Knowledge sources

Websites
Enter a home page or a sitemap. We read the sitemap first, then follow links on the same site, respect robots.txt and pages marked noindex, skip duplicates, and stop at the page count you set (up to your plan's limit). Use include and skip paths like /help/* to narrow it.

Files
PDF, Word (.docx), TXT, Markdown, HTML and CSV, up to 8 MB each. Scanned PDFs without selectable text can't be read yet. A CSV with question and answer columns becomes Q&A pairs.

Text and Q&A
Paste policies or write question-and-answer pairs. Q&A pairs are matched one to one, so they're the most reliable way to teach exact answers.

Notion and Google
Publish the Notion page to the web, or share the Google Doc, Sheet or Slides as "Anyone with the link can view", then paste the link.

For a private Notion workspace, create an internal integration at notion.so/profile/integrations (read content only), add it to the pages you want under ··· → Connections, and paste its secret. For private Google files, choose Drive folder, connect Google Drive (read-only) and paste a folder link: we read Docs, Sheets, Slides, PDFs and Word files in it and one level of subfolders.

YouTube
Paste video links, one per line. We read each video's captions, English first. Videos without captions are skipped with a note.

Zendesk, Freshdesk and Confluence
Enter your help-centre address. Zendesk help centres that are public need no token; private ones need an agent email and API token. Freshdesk needs an API key. Confluence needs your Atlassian email, an API token and the space key. Only published articles and current pages are read.

Past tickets
Connect Zendesk or Freshdesk to learn from up to 200 recently solved tickets. We keep only the customer's question and your team's public replies, and remove emails, phone numbers, card, Aadhaar and PAN numbers before anything is stored.

Keeping it fresh
Websites, Notion, Google, help-centre and ticket sources can re-read daily or weekly. You can also press Re-learn anytime.

Limits
Each agent has a character budget (100 million characters during launch). The sources page shows what you've used.