An email from a customer, an answer from your documents

Email is still the channel where customers ask the same questions as in chat and wait the longest for an answer. Document search works here exactly as it does there; only the door changes.

What is inside

1. Scheduled trigger — the entry.
2. New messages fetches the unread and marks it read.
3. The question from the email joins the subject and the body into one query.
4. Search across documents pulls matching fragments from the knowledge base.
5. A draft answer — an agent with the same grounding rule as the support bot.
6. Send the answer goes back to the same sender.

Start by sending it to yourself

This is the riskiest scenario on the page: it writes to a living person without you. Change the recipient to your own address and watch for a week what the bot was going to say. Keep the “only from what was found” rule in the prompt regardless — without it the model fills the gaps in your documents with its own ideas, and the customer cannot tell the difference.

What to set up for yourself

• The recipient. Yours, until you trust it.
• The knowledge base and top_k on the search.
• What to do when there is no answer. The prompt asks it to say the question has been passed to a specialist. That is more honest than silence and safer than a guess.
• The subject line. Currently Re: original subject, so the reply lands in the same thread on the customer’s side.

What to connect

A mailbox with IMAP and SMTP (reading and sending), a knowledge base under Documents and the Schedule target. A separate mailbox for the bot beats a shared one: then its mistakes do not mix with your own correspondence.