Knowledge → Documents holds two things: files you upload, and notes you write yourself. Both are indexed the same way and answer questions the same way.
What file types can I upload?
Text and documents: .txt, .md, .pdf, .docx, .xlsx, .csv.
Images: .png, .jpg — these are read with OCR.
A few things worth knowing before you upload:
- Text-based PDFs and Word/Excel/CSV files are read in full. Prefer these when you have the choice.
- Scanned PDFs and images go through OCR. A very long scanned PDF may only be captured in part, so a text-based original always beats a scan.
- There's a size limit per file. If a file is over it, split the document and upload each part — the dashboard tells you the current limit and the size of the file you tried.
- We don't keep your original file. The text is extracted to train the chatbot and the file itself is discarded, so it can't be downloaded back. Keep your own copy of anything you may need later.
Upload status moves from pending → processing → ready. Once it says ready, the content is live.
When should I write a Note instead of uploading a file?
Whenever the knowledge lives in your head rather than in a document. Notes are the fastest way to fix a bad answer: write what the assistant should have said, save, done.
Click Write a Note, give it a title (the title also helps the assistant find the note by topic), and write the body in Markdown with a live preview beside it.
Use bullet lists for sets of values. Hours, service areas, plan tiers, supported models — a list keeps each item distinct when the note is chunked for indexing. A wall of prose blurs them together.
## Hours
- Mon–Fri: 9–5
- Sat: 10–4
We are at **123 Main St**.
How do I stop the assistant using a document?
Toggle it off in the list. It stays in your account, but the chatbot stops using it. Toggle it back on to make it live again. Deleting removes it entirely.
Why isn't my document being used in answers?
- It may still be processing — check the status chip.
- It may be disabled — check the toggle.
- The content may not match the question. Test the exact visitor question in Knowledge → Live Data and look at the source chunks that come back. See Testing your knowledge base.
- If it was a scan, OCR may have captured less than you think. Open View ingested content to see exactly what we extracted.