Build a chatbot trained on your own content
Paste your text, upload a PDF, or import a URL, and your agent answers from your content — no chunking, no embeddings, no vector database to host. This guide covers what “trained on your content” really means, the four ways to load it, and where the approach stops working.
What “trained on your content” actually means
Nobody re-trains a language model on your website. Training a model costs millions and takes weeks — that is not what happens when you add a PDF to a chatbot, whatever the marketing page says.
What actually happens is simpler and better: your content is handed to the model as context, every time it answers. The model already knows how to read and reason; you supply the facts. This is why an edit takes effect on the very next message, and why the agent can quote a price you changed five minutes ago.
The set of things you hand over is the knowledge base— a list of entries, each with a title and a body. Filling it well is the single highest-leverage thing you can do for answer quality. Everything else on this page is about how to fill it.
The four ways to feed your content
All four live in the same place in your dashboard, and they mix freely: paste a few answers, upload last year’s price list, and let WordPress push the rest.
- 1
Paste your text
Write or paste anything the agent should know — your services, your pricing, your policies, the answers you repeat every week. Each entry gets a title and a priority.
- 2
Upload a file
Drop a PDF, a Markdown file, a text file, an HTML export or a CSV (up to 5 MB). The text is extracted and stored as a knowledge entry. A CSV becomes one entry per row.
- 3
Import a URL
Give the address of a page and its content is fetched, stripped of navigation and scripts, and converted to clean Markdown. One page per import.
- 4
Connect WordPress
With the official plugin, every post and page you publish is pushed to the knowledge base automatically, and updated whenever you edit it. Nothing to re-import by hand.
What the import does not do
- No site crawl, no sitemap import. The URL import reads one page. To cover a whole site automatically, use the WordPress sync or the bulk API.
- No OCR. A scanned PDF with no text layer comes in empty.
- No file storage. A PDF is read for its text; the file itself is not kept.
Why there is no vector database
The standard recipe for a “chatbot on your documents” is to split them into chunks, turn each chunk into an embedding, store those in a vector database, and at question time retrieve the chunks closest to the question. It is a good design — for a corpus of hundreds of thousands of documents.
A website’s knowledge base is not that. It is a few dozen entries. They fit in the model’s context window, so we put them there. No chunk boundaries cutting a sentence in half, no re-indexing job to run after every edit, no similarity threshold to tune, no vector store to host and pay for.
Each entry can carry a short summaryalongside its full body. The summaries go into the agent’s context, so it always knows what it knows. The full body is pulled on demand, only for the entry a question actually touches. You get the breadth of the whole knowledge base at the price of a table of contents.
Where this stops working. Somewhere past a few hundred substantial entries, the context stops being the right place to keep everything, and vector search earns its complexity. If you are indexing a decade of support tickets, use a vector store. If you are answering questions about your own site, you do not need one, and paying for one buys you nothing.
How the agent finds the right answer
- It starts with the map.Your entries — summarised where you gave a summary — are in front of the agent before the visitor even types.
- It opens what it needs. When a question calls for detail, the agent looks the entry up by title and reads the full body. If it finds nothing under that title, it gets back the list of titles it doeshave — so it can correct itself instead of guessing.
- It answers from a named source.The answer traces back to an entry you wrote, which means you can fix a bad answer by fixing the content — not by wrestling with a prompt.
One practical consequence: titles matter. Call an entry “Refund policy” rather than “Document 12”. It is how the agent finds it.
Keeping answers grounded
A model with nothing to go on will fill the silence. A model holding your content answers from it, and is instructed to say it does not know rather than improvise. That is the whole point of the knowledge base, and it is why an hour spent writing good entries beats a week spent tuning a prompt.
Two safeguards sit on top: the agent can escalate to a humanthe moment a question leaves its scope, and visitors can be shown a disclaimer that they are talking to an AI. Be honest with yourself about the rest — no configuration makes a language model incapable of ever being wrong. Grounded in your content, with an escape hatch, it is reliable enough to put in front of customers. Read more in the trust & safety page.
Limits and pricing
Plans cap the number of entries, not their size. A 40-page handbook is one entry. Uploads are capped at 5 MB per file.
| Plan | Knowledge base | Price |
|---|---|---|
| Free | 5 entries | €0 |
| Starter | 20 entries | €29/month |
| Builder | 100 entries | €49/month |
| Pro | 100 entries | €79/month |
| Agency | Unlimited | €199/month |
Running WordPress?
The official plugin pushes every post and page to the knowledge base when you publish, and updates it when you edit. No crawler, no re-import, no stale answers. See the AI agent for WordPress.
Frequently asked questions
Can I train a chatbot on my own website?
Yes. You import your pages into the agent’s knowledge base — one URL at a time, or automatically if your site runs on WordPress — and the agent answers from that content. Nothing is re-trained: your content is given to the model as context every time it answers, which is why an edit takes effect immediately.
Does it read PDFs?
Yes. Upload a PDF (up to 5 MB) and its text is extracted and stored as a knowledge entry. Only the extracted text is kept — the original file is not stored, and images or scanned pages without a text layer will not be readable.
Can it crawl my entire site?
No. The URL import fetches one page at a time; there is no crawler and no sitemap import. If your site runs on WordPress, the official plugin syncs every published post and page automatically, which covers the whole site without a crawler. Otherwise, import the pages that matter — in practice a handful of them carry most of the answers.
Do I need a vector database or embeddings?
No. Your knowledge base is passed to the model as context, and the agent looks up the full text of an entry on demand. There is nothing to chunk, no embeddings to compute, no vector store to host or pay for. This works because a website’s knowledge base is a few dozen entries, not a few hundred thousand.
Is this RAG?
It depends what you mean. If RAG means "the answer is grounded in your documents rather than the model’s memory", then yes — that is exactly what happens. If RAG means "embeddings and vector similarity search", then no: we retrieve by loading your content into context and by looking entries up by title, not by cosine distance.
How many documents can I add?
The limit is on the number of entries, not their size: 5 on the Free plan, 20 on Starter, 100 on Builder and Pro, and unlimited on Agency. A single entry can hold a long document, so a 40-page handbook counts as one entry.
What happens when I update a page?
On WordPress, the change is pushed automatically when you hit publish. Everywhere else you re-import the URL or edit the entry from the dashboard. There is no re-indexing step to wait for: the next conversation already sees the new version.
Will the agent invent answers about things I never wrote?
That is what the knowledge base is there to prevent, and it is the main reason to fill it properly. The agent answers from the content you gave it and is instructed to say it does not know rather than improvise. You can also have it escalate to a human as soon as a question falls outside its scope. No setup makes a language model literally incapable of being wrong — but grounded in your own content, and with an escalation path, the risk becomes small enough to run in production.
Can I keep some content for a specific audience?
Yes. Entries can be scoped to a segment, so a given agent — or a given part of your site — only sees the knowledge it is supposed to see. Useful when the same account serves several brands, languages or customer tiers.
Does it work with my documentation site?
Yes. Import the pages that answer real questions — getting started, pricing, limits, troubleshooting — rather than the whole reference. A documentation site’s value for an agent is concentrated in a small number of pages; the API reference is usually better served by a link than by a paraphrase.
Can I add content programmatically?
Yes. The bulk endpoint accepts up to 50 entries per call and upserts them on your own identifier, so re-running your script updates entries instead of duplicating them. That is how the WordPress plugin itself stays in sync. See the developer guide.
Is my content used to train AI models?
No. Your knowledge base is sent to the model provider to produce an answer, and it is not used to train their models. It stays yours: you can edit or delete any entry at any time, and deleting it removes it from every future conversation.
Train your agent on your content
Five entries are enough to start, and the Free plan costs nothing. Paste what you answer every week, and see what your agent does with it.