Help sources
Give a chatbot your help website or a Markdown help file, and keep what it knows up to date.
A help source is where a chatbot learns what to say. It answers only from what it read there. A chatbot can't answer anyone until at least one source has been read.
You can add up to 10 sources per chatbot. Each one is either:
- A help website: the address of your help centre's home page, for example
https://help.example.com. - A help file: the address of one public Markdown or text file, for example
https://example.com/llms-full.txt.
Add a source#
- 1
Open the chatbot. It opens on the Help sources tab.
- 2
Click Add help website. A pop-up opens.
- 3
Paste the address in Address. A help file's address works here too.
- 4
Click Add help website. Reading starts straight away.
- 5
Wait for the status to change to Ready. A website can take a few minutes. The page updates by itself.
Every source is a line in a table:
| Column | What it shows |
|---|---|
| Help website | The address, which opens the site. If reading failed, the reason shows under it. |
| Status | Where reading is, in a coloured label. See below. |
| What it read | Once ready, how many pages and passages it found, and when it was last read. |
What the statuses mean#
| Status | Meaning |
|---|---|
| Waiting to be read | Queued. Reading starts shortly. |
| Reading the site… | Pages are being read now. |
| Ready | Done. What it read shows how many pages and passages it found. |
| Could not be read | Something went wrong. The reason shows under the address. If it read the source before, it says It still answers from what it read last time. |
A passage is a short piece of a page, about a paragraph or two. The chatbot finds the passages that best match each question and answers from those.
Help websites#
From the address you give, every page on the same site, under the same folder, is read. For example, from https://example.com/help/ it reads https://example.com/help/billing but not https://example.com/blog/.
End a folder address with a slash
https://example.com/help/ reads the help folder. https://example.com/help (no slash at the end) is treated as a page in the site's top folder, so the whole site is read. If your help centre has its own address, like https://help.example.com, this doesn't matter.
How pages are found:
- By following links from page to page.
- From your sitemap, if you have one. It looks for the sitemap listed in your
robots.txt, or at/sitemap.xml.
What's read, and what's skipped:
| Read | Skipped |
|---|---|
| Normal web pages (HTML) | PDFs, images, videos, spreadsheets and other downloads |
The page's main content, when the page marks it (a <main> or <article> area) | Menus, headers, footers, sidebars, forms and buttons |
| Up to 400 pages per source | Pages larger than 2 MB |
Pages your robots.txt blocks, and pages marked "noindex" | |
| The same page found under two addresses (read once) |
Sites built in the browser show no text
Some help sites are built by JavaScript after the page loads (single-page apps). The page we download is empty, so nothing can be read. If your help centre works this way, use a help file instead.
Help files#
A help file is one public Markdown (.md, .markdown) or text (.txt) file. Many sites already publish one for AI tools, often called llms-full.txt.
How the file is split up:
- Each top-level heading (
# Title) starts a new page. - A line
URL: https://…straight under that heading says where the page lives on your website. The chatbot uses it to link to the page.Source:orLink:work too. ##and###headings split a page into sections.- Images are ignored.
- If the file has no
#headings, the whole file is one page.
# Changing a booking
URL: https://example.com/help/bookings/change
Open the booking and click Edit.
## Moving it to another day
Pick the new date and save.
# Cancelling a booking
URL: https://example.com/help/bookings/cancel
Open the booking and click Cancel booking.
The address must end in .md, .markdown or .txt. The file must be public and under 2 MB. If the address returns a web page instead (for example, your site's "page not found" screen), it's refused.
Keeping it up to date#
Every source is read again every night, so the chatbot keeps up with changes to your help centre.
To pick up a change now, open the source's ⋯ menu and choose Read again. Reading again replaces what the chatbot knew from that source. If reading fails, it keeps answering from the last good read.
To stop using a source, choose Remove… in the same menu and confirm with Remove. The chatbot forgets everything it read from it.
How answers stick to your sources#
For each question, the chatbot:
- Searches your passages for the words in the visitor's question, and the question before it (so follow-ups like "and for staff?" still work).
- Takes the best matches, up to eight passages.
- Answers from those passages only.
If nothing matches, it says it doesn't know and follows your instructions for what to suggest instead. It never fills gaps from general knowledge about your product.
Tips for good help content#
- Use the words people type. The search matches words, so a page that says "cancel a booking" is found for "how do I cancel". Include the words people actually type, including common alternatives ("refund", "money back").
- One topic per page or section. Clear headings help, because passages are kept within one heading.
- Put the answer in text. Text inside images, videos or PDFs can't be read.
- Write out the steps. "Open Settings, then click Billing" is more useful than "see the screenshot below".
- Include a contact or support page. The chatbot only links to addresses in your sources. If you want it to point people to support, that page needs to be there.
- Check the Conversations tab. Questions it couldn't answer show you what to write next.