Sources & training
An agent answers from what you give it, and nothing else. This page covers the source types, their limits, and how to fix a wrong answer.
Source types
| Type | Added from | Limits |
|---|---|---|
| Website | Onboarding | Up to 50 pages per crawl. Main content only — navigation and footers are stripped. |
| File | Sources tab | PDF, DOCX, TXT, and Markdown. 20 MB per file. A file with almost no extractable text is rejected rather than silently embedded. |
| Text | Sources tab | A title and 20 to 200,000 characters. Good for policies that live in someone's head. |
| Q&A | Sources tab | 1 to 200 pairs at a time. Questions up to 500 characters, answers up to 8,000. |
| Notion | API only | Up to 50 pages per import, 25 by default. Citations link back to the Notion page. |
Everything is chunked and embedded into an index scoped to your workspace. Your content is used to answer your visitors' questions and nothing else — it is not used to train models.
Keeping a website source current
Retraining re-crawls a website source and compares each page against the copy it already holds. Only pages whose content actually changed are re-embedded; unchanged pages cost nothing. Pages that have vanished are removed.
Turn on auto-sync and this runs daily without you. One caveat worth knowing: when a crawl hits the 50-page cap, removals are skipped for that run. At the cap the set of pages found can vary between crawls, and pruning on absence would delete and re-add the same pages forever.
Fixing a wrong answer
Analytics ranks the questions that hit the fallback most often. Work down that list:
- The answer is not in your content. Add it, usually as a Q&A pair or a text snippet. This is most of them.
- It is in your content, but retrieval missed it. The retrieval inspector in the Playground shows what was retrieved and what it scored. Content that says the same thing in the visitor's words, rather than yours, usually fixes this.
- It was retrieved but answered badly. Use the correction flow on the conversation: it turns that exact question and a better answer into a durable Q&A pair.
Grounded answers
Each agent has a confidence threshold calibrated against its own content. Below it, the agent says it does not know instead of guessing, and can collect an email instead. Those fallback messages cost no credits — you are not charged for the agent declining to answer.