Knowledge base
Upload your team's internal docs (PDF, DOCX, MD, TXT) so the AI can answer questions about them. Useful for SOPs, refund policies, onboarding playbooks, compliance frameworks, product specs — anything you wouldn't want to (or can't) publish on your marketing site.
Upload a document
Open Dashboard → Settings → Knowledge. Drag a file onto the upload zone, or click browse to pick one.
- Accepted types: PDF, DOCX, TXT, MD.
- Max size: 10 MB per file.
- Files are stored privately. Visitors never see them or get a link to them.
Once uploaded, the document goes through a brief processing step (parse → chunk → embed). The status pill on the row shows where it is:
processing— being parsed and embedded right now.ready— chunks are indexed; the AI can search them.error— something went wrong (corrupted PDF, scanned image with no text, etc.). The error message is shown inline. Click Retry to re-process.
Turn on AI knowledge search
Uploading a doc doesn't automatically make the AI use it. Flip the switch in Settings → AI Agent:
- Find Let AI search the knowledge base in the Tools section.
- Toggle on, save.
From the next message onward, the AI runs a knowledge-base search before every reply and uses the matching chunks to ground its answer.
How the AI uses it
On every visitor message (when the toggle is on), we:
- Embed the visitor's question into a vector.
- Find the top 6 most-similar chunks across all your uploaded docs.
- Pass those chunks to the model as context.
- The model answers from that context, falling back to web-search on your public site (if also enabled) for anything not in the KB.
The model is instructed to answer naturally, like a knowledgeable teammate would — visitors never see filenames or chunk numbers in the reply.
Update or delete a doc
- Update — delete the old version and upload a new one. (Re-uploading the same filename creates a separate row; delete the obsolete one to keep things tidy.)
- Download — get a short-lived signed URL to fetch your own copy.
- Delete — removes the file from storage and wipes every vector chunk associated with it. The AI stops using that content immediately.
Practical tips
- Smaller, focused docs work better than one giant manual.A 30-page “everything” PDF still works, but several smaller docs (refund-policy.pdf, shipping.md, returns-faq.docx) make retrieval more precise — the relevant chunks have less noise around them.
- Use clear headings inside the doc. Our chunker is paragraph-aware; well-structured docs split into clean, self-contained chunks. Wall-of-text PDFs still work but retrieval quality drops.
- Scanned-image PDFs won't parse— we extract text, not OCR images. Run the PDF through your scanner's OCR feature first, or export the source doc as text.
- Sensitive data:only upload what you'd be comfortable showing a customer through a chat reply. The AI won't paste raw chunks verbatim, but if a doc contains a secret, treat that secret as exposed to your visitors.
Scope
Knowledge documents are per-site, like every other setting. If you operate three sites with different products, each maintains its own knowledge base.