a tool that has a free plan, and connect your help center. Wonderchat’s free plan refills 50 AI replies every month with no credit card, which is enough to run the four-question test before you pay anything.

How do I program my own chatbot?

With code, you choose a model, a way to search your documents, and a chat interface, then connect them yourself. Without code, a no-code tool handles all three. You connect your content and set the rules.

Will a chatbot trained on my docs make things up?

A chatbot trained on your docs can still be wrong, and the most common cause Wonderchat’s support team finds is an outdated page. Pick a tool that answers only from your content, shows the source on every reply, and says when it can’t find an answer.

How long does it take to train a chatbot on my data?

80% of Wonderchat agents that go live do so the same day, while knowledge bases over 500 pages take a median of 15 days to go live.