How to train an AI chatbot on your own data

When people ask how to train an AI chatbot on their own data, they usually picture the model studying their website until it knows it by heart. That is not how most support agents work, and knowing what happens instead explains most good and bad answers. This guide covers the mechanism, the mistakes that cause wrong answers, and how to set up your sources in Centillion.

​

What “training” means for a support agent

A support agent does not rewrite the language model. The model stays the same for every business. What changes is the material the agent reads before it answers. The process has two halves. When you add a source, the text is extracted, split into passages and indexed, so it can be searched later. This indexing step is what most products call training. When a customer asks a question, the agent searches that index, picks the passages that look most relevant, and writes its answer from them. The technique is called retrieval: the model is given the right page at the right moment, every time. Two things follow from this design:

• Updates are fast. Change a page, re-index it, and the next answer uses the new text. No model has to be retrained.
• The agent only knows what it finds. If the answer is not in your sources, or the search picks the wrong passage, the answer suffers. Most quality work happens in the sources.

​

What to feed it

Good sources are the ones your team already uses to answer customers:

• Your website. Product pages, pricing, shipping and returns, the FAQ.
• Policy documents. Terms, warranty conditions, internal guides your team follows.
• Saved answers. The replies your team types out every week, written once as a question with variations and an approved answer.
• Working notes. Pages in a tool such as Notion where your team keeps current procedures.
• Closed support tickets. Real questions and how your team solved them.

Leave out what a customer should never see: internal pricing notes, personal data, drafts that were never approved. Anything in the index can end up in an answer.

​

Why answers go wrong

Two pages disagree. An old pricing page and a new one both match the question, and the agent quotes the old one. Contradictions are the most common cause of a wrong answer. Remove or exclude the outdated page; do not rely on the instructions to pick the right one. Noise crowds out the answer. The search keeps only a few passages per question. Navigation menus, tag pages, duplicate collection pages and legal boilerplate all compete for those slots. Import the pages that hold answers, and skip the rest. The content is not text. Images inside documents are usually not read, and a scanned PDF is a picture of text. Convert it to selectable text first. The page changed and the agent did not. Websites change outside the support tool. Without a scheduled re-check, the agent keeps answering from the version it saw last. Nothing to answer from. An agent with no knowledge still writes fluent sentences. That is worse than silence. Add sources before you switch a channel on, and tell the agent in its instructions to say when it does not know.

​

How to check it

Ask the questions your customers ask most, in their words. Then ask a question the agent should not be able to answer, and check that it says so. When an answer is wrong, look at which sources it used before you touch the instructions. The fix is almost always in the knowledge.

​

Set it up with Centillion

In Centillion, knowledge is what your agent answers from. You add sources under Knowledge base → Data sources, and Centillion imports them and trains the agent automatically after every change, on every plan. The full reference is Adding knowledge. How an answer is put together in Centillion:

• Exact questions first. If the question matches a variant of a Q&A entry, that entry is used without searching. Case and punctuation are ignored.
• Otherwise a search. The agent looks for passages by meaning and by keyword at once, merges the two lists, and keeps the best six. Those passages are what it writes from.
• Draft edits stay on the branch. Removing a source from a draft or excluding its pages changes that draft. Publish to use the change in new customer conversations. Removing a source from the workspace stops its retrieval in every version.
• Only changes are re-indexed. Unchanged passages are reused, and a failed build keeps the previous training available.

​

Questions

​

Do I need to retrain the agent after adding a source?

No. Adding, editing, removing or restoring a source starts training on its own. The source shows Synced when the branch draft can use it, and you can close the page while it runs. Publish the branch to use its updated knowledge in new customer conversations.

​

How much can I add?

Storage is counted on the text Centillion extracts, per agent: 1 MB on Free, 10 MB on Hobby, 20 MB on Standard and 40 MB on Pro. The size of the files you upload does not count.

​

Why does the agent quote an old price?

Most likely two sources disagree. Find the page that still holds the old figure, then remove it from the branch draft or exclude it from the website source. Wait for Synced, then Publish. New customer conversations use the updated knowledge; existing conversations keep their saved version.

​

Can it read images or scanned PDFs?

Images inside files and Notion pages are not read, and scanned PDFs need converting to selectable text before you import them.

​

Does new knowledge reach a published release?

Every agent starts with version control on. Source changes apply to the selected branch draft. Published versions keep their saved knowledge; publish a new version to use the new content in new customer conversations.

​

Try it

Add your website through its sitemap and two policy documents, wait for Synced, and ask the three questions your team answers most. Then ask one it should not know. Or ask the assistant on this page.

​

Related