Build the agent

Improve answers from past chats

Turn any conversation in your transcript log into a canonical Q&A pair the agent will use the next time someone asks the same thing.

When to use it

Use this whenever the agent's reply was close-but-not-right, or when you want to nail down the exact wording for a sensitive answer (refunds, pricing edge cases, anything legal-adjacent). Improvements join your knowledge base immediately — no re-crawl needed.

How to do it

1. Open Bot → Conversations and click any chat.
2. On any assistant reply, hit Improve answer. An inline editor opens with the original question and the agent's reply pre-filled.
3. Edit either field. Keep the question phrased the way a real visitor would ask it; the agent will match against the wording.
4. Save training pair. We embed the question and add the Q&A pair to this bot's knowledge index right away.

How the agent uses your pairs

Saved pairs are retrieved through the same vector search as crawled pages. The retrieval payload marks them as source: trained_answer so the model knows to prefer them when both a trained pair and a crawled passage match the same question.

In the dashboard transcript view, replies that pulled from a trained pair show a green Trained answer chip instead of a URL citation.

Editing or removing a pair

Improvements are scoped to the assistant message you opened the editor on, so re-opening the same message lets you adjust the pair or remove it entirely. Removed pairs leave the knowledge base immediately.

Quotas

Trained pairs count against your workspace's training pages quota — see Billing & quotas. A single pair is one row, so it's the cheapest way to add knowledge.