Skip to content

Snippets and files

The crawler can only index what your website publishes. Everything else — prices you never put on a page, the answer to a question that keeps coming up, a policy that lives in a PDF on someone’s desktop — goes in by hand.

Workspace → Knowledge → Text snippets.

A snippet is a title and a body. It is chunked and embedded exactly like a crawled page, so it competes with your website content on relevance rather than overriding it.

Good snippets are specific and self-contained:

Title: Refunds outside the 30-day window Body: After 30 days we do not refund, but we do credit the amount against a renewal for up to a year. Ask support to apply it; there is no form.

Bad snippets are the ones that need context you did not give:

Title: Pricing Body: See the pricing page.

Snippets do not expire and nothing re-checks them, which makes them the most likely thing in the index to go stale. Two habits help:

  • Put a date in the body when the fact has a shelf life — “as of March 2026, the trial is 14 days”. The assistant will quote it, and a visitor reading a dated fact treats it correctly.
  • When something changes on the website, search the snippets for it. The crawler updated your pages; it did not update your notes.

If visitors keep asking something, and the answer is not on the site, the snippet is a patch and the page is the fix. Content gaps exist to make that choice visible — see Content gaps. Answering it in a snippet helps the people who reach the chat; publishing the page also helps everyone who searched for it and never arrived.

Separately from all of this, visitors can attach images and documents to their own messages — up to 8 MB each. Those attachments belong to the conversation: they are read to answer that visitor’s question and shown to you on the conversation, but they never enter the agent’s index and never influence anyone else’s answers.