What documents should I give the agent?

Open Workspace → AI Agent → Knowledge → Documents. These are the price sheets, FAQs and policies the agent may quote from. It only claims what it finds in a document or in your facts — nothing else — so what you upload here is the ceiling on what it can tell a customer.

What you can upload

PDF (with a text layer), Markdown, CSV, JSON or plain text, up to 8 MB per file.

What makes a good document

Short and structured beats long and complete. The agent retrieves a passage, not a file, so a document with a clear heading per topic gives it something precise to find and quote.

• Good: a FAQ where each question is its own heading, one answer under it.
• Poor: a twelve-page brochure where the return policy is a sentence in the middle of a paragraph about the company's history.

Split a long document into several short ones rather than uploading one big file. Retrieval improves and you can replace one topic without re-uploading everything.

Reading the status

Each document shows what happened to it:

• Processing… — being indexed. Give it a moment.
• Ready · N chunks — indexed and usable. The chunk count is roughly how many separately retrievable passages it produced.
• Failed — the file could not be read at all. Re-upload in another format.
• Indexed but empty — the file was read but no text came out.

"Indexed but empty" — the scanned-PDF trap

This almost always means a PDF without a text layer: a scan or a photo of a page, where the words are pixels rather than characters. The agent cannot quote it, and because it indexed without error it can look fine at a glance.

The fix is to re-export the PDF from whatever produced it rather than scanning a print, or to paste the content into a Markdown or text file.

Removing a document

Delete document removes it and its passages. The agent stops being able to quote it from that moment. Delete an out-of-date price sheet rather than leaving it beside the new one — the agent has no way to know which of two contradicting documents you meant.

The agent quotes something I have deleted

Check the Learning tab. An answer a colleague wrote that repeats the old information may have been approved for reuse, and it survives the document being deleted. See How does the agent learn from my team's replies?.