Building a customer bot on Chatbase: sources, instructions and the retraining habit

Chatbase is popular with small businesses and agencies because it’s quick: point it at a website, upload a few documents, paste the widget code, and a bot is answering questions the same day. It’s flexible too, letting you choose the underlying model, write your own instructions and compare versions side by side.

Speed has a side effect. The bot knows your content as it was on the day you trained it. This note covers how to set up a Chatbase bot for a small service business or shop, and the routine that keeps it current.

What it reads and what it costs

A Chatbase agent learns from sources you add: website pages it crawls, uploaded files, plain text you paste in, and question-and-answer pairs. Plans come with monthly message credits, and each AI response uses credits depending on the model you choose, from one credit on basic models to several on the largest ones.¹²

Two practical consequences:

• choosing the largest model doesn’t fix wrong sources, and it uses credits faster;
• a bot that answers every visitor’s “hi” still uses credits, so a clear greeting and quick hand-over for small talk saves money.

1. Choose sources on purpose

The easiest setup is to crawl the whole site. It’s also the one that causes most wrong answers, because the crawl picks up old news posts, last year’s price list in a PDF, and every variant of every page.

For a small business, start narrow:

Add	Leave out
Services or product pages with prices	Blog posts and news, unless evergreen
Opening hours and locations	Old PDFs: menus, price lists, brochures
Booking, cancellation and payment rules	Team pages and job ads
Delivery and returns, for shops	Duplicate pages for other regions
A short FAQ written for the bot	Legal pages as the only source on a topic

Then add a text source written specifically for the bot: ten to twenty plain sentences with the facts customers ask about most, each with its condition. “A first appointment costs €45 and lasts 30 minutes. Cancellations less than 24 hours before are charged in full.” This becomes the piece the bot finds first.

2. Write instructions that limit, not decorate

Chatbase lets you set the agent’s instructions. Spend them on behaviour, not personality:

• “Answer only from the provided sources. If the sources don’t answer the question, say so and give our email address.”
• “Never offer discounts or prices that aren’t in the sources.”
• “For questions about an existing booking or order, ask the customer to email us.”
• “Don’t guess opening hours for holidays; say that holiday hours are on the contact page.”

The last one matters more than it seems. Asked about parking, a holiday opening time or anything the sources don’t cover, bots on most platforms tend to give a plausible guess. A specific instruction for the common gaps is more reliable than a general one.

3. Test with Compare, then as a customer

Chatbase’s Playground lets you chat with the agent before deploying it, and its Compare mode runs several versions side by side with different models, temperatures and instructions.³ Use it to choose a model and tune instructions.

Then test the deployed bot on your live site as a customer would, with the questions your customers really ask and the follow-ups they add. For each answer, note the page that should have answered it. Compare mode tells you which version answers better from your sources; only this second test tells you whether the sources match your site.

4. Retrain when pages change

This is the habit most Chatbase bots are missing. When you change a price, a schedule or a policy on your website, the bot’s copy of that page doesn’t necessarily change with it. Until you retrain or refresh the source, the bot keeps quoting the old version, confidently and consistently.

Make it a rule: every website change to prices, hours or policies is followed by a source refresh and three test questions on that topic. Write it on the same checklist you use to update the website, so it happens in the same sitting.

5. A made-up example

A cleaning company offers regular and one-off cleaning. In August it raised the one-off price from €28 to €32 an hour and added a three-hour minimum. Its Chatbase bot was trained in June.

In September a customer asks: “How much for a one-off clean of a two-bedroom flat, about two hours?” The bot answers: “One-off cleaning is €28 per hour, so about €56 for two hours.” The correct answer is three hours minimum at €32: €96. The customer books expecting €56.

Nothing was wrong with the model, the instructions or the platform. The source was three months old. A refresh in August and one test question would have caught it.

The short version

1. Add sources on purpose; leave out old PDFs and blog posts.
2. Write a short bot-only text source with your key facts and conditions.
3. Use instructions to forbid guesses and discounts.
4. Test with Compare, then as a customer on the live site.
5. Refresh sources every time a page changes.

Our Chatbase page lists the mistakes we see most on Chatbase bots. If you’d like the checking done every week, that’s Assistant Care; for the general mechanism, see how an AI assistant answers.

Footnotes

1. Chatbase documentation, FAQ on message credits, chatbase.co/docs. ↩

2. Chatbase, pricing, chatbase.co/pricing. ↩

3. Chatbase changelog, “Introducing Compare mode in the Playground”, chatbase.co. ↩