Chatbase Tutorial: Build a Custom ChatGPT for Your Website
Uploading a website and copying an embed code creates a chatbot, but it does not create a trustworthy support process.
Some links on this site are affiliate links. If you buy through them, we earn a commission at no extra cost to you. We only recommend tools we would deploy ourselves.
$ ls ./sections
This Chatbase tutorial builds a narrow website assistant that answers from approved material and admits when the answer is missing. That last behavior matters more than the chat bubble color.
Chatbase tutorial setup
Create an agent in Chatbase, add only current support pages and documents, then set the instruction to answer from those sources. Retrieval means the system finds relevant passages before asking the model to respond; explain that once to stakeholders and skip the fake “trained our own AI” language.
Use a boundary like this:
Answer only from the supplied sources. If the sources do not contain the answer,
say that you do not have enough information and offer the human support route.
Never invent prices, policies, availability, or account details.
Build the custom ChatGPT test set
Write ten normal questions, five ambiguous ones, and five the sources cannot answer. Confirm citations point to the intended page and that unknown questions escalate cleanly. Dante AI is the adjacent option worth comparing for an agency-style knowledge assistant.
Embed and troubleshoot Chatbase
Copy the provided embed snippet into the site’s global layout or tag manager. If the widget does not appear, inspect Content Security Policy and consent-manager blocking. If answers are stale, resync the source and remove obsolete documents rather than adding a prompt that asks the model to guess which policy is newer.
Connect the assistant to a human only after the no-code agent guide defines what data may leave the conversation.
Chatbase
Custom GPT chatbots trained on your own data.
Dante AI
Custom AI agents trained on websites, files and business knowledge.
Found the fix? The tool that ends the problem is one click away.
The Stack