How to Build a Chatbase Chatbot From Your Docs
Chatbase builds a support chatbot from material you already have. Upload docs or point it at URLs, wait for indexing, write a short system prompt, and you have a working widget to embed in roughly 20 minutes.
No training pipeline, no vector database to run. That is the appeal, and it is also the limit: the bot is only as good as the pages you feed it. Garbage docs produce confident garbage answers.
What is Chatbase and what can it ingest?
Chatbase is a hosted service for building retrieval-augmented chatbots. Your content gets chunked and indexed, then answers cite chunks at runtime.
Accepted sources:
- PDF files and DOCX uploads.
- Website URLs, crawled on demand.
- Plain text pasted directly.
- Notion pages and help-centre connectors on paid plans.
- Q&A pairs you write by hand for edge cases.
Each source counts against a character or page allowance tied to your plan. The free plan is small; the entry paid tier is in the low tens of dollars per month with a monthly message cap.
How do you create and train a bot?
- Sign up and create a new chatbot. Give it a name customers will recognise, not an internal codename.
- Add sources one at a time. Start with your help centre root URL, then the three PDFs that answer the top support questions.
- Wait for indexing. A 50-page site takes a few minutes; a 400-page one can take the better part of an hour.
- Write the system prompt. Two or three sentences: what the bot is, what it must never do, and what it should say when unsure.
- Test with your ten most common support questions before anyone else sees it.
- Embed the widget snippet on your site or connect the hosted chat page.
Step five is not optional. The first test round almost always reveals two docs pages that contradict each other. The bot will then alternate between both answers depending on phrasing.
What should the system prompt actually say?
Keep it short and restrictive. A useful default:
You are the support assistant for [Product]. Answer only from the provided documentation. If the docs do not cover the question, say you do not know and suggest emailing support. Never invent pricing, deadlines, or policy details.
That last sentence does most of the work. Unconstrained bots invent refund policies, which is a legal problem, not just an annoyance.
Add a tone line if your brand has one. Skip personality instructions; they consume prompt space and change nothing customers notice.
How do you embed the bot without slowing your site?
Chatbase gives you a script tag for the floating bubble, plus options for a full-page chat or an inline panel.
Practical notes:
- Load the script with
deferso it does not block first paint. - Set the launcher colour and position in the console to avoid layout jumps.
- For a docs site, an inline panel beats a bubble; people expect chat in context.
- Hide the bubble on your checkout flow unless the bot is trained on billing answers.
Embed the script once in your global footer, not per page. Duplicated snippets cause two bubbles to appear and confuse the analytics.
How accurate is a Chatbase bot in practice?
For a well-documented product, quality is high on factual questions and weak on questions requiring synthesis across three pages. It answers “what are your API rate limits” reliably. It stumbles on “why did my last invoice go up” because that needs data, not prose.
Expect to add hand-written Q&A pairs for the five questions users ask most where your docs are thin. Those pairs outrank uploaded PDFs in retrieval, and they are the fastest fix for wrong answers.
Who should not use Chatbase?
Anyone whose support answers depend on live account data. Chatbase answers from static content; it cannot look up an order or a subscription. You need a bot wired into your backend for that.
Also skip it if you cannot maintain docs. A chatbot trained on a stale help centre confidently repeats last year’s pricing. And if your content is highly regulated, review every answer path: retrieval bots can quote outdated compliance language verbatim.
Quick Summary
- Chatbase builds a support bot from PDFs, URLs, Notion pages and pasted text in about 20 minutes.
- Write a restrictive system prompt; the “never invent policy details” line prevents most bad answers.
- Test your ten most common support questions before embedding the widget.
- Hand-written Q&A pairs fix wrong answers faster than adding more documents.
- It answers from static docs only; anything needing live account data needs a different integration.
