Training your chatbot — supported sources
Add knowledge from websites, PDFs, YouTube videos, or plain text.
Updated 2 Mar 2026
How training works
Pivra splits your content into chunks (~1,500 characters with overlap) and creates vector embeddings. When a visitor sends a message, the most relevant chunks are retrieved and passed to the AI as context.
Supported source types
Website URL
Pivra crawls the URL and up to 50 sub-pages. It respects robots.txt and extracts visible page text (ignores navigation, footers, scripts).
Tip: Add your homepage — the crawler discovers linked pages automatically.
PDF / DOCX
Upload a document up to ~50,000 characters. PDFs are parsed with layout preservation; DOCX files use Microsoft Word formatting.
YouTube video
Paste a YouTube video URL. Pivra downloads the transcript (if captions are available) and trains on it.
Plain text / FAQ
Paste text directly — ideal for FAQs, policies, pricing tables, and product descriptions.
Source status
| Status | Meaning |
|---|---|
| Training | Processing in progress |
| Trained | Ready to answer questions |
| Failed | An error occurred — hover for detail |
Character limits by plan
| Plan | Limit |
|---|---|
| Free | 500K chars |
| Starter | 2M chars |
| Pro | 10M chars |
| Business | 30M chars |
Tips for better responses
- Be specific: narrow topic sources outperform a single giant document
- Use headings: the chunker uses headings to keep related content together
- Re-train after updates: edit a source and click Re-train to refresh embeddings
- Test after training: use the Preview chat to verify the bot finds the right answers
Was this helpful?
If you have questions or suggestions, email us at support@pivra.ai .