Train an AI agent on your website and docs

Knowledge is what your agent answers from. This page shows how to train an AI agent on your website and documents: add a source under Knowledge base → Data sources and Centillion imports it and trains the agent on it automatically. You can close the page while imports and training run.

​

What it does

• Imports six kinds of source: files, websites, text snippets, Q&A, Notion pages and closed support tickets from Zendesk or Salesforce.
• Reads files up to 20 MB each, 20 files per import, in PDF, DOC, DOCX, TXT, Markdown, CSV and JSON.
• Scans a website for up to 500 pages, and imports up to 50 selected pages per website source.
• Trains automatically after every change on every plan. Only new or changed text is embedded again.
• Checks websites, Notion pages and tickets for changes every week with Auto-resync, on Standard and up.
• Counts extracted text against a storage allowance per agent: 1 MB on Free, 10 MB on Hobby, 20 MB on Standard and 40 MB on Pro.

​

What you need

Every plan can add sources. If you are new to the product, start with what Centillion is. Beyond that you need:

• The Owner or Member role to add, edit, remove and restore sources. Viewers can read and export them. See workspace roles.
• Standard and up for weekly Auto-resync. Any plan can refresh a source by hand with Resync now.
• For tickets: a connected Zendesk or Salesforce integration that can read tickets and comments. Tickets as a knowledge source is part of Pro and up.
• For Notion: access to the Notion workspace, or a workspace administrator who can grant the integration access.

​

Set it up

​

Use it every day

​

Files

Choose Add files or drag documents into the upload area. PDF, DOC, DOCX, TXT, Markdown, CSV, and JSON are supported, up to 20 MB each and 20 files per import. Files upload one at a time. PDF text must be selectable; scanned PDFs require conversion to readable text before importing. Images are not training content. Each file has its own source and extracted-content preview. You can edit the extracted text. The original file is discarded after extraction; keep your own copy. Exporting a file produces a .txt file containing its extracted text.

​

Text snippets

Give each snippet a title, then write or paste its content. The editor supports headings, bold, italic, underline, strikethrough, ordered and bulleted lists, links, and emojis. Snippets export as Markdown (.md). To keep a short document in front of the agent on every turn, open its … menu and choose Always keep in prompt. The source shows Always in prompt, and the agent reads its full text without searching. This works for files, text snippets and Q&A sources, up to 20 sources and 50,000 characters per agent. Long prompts cost more and can slow replies, so keep this for material the agent needs on most turns. Choose Search only when needed to go back. A published agent uses the text from its release, so an edit to one of these sources reaches customers when you publish. A workflow step that doesn’t inherit the agent’s knowledge doesn’t get them, and they stay behind when a conversation moves to another agent, which searches its own sources as usual.

​

Websites

Choose Add website, then one of:

• Crawl links: discover public pages from the website.
• Sitemap: read an XML sitemap, including indexes that reference child sitemaps.
• Individual link: import specific URLs from the same website, one per line.

Choose Fetch links and select the pages to import. The site and its www. version count as the same website. Other subdomains and private or internal addresses are refused. Several website sources can import concurrently. Open a website source to see Fetched links, grouped by path. Trained shows included links, Not found shows failed pages when present, and Excluded shows pages you left out. Search and sort the tree, select an individual page or a folder, then choose Exclude selected links or Include selected links. Membership changes save immediately. Discard clears the selection without making changes. Excluded pages stop contributing to draft answers. Publish the branch to change the pages used by new customer conversations.

​

Advanced options

Saving options starts a resync. Available options are:

• Exclude paths: starts with, ends with, contains, exact match, or wildcard patterns with * placeholders. Add up to 30 patterns; Clear all removes them.
• Scraping location: choose the country used by the scraper; US is the default.
• Slow-scraping: read one page at a time to reduce load on the site.
• Include query parameters: treat pages such as ?page=1 and ?page=2 separately.

Fragments such as #pricing are part of their containing page.

​

Resync now

Choose Resync now from a website or Notion source’s menu. Website refreshes read existing pages and discover new ones. Sitemap refreshes remove pages that left the sitemap. A failed crawl page stays visible with its error for review. Manual page exclusions survive future refreshes. Retry failed pages reads only pages that failed without rediscovering the site.

​

Q&A

Each Q&A has a title, an answer, and up to 50 question variations. Answers use the same rich editor as text snippets. Matching ignores capitalization and punctuation. Add variations for other ways users ask the question. On a surface with no actions configured, an exact match can return the saved answer directly. Actions or customer attachments may require the answer to go through the model. Other questions can retrieve Q&A content through search. Bulk upload accepts .xlsx files up to 1 MB with 1 to 100 data rows plus a header. Include Questions (or Question) and Answer columns. Title is optional. Separate question variants with line breaks. Exports also include Source ID: keep it to update an existing Q&A in this agent, or leave it blank to create a new entry. Rows are validated before importing; a later storage failure reports how many rows were already saved. Open a Q&A preview for Usage insights: total matches, last asked time, added time, and daily activity. Counts include exact and search matches to that Q&A. Daily totals use UTC.

​

Notion

Choose Add Notion pages, authorize the workspace, and choose the pages or databases to import. Private pages do not need to be published. If permissions are restricted, ask a workspace administrator to grant integration access. Expand a page to choose its subpages individually. Import up to 50 pages or databases at a time. Existing pages are marked Added. Each database source supports up to 500 entries; a source supports 2,000 blocks and 20 nesting levels. Text, lists, tables, and nested blocks are read. Images and attachments are not. Edit the original content in Notion and use Resync now, or enable weekly Auto-resync. Imported content is trained automatically. Removing a source from one agent does not change the connection’s access for other agents.

​

Support tickets

Choose Add tickets, select Zendesk or Salesforce, turn on Sync tickets into your AI agent, and save. A Connected label confirms an active integration. Otherwise follow Open Integrations to connect it first. Only one platform can provide ticket knowledge for an agent at a time. All eligible closed tickets are imported and summarized; there are no date or status filters. The connection must be able to read tickets and comments. Individual tickets cannot be previewed, edited, or deleted in Data sources. Switching platforms replaces the ticket source. Turning syncing off removes its knowledge immediately. Auto-resync imports newly closed tickets each week.

​

Sharing knowledge across agents

To reuse knowledge across agents, create it in the workspace Knowledge base, then attach it under an agent’s Knowledge base → Shared library. You can also import a copy of an existing source. Automatic source changes apply to the selected branch draft. Published versions retain their saved knowledge; publish another version to update new customer conversations. Removing a source from a draft leaves published versions unchanged; removing it from the workspace stops retrieval in every version. See releases and shared resources.

​

Finding sources

Search by title or URL, filter by type, sort by age or name, and choose 10, 25, or 50 sources per page. Bulk select selects sources on the current page. Changing pages or filters clears that selection.

​

Automatic training and weekly resync

Local changes train automatically for every plan. Files, snippets, and Q&A do not need an external refresh schedule. Remote websites, Notion pages, and tickets can change outside Centillion. On Standard and up, open Auto-resync in the header and flip the switch. The next check is seven days after enabling it; later checks run weekly in the background. The menu shows the next check. A plan without this feature can still use Resync now. Scheduled checks revalidate the plan entitlement, so a downgrade stops scheduled imports. Only new or changed text is embedded again. Unchanged passages reuse their embeddings. Imports resume from saved progress, training retries safely, and concurrent workers cannot claim the same job. A failed build leaves the prior training available. Source membership and page changes affect the draft; published versions keep their saved knowledge. A Sync failed alert offers Retry sync after automatic retries are exhausted. Each source shows one of these statuses:

• Syncing: the source is queued, importing, or being applied to the agent.
• Synced: the source is ready in your draft. Publish to use it in new customer conversations.
• Partly imported: some pages failed; review their errors and retry.
• Failed: import or training failed; review the reason.
• Needs import: a legacy source has not been fetched; use Resync now.
• Removed: the source is retained for restoration but is no longer used.

​

Removing and exporting sources

In an agent, Remove detaches the source from the selected branch draft and keeps it available to restore. Published versions keep using their saved copy. Restore reattaches it and automatically syncs its knowledge again. In the workspace Knowledge base, Remove stops retrieval of that source in every version. Permanently delete is available there after removal and deletes the stored content and training records. Bulk selection supports these operations across the current page. Exports by source type:

• File: extracted text (.txt).
• Text snippet: Markdown (.md).
• Q&A: all selected Q&As in one Excel workbook (.xlsx).
• Website, Notion, tickets: not exportable.

Multiple output files are bundled into .zip. Select up to 50 sources at a time; exports cannot exceed 200 MB. With the Q&A filter active, Export all Q&As includes up to 5,000 entries matching your search.

​

Limits and good to know

• Storage allowances are Free 1 MB, Hobby 10 MB, Standard 20 MB, and Pro 40 MB per agent. Enterprise allowances are configured separately.
• A source supports 10 MB of extracted text; a snippet or Q&A supports 1 MB. Source titles support 250 characters.
• A website scan can return up to 500 pages, and one source imports up to 50 selected pages. A larger site needs multiple sources. Sitemap indexes support up to 100 child sitemap files.
• Resync reports a limit error instead of silently truncating an expanded source.
• Images inside files and Notion pages are not read, and scanned PDFs need converting to text first.
• Shopify actions and Stripe actions retrieve live customer data separately from shared knowledge.
• For questions that need a person, configure Escalations or Centillion live chat under Tools, with Helpdesk enabled and the customer’s channel selected, so a person takes over from the helpdesk inbox.

​

Questions

​

Do I need to retrain the agent after adding a source?

No. Adding, editing, removing or restoring a source starts training on its own. You can close the page while it runs, and the source shows Synced when it is ready. Publish the branch to use its new knowledge in customer conversations.

​

How often does Centillion check my website for changes?

With Auto-resync on, every seven days, starting seven days after you turn it on. It is available on Standard and up. On any plan you can choose Resync now from the source’s menu.

​

Why was part of my website refused?

Within one website source, the site and its www. version count as the same website. Pages on other subdomains and private or internal addresses are refused. Fragments such as #pricing are part of their containing page.

​

What happens to the file I uploaded?

Centillion extracts its text and discards the original, so keep your own copy. You can edit the extracted text, and exporting the source gives you that text as a .txt file.

​

Can several agents use the same knowledge?

Yes. Create the source in the workspace Knowledge base and attach it under each agent’s Knowledge base → Shared library, or import a copy of an existing source.

​

Does new knowledge reach a published release?

Automatic source changes apply to the selected branch draft. Published versions keep the knowledge they were saved with, so publish another version to use the new content in customer conversations.

​

What counts toward my storage allowance?

The text Centillion extracts from your sources, measured per agent. The size of the files you upload does not count. The counter on Data sources shows how much of the allowance you use.

​

Try it

Add your website and a few documents, wait for Synced, and ask the agent the questions your customers ask most. Or ask the assistant on this page.

​

Related