AI Training Hub: teaching Fynn

Before you start

• The Training Hub is part of the Startup, Business and Enterprise plans, because it trains your AI bot.
• Create a bot first under Settings > Bots. Without a bot, the Training Hub shows Create your AI bot first.
• In a new workspace the guided setup creates Fynn and 2 sources for you: a Website Crawl named after your website and an FAQ List called Setup answers, which holds the answers you kept or corrected in the setup (up to 20). The website's first read is quick, up to 25 pages within 90 seconds with the pages customers ask about first, and then the full crawl follows as described below.
• Administrators can add and delete sources, and on Enterprise also members whose custom role has Manage Training Hub sources.

The 4 source types

• Website Crawl reads up to 500 pages of a website, starting at the URL you enter.
• Help Center imports help center articles. Pick one of your Deskwoot help centers under Your Deskwoot help center, or choose Enter an external URL for a help center somewhere else. It is crawled like a website.
• Document Upload takes a PDF, Word (.docx), CSV or TXT file of up to 10 MB.
• FAQ List holds questions and answers you type in. Write each pair as Q: question and A: answer, with a blank line between pairs.

Add a source

1. Open Settings > Training Hub and click Add Source.
2. Pick one of the 4 types.
3. Give the source a Name, for example "Main website".
4. Under Assign to Bot, pick one bot, or keep All bots (shared) so every bot can use it.
5. Enter the URL, upload the file or type your FAQ list.
6. Click Add Source. Deskwoot checks that the content can be read, then creates the source.

Documents and FAQ lists are processed right after you add them and then show Ready. A website or help center source shows Processing and "N pages read so far" until its first crawl is done.

How the crawl works

The crawler starts at your URL and also reads the sitemap: the one your robots.txt names, or /sitemap.xml. Then it follows links on the same site up to 3 levels deep. Pages in the language of your start URL come first, other languages after them. It keeps out of paths your robots.txt excludes, pauses between pages as robots.txt asks, and never fetches internal or private addresses. One source holds up to 500 pages.

The crawl runs in the background. It starts as soon as you add the source, reads for up to 4 minutes, and then goes on about every 10 minutes until it is done. A large site can take a while. A crawl that is still not finished after 24 hours is stopped.

Automatic re-crawl every 48 hours

Website and help center sources are crawled again every 48 hours, so Fynn answers from the current version of your pages. The row of each source shows Last crawled and when the next crawl is due, for example Next crawl in 12 h. During a re-crawl it shows Re-crawling... and the pages read so far. Fynn keeps the old content until the new crawl is complete, then the new pages replace the old ones in one step. If nothing changed, nothing is replaced.

Need the update sooner, for example right after you published a new article? Click Re-crawl next to the source.

If a re-crawl fails, for example because the site was unreachable, blocked the crawler or returned far less content than last time, Fynn keeps the previous content and the automatic re-crawl pauses for that source. The row shows the error and Automatic re-crawl paused after an error. Fix the cause, then click Re-crawl. A successful crawl starts the 48 hour cycle again.

Everyone who may manage Training Hub sources also sees a notice at the top of the dashboard. It names the source and the reason, says from which date Fynn still answers, and links to the source, whose row in the Training Hub says the same. If more than one source is affected, it gives their number and links to the Training Hub. Closed, the notice stays away in that browser tab and comes back in a new one while the re-crawl still fails. It goes away by itself after a successful re-crawl or once you delete the source. Did you move that help center into Deskwoot, for example with the Zendesk import? If your bot has Use Help Center articles as knowledge base switched on, Fynn reads the published articles there, and you can delete the old source.

A source whose first crawl or processing failed shows Failed and a Retry button. Documents and FAQ lists are not crawled again. To change one, delete it and add the new version.

Tips for better answers

• Keep content current. Out of date material leads to wrong answers.
• Be specific. Short pieces about one topic each work better than one long document about everything.
• Avoid duplicates. When two sources say different things, Fynn can give different answers. Remove the old one.
• Test after a change. Ask Fynn a question about the new content in a test chat.

Frequently asked questions

Which files can I upload?

PDF, Word (.docx), CSV and TXT, up to 10 MB per file. A PDF needs selectable text: a scanned PDF without a text layer cannot be read. Other formats, such as images, Excel files or Markdown, are not accepted, so save them as one of the 4 formats first or paste the text into an FAQ list.

How many sources can I add?

There is no fixed limit on the number of sources. Each website or help center source reads up to 500 pages. Quality matters more than quantity: sources that contradict each other make answers worse.

Does Fynn read my help center without a Training Hub source?

It can. In the bot settings under Settings > Bots, switch on Use Help Center articles as knowledge base. Fynn then searches the published articles of your live help centers directly. A Help Center source in the Training Hub is the other way: it is crawled and refreshed every 48 hours.

Why does my source say Automatic re-crawl paused after an error?

The last crawl of that website or help center failed, so the automatic re-crawl stopped and Fynn still uses the previous content. Check that the URL works, that the site is reachable and that it does not block DeskwootBot. Then click Re-crawl.

What does Fynn do when my sources do not answer a question?

Fynn is told to answer only from your content and never to make things up. When your content does not answer the question, it says so. In an inbox whose Bot Mode is Support Bot, it also offers to bring in a teammate. An Info Bot does not hand over.

Are my whole files sent to the AI model?

No. Deskwoot splits every source into small pieces, and Fynn sends the AI model only pieces, never whole files.

Does the Training Hub cost extra?

No. It is part of the Startup, Business and Enterprise plans. Fynn's replies count toward your plan's AI conversations, see AI conversations and billing.

What happens to the automatic re-crawl after a downgrade to Hacker?

It stops, because Hacker has no AI bot. Your sources are kept. After you upgrade again, click Re-crawl on a website or help center source to start its 48 hour cycle again.