Auto-answer Telegram DMs from a Notion FAQ

Build me an agent that auto-answers common questions in my Telegram DMs from my Notion FAQ database, replying as me through my connected secretary bot. Anything the FAQ cannot confidently answer should be queued for me to reply to instead.

Trigger: run on a cron every 10 minutes.

On each run:

1. Use telegram-secretary Receive Updates to pull new incoming updates. Only consider new business_message updates from managed private chats. Advance the offset so each message is processed once.

2. Skip any message whose chat id is on my per-chat opt-out list (private threads with friends and family). Also skip anything the agent has already answered in this run.

3. Decide whether the message is question-shaped: it ends in a question mark, uses question phrasing (how do I, can you, what is, do you, when, where, why, pricing, refund, hours, etc.), or is otherwise a clear ask. Small talk, thanks, reactions, and forwarded content should be ignored.

4. For each question, look for a matching FAQ entry in my Notion FAQ database. First try notion Query a Data Source with a text filter on the Question property; if that returns nothing useful, fall back to notion Search by Title scoped to that database. Pick the single best match and score how confident you are that it actually answers this question.

5. If the confidence is high, draft a short, friendly reply in my voice (two to four sentences, no corporate tone, no emoji unless the sender used one) that pulls the answer from the FAQ entry. Send it with telegram-secretary Send Message as User in the same chat, using the business_connection_id from the incoming update and quoting the original message with reply_parameters. Then call telegram-secretary Mark Messages as Read on that message so my unread count reflects reality.

6. If the confidence is low, do NOT send anything. Instead use notion Create a Page to append a new row to my 'needs my reply' Notion page (I will give you its id) with these properties: Sender name, Chat id, Received at (ISO timestamp), Message (the original text), Suggested draft (your best attempt at a reply), and Closest FAQ (the FAQ entry title you considered, if any).

Inputs I want to configure on the workflow:

- FAQ data source id in Notion (the database of questions and answers).

- 'Needs my reply' Notion page or database id.

- Opt-out list: an array of Telegram chat ids that must never be auto-answered.

- Confidence threshold (default: only auto-reply when the match is clearly correct).

- Reply voice guidelines (default: warm, concise, first person, no jargon).

Guardrails:

- Never send more than one auto-reply per chat per run, even if the sender fired off several questions in a row. If there are multiple, answer the most recent one and mark the rest as read.

- Never auto-reply if the message contains anything sensitive (legal threats, medical questions, payment disputes, refund escalations, complaints). Log those to 'needs my reply' instead.

- Never invent facts. If the FAQ does not explicitly cover something, treat it as low confidence and route to 'needs my reply'.

- Respect Telegram flood limits (do not blast; pace sends).