Lab build: a public Telegram support bot that routes 30/30 and cites its source docs

The problem

Public support bots are easy to demo and hard to run: they hallucinate answers, can't actually check a ticket, never admit when a human is needed, and fall over the moment real traffic (or Telegram's retry storms) arrives.

The approach

1. Answers from the product's help docs with RAG and cites the source article inline, so a customer can verify the answer instead of trusting it. Out-of-scope questions get an honest refusal, not a guess.
2. Looks up real ticket status on request, and routes billing disputes or an explicit 'get me a human' to honest escalation: the transcript is forwarded and the customer is told a person will follow up, rather than the bot pretending to resolve it.
3. Hardened for public traffic: webhook secret auth, update_id idempotency so Telegram's retry storms can't double-process a message, per-chat rate limits plus a global daily LLM-call budget, and a degraded no-DB mode that keeps answering docs questions when the database is down.
4. Guarded by two repeatable evals: npm run eval:routing over docs/ticket/escalate/refuse intents, and npm run eval:retrieval over the golden question set.

The results

A public bot you can actually message: it answers from real docs with citations, checks real tickets, hands billing disputes to a human honestly, and stays up under retry storms and load. Measured, not asserted, and running at $0 on the free tier.