The seven knowledge source types
What each one takes, and which to use when.
| Type | Takes | Best for |
|---|---|---|
| Upload (documents) | PDF, DOCX, TXT, 10MB per file | Books, workbooks, guides you already have |
| Website | A URL to crawl | Services, positioning, pricing pages |
| Text | Pasted text | Anything you write directly |
| Q&A | A question and your answer | The questions you actually get asked |
| Your profile | Background and credentials, quickly | |
| Video | A video file to transcribe | Talks, recorded calls, workshops |
| YouTube | A YouTube URL | Anything already published there |
Which to reach for
Q&A is the highest value per minute spent. It is the only type where you choose both the question and the answer, so it maps exactly onto what people ask.
Video and YouTube are the only types that capture how you speak rather than how you write. If your clone sounds stiff, this is the fix.
Website is the fastest to add and the weakest per character. Marketing pages are written to persuade, not to explain, so a crawled site often teaches your clone to pitch.
Two different size limits
This catches people out, because the app does not spell it out anywhere.
| What you are adding | Limit |
|---|---|
| Documents: PDF, DOCX, TXT | 10MB per file |
| Video and audio | 500MB per file |
So a 40MB PDF is refused while a 400MB recording of the same material is fine. If a document is too big, split it by chapter. That also makes retrieval more accurate, so it is worth doing even when you are under the limit.