Knowledge
Train the AI on the business's real content: deep website crawls, files, text, and Q&A pairs. Retrain any time; answers are grounded in what you loaded, not guessed.
View as MarkdownKnowledge is what makes the AI answer like a local instead of a chatbot: the menu, the price list, the service area, the insurance list, the return policy. It's trained once at the business level and shared by the agents, so every door quotes the same facts.
Sources
| Source | How it works |
|---|---|
| Website (deep crawl) | Paste the URL. OpenAgent reads the sitemap and robots rules, then breadth-first crawls linked pages, and reports exactly which pages were found and trained — with a per-page list you can prune. |
| Files | PDFs and documents — price sheets, service menus, policy docs. |
| Text | Paste anything: "wait times tonight are 25 minutes", the things that change nightly. |
| Q&A pairs | The exact answer to an exact question — strongest signal, use for the questions customers ask daily. |
Training and retraining
- 1Add sourcesBusiness → Knowledge → add a URL, file, text, or Q&A pair. A crawl shows a scan preview before it trains.
- 2TrainContent is chunked and embedded into the retrieval index. The page reports pages found vs. pages trained.
- 3TestAsk the chat preview real customer questions. Wrong answer? Fix the source — add a Q&A pair for anything the site says ambiguously.
- 4Retrain when things changeEdit or re-crawl, then Retrain. Retraining re-embeds the agent's sources so the answers move with the business.
When an answer is wrong, the fix is almost never the persona prompt — it's a missing or stale source. Q&A pairs are the scalpel: one pair per recurring question.
Knowledge in flows
A workflow can pin reference material to a single step — the menu on the order step, the shut-off instructions on the emergency step. Step-pinned knowledge is only in play on that step, which keeps long documents from bleeding into unrelated answers.