A AtomicX Chat
Help Sign in

Updated October 7, 2026

Add documents and websites

Overview

Once you've created a knowledge base, you fill it with content the AI can use to answer questions. You can upload documents such as manuals, price lists and policies, and you can add websites so pages are scanned and added automatically.

All content is managed on the Content Management page of a knowledge base. To get there, go to AI → Knowledge bases and click Manage Content on the knowledge base card. The page has three tabs: Documents, Websites and Email Sources. This article covers the first two.

Steps

Upload documents

  1. On the Documents tab, click Upload Document. You can also use Add Document at the top of the page.
  2. Under Select Files, choose one or more files from your computer (hold Ctrl or Cmd, or Shift, to pick several). You can upload up to 20 files at once. Supported formats are TXT, Markdown, PDF, DOCX and HTML, up to 10 MB per file. Files over 10 MB are marked in red and skipped.
  3. Optionally, tick Available for download. The setting applies to every file in this upload. When this is on, visitors who get an answer based on this document see a download link below the AI's reply.
  4. Click Upload. If some files fail, you get a list of them and they stay selected so you can try again. The other files are still added.

Uploaded documents count toward your plan's storage. A file that would go over it is listed with "This upload would go over your plan's storage" and how much you've used, and isn't added. Delete documents you no longer need or upgrade your plan.

The document is processed in the background. Its card shows one of these statuses:

Status Meaning
Processing The document is being read and prepared.
Processed The document is ready and the AI can use it.
Failed Something went wrong. The reason is shown on the card.

The page refreshes on its own while content is processing, so you don't need to reload it.

Manage uploaded documents

Each document card shows its type, size, upload date and, once ready, the processed date.

  • View opens a read-only window showing the file name, its status and the text that was extracted from it. Use this to check what the AI can actually read. If processing hasn't finished, the window says no extracted text is available yet.
  • Retry processes the document again. Use it if a document shows Failed.
  • Available for download can be switched on or off at any time from the card.
  • Delete removes the document from the knowledge base after you confirm.

The top of the tab shows totals for Documents, Processed and Total Size.

Learned answers from the mobile app

When someone on your team hands a conversation back in the mobile app with Teach the agent this answer ticked, a short text document named **Learned answer – ** followed by the question is added to the agent's knowledge base. It holds the question, your team member's answer, and who taught it and when. Personal details are left out.

Learned answers are added without review. Use View to check one, and Delete to remove any you don't want the AI to use. To change an answer, delete it and upload a corrected document. See Mobile app.

Add a website

  1. On the Websites tab, click Add Website.
  2. Enter the Website URL, for example https://example.com.
  3. Choose a Maximum Depth. 1 level (current page only) scans just that page. Higher values also follow links to scan more pages of the site. The depths you can choose depend on your plan.
  4. Tick Follow external links only if you also want pages on other websites to be scanned.
  5. Choose Rescan frequency: Manual only (the default), Daily, Weekly or Monthly.
  6. Click Add Website.

Each website counts toward your plan's number of websites, across all your knowledge bases. When you've reached it, you'll see "Your plan includes N websites. Upgrade your plan to add more." and the website isn't added.

The site is scanned in the background. Its card shows one of these statuses:

Status Meaning
Queued A scan is waiting for the background worker.
Scanning The background worker is scanning the website. Once pages come in, it shows how many have been fetched so far, for example Scanning · 120 pages fetched.
Processing The initial scan is still being processed. It also shows how many pages have been fetched so far.
Processed All scanned pages were added successfully.
Partial Some pages were added and some failed, for example 18/20.
Failed The scan could not be completed. The reason is shown on the card.

Review and rescan a website

Each website card shows Pages Scanned, how many succeeded, Max Depth, whether External Links were followed, and the Last Scan date.

  • Click Pages to see every page that was scanned, with failed pages first. Search by title or URL. Click a page to see what was taken from it: its status, whether it was recognised as a product page, the Price, Image and Description found, and the Extracted text the AI can use. This works like View for documents. If a product card in the chat shows no photo, check the page's Image here.
  • If some pages failed, click the red failed count to open Failed Page Scans. It lists each page with its error. The card may also show Top failure reasons.
  • Click Rescan to queue a scan in the background. The card shows Queued, then Scanning. You can close the page; the scan continues without an open browser. While open, the card refreshes automatically, including for scheduled scans.
  • Use Rescan frequency on the website card to change or turn off its schedule. Daily means 24 hours, Weekly means seven days, and Monthly means one calendar month. The first scan is scheduled from when you save the setting; later scans are scheduled from the previous scan's completion. Manual only turns off future automatic scans. An already queued or running scan can still finish.
  • Next scan shows the due time in UTC. Missed runs are combined into one scan. Schedules apply to active knowledge bases and continue at the chosen interval after a failure.
  • During a scan, the status shows how many pages have been fetched so far. Pages Scanned and Last Scan still show the previous results until the scan finishes. Large sites can take a while: pages are only added to the knowledge base once every page has been fetched. When it finishes, the card shows how many pages were added, updated, unchanged, removed and failed. Updated includes changes to product photos, prices and descriptions, even when the main page text stays the same.
  • If our service restarts during a scan (for example during an update), the scan continues where it stopped; pages already fetched are not fetched again. A scan that was stopped for more than a day starts over.
  • Scans follow your plan's page and depth limits. Previously indexed pages are revisited in bounded batches across runs. A confirmed 404 or 410 response removes a page's indexed content. Timeouts, bot challenges and pages not reached within the scan limit keep their last usable content; a sold item must actually be removed or updated on your site to be reflected.
  • If your site blocks our scanner (for example a firewall or a bot check), each blocked page is tried once more through an external page-fetching service before it counts as failed. Pages fetched this way use AI credits, shown as Website fetching on AI credits. If pages still fail with a timeout or 403, ask your web host to allow our scanner.
  • If every page of a website could only be reached through that service, its Rescan frequency is set to Manual only and locked, so scheduled scans don't keep using credits. The card says the pages come from a third-party service and that every scan deducts AI credits. Click Rescan to update it yourself.
  • If it's your own website, ask your web host to allow our bot (AtomicXChatBot) and our server's IP address, shown on the website card. Once a scan reaches your site directly again, it's free and you can turn the schedule back on.
  • Click Delete to remove the website and its pages from the knowledge base.

When a visitor asks about one of your products or services, or asks for a link or where to read more (for example "Where can I read about it?"), the chat shows a link to the most relevant page of your website below the AI's reply. The link appears as a card with the page's image, title and short description, taken from the page's Open Graph tags (og:title, og:description, og:image, or twitter:image). If a page has no Open Graph tags, the card uses the page title and description instead.

  • Extra pages are shown only when they are clearly relevant to the question, and at most three links appear. When there is more than one, the cards form a carousel that visitors can swipe or scroll with the arrow buttons. Pages without a preview and documents appear as small link buttons.
  • Greetings and general questions usually get no link.
  • Links are chosen for the visitor's latest question only. Earlier messages count only when the visitor asks for more information or clarification, such as "Tell me more" or "What do you mean?". A "Thank you" at the end of a conversation gets no link.
  • Only pages from websites added to this knowledge base are linked. The AI never writes links into its replies itself, and never mentions "the link below", because it can't know whether a link will be shown.
  • To turn page links off for a widget, clear Show Page Links in Answers under Additional Settings in the widget's configuration. Downloadable documents still show their download link. See Configure your chat widget.
  • To make your cards look good, give each product and service page an og:image and og:description. Changes to these tags can take up to a day to show in the chat.
  • If your site uses bot protection (a "checking your browser" or "establishing a secure connection" page), the card uses the image and title saved during the last scan. Rescan the website after you add or change page images.

Test your indexed content

After processing or rescanning, test questions whose answers depend on the updated content. Include a question requiring information from more than one page and questions in other languages. Verify the facts and any page links separately: the passages used for an answer and the navigation links shown beneath it serve different purposes.

If content is waiting for a new indexing model, wait for processing to finish before testing it. An unchanged page may still need fresh indexing after a model change.

Tips

  • Clean, well-structured documents with clear headings give the best answers.
  • Start with a low Maximum Depth and increase it only if important pages are missing. Deep scans take longer and can pull in pages you don't need.
  • Leave Follow external links off unless you really need content from other sites.
  • If a document keeps failing after Retry, read the error on its card. If you can't resolve it, contact support.
  • The number of pages a single scan can cover depends on your plan. See Plans and subscription.
An unhandled error has occurred. Reload ×