Backoffice

Train a bot on your content

Add URLs, files, plain text, and question and answer pairs as sources.

Ingest data sources

Train each bot on one or more data sources. You can mix URLs, files, text, and Q&A.

URL sources

Use this to train on existing web pages.

Workflow:

  1. Go to Data Sources > URL.
  2. Click Add URLs.
  3. Paste one or more site addresses, one per line.
  4. Click Read pages. Each address is read along with the pages it links to, and the list shows what was found.
  5. Review the list and remove any pages you do not want.
  6. Click Train.

How far an address reaches:

  • An address is read together with the pages it links to directly, on the same site. Pages those pages link to are not read. If your site has a section a visitor reaches in two clicks, such as /docs/getting-started behind /docs, add the deeper page as an address of its own, or add the section's index page so its links are the ones followed.
  • Links to other sites are never followed.
  • One read stops after 25 pages. For a larger site, add its sections as separate addresses and read them one at a time.
  • Only the pages you leave in the list are trained. A page that was not read is not something the bot can answer from, however clearly the rest of your site points to it.

Training an address again:

  • An address you have already trained is the same source, not a second one. If the page has not changed, the run reports it as Already up to date and nothing is charged. If it has changed, the source is replaced with what the page says now and the run reports it as Updated.
  • To pick up a changed page without going through Add URLs, use Refetch and retrain on the source in the list. It does the same thing.
  • A page that can no longer be reached leaves the source as it was and says why. Nothing the bot knows is lost by trying.

Notes:

  • A page whose content is identical to a source already trained is skipped, whatever its address.

File sources

Use this to train on local files.

Workflow:

  1. Go to Data Sources > Files.
  2. Click Add more files.
  3. Drag and drop files into the upload zone.
  4. Click Train bot.

Supported file types:

  • TXT
  • PDF

Uploading a file again:

  • A file with the same name as one the bot already holds is treated as a new version of that file. If its content is unchanged, the run reports it as Already up to date and nothing is charged. If the content differs, the existing source is replaced with the new file and the run reports it as Updated.
  • This means two different documents that share a filename cannot both be trained: the second replaces the first. Rename one of them before uploading if you need both.
  • A file whose content is identical to a source already trained is skipped, whatever its name.

Notes:

  • Files with unsupported formats will show warnings.

Text sources

Use this for plain text content.

Workflow:

  1. Go to Data Sources > Text.
  2. Click Add more text.
  3. Enter a title and paste content.
  4. Click Train bot.

Editing:

  • You can edit a text source and save changes. The bot is trained on the new text as you save; there is nothing further to do.
  • If training could not run, for example because the Workspace's AI allowance is used up, the save still lands and the source shows Needs training in the list with a Retry training action.

Limits:

  • Max 10,000 characters per text source.

Q&A sources

Use this to train on structured questions and answers.

Workflow:

  1. Go to Data Sources > Q&A.
  2. Click Add more Q&A.
  3. Enter a title and add pairs.
  4. Click Train bot.

Editing:

  • You can edit Q&A sources and save changes. The bot is trained on the new pairs as you save.

Limits:

  • Max 50 pairs per source.
  • Max 200 characters per question.
  • Max 1,000 characters per answer.

Managing sources

On each data source tab you can:

  • Refresh the list.
  • Delete sources to remove learned data.
  • Refetch and retrain a URL source, to pick up a page that has changed.
  • Retry training on a source marked Needs training, whose last training did not finish.

Training tips:

  • Keep content relevant and focused.
  • For files and text, use headings and paragraphs when possible.
  • Avoid tabular or unformatted text; it can lead to inaccurate responses.

Usage

Training is measured in two ways, and both appear in your Workspace usage:

  • Training volume counts the source characters processed.
  • Tokens count the estimated embedding usage needed to index the content.

Chat consumes tokens but no training volume. Review either under Account > Usage, which reports for the whole Workspace rather than for you individually.