Key takeaways
- Add knowledge by crawling your site and help centre or by uploading files.
- Answers come from your passages, with the source kept for review.
- When the knowledge has no answer, the agent offers a follow-up instead of guessing.
- Your documents, the prospect's words and screen text are treated as data, never as instructions.
What counts as knowledge
- Your public website and pricing page, crawled from the address you give.
- Help centre articles, API docs and release notes.
- Uploaded files such as FAQs, security questionnaires and one-pagers. Files are scanned for malware before indexing.
- Notes you write for the agent: positioning, objections and things it must not promise.
How answers from your docs are kept honest
Each answer is built from the passages the search returns, and those passages are kept with the session so you can see where an answer came from. The agent is told to stay inside them. When a prospect asks about something your knowledge does not cover, such as a custom contract term, it says it will check with your team and records the question for follow-up.
Knowledge passages, web pages and the prospect's speech are fenced as untrusted data, so text inside a document cannot instruct the agent to do something else.
What is on the roadmap
A knowledge graph that links screens, flows and docs is on the roadmap, along with tracking product changes so answers stay current after you ship.