n8n

SupportRAG

An AI customer-support email agent on n8n that drafts a policy-grounded reply to every email, then waits for human approval on Telegram before anything is sent.

View on GitHub (private) →

Overview

SupportRAG is an AI assistant for customer-support email, built end to end in n8n. It reads every message that arrives, understands what the customer needs, writes a helpful reply grounded in the company's own policy documents, and waits for a person to tap Approve before anything is sent. Angry or urgent emails skip the queue and alert the team instantly.

The design goal is a support teammate that drafts every reply in seconds, never invents answers, and always asks before hitting send.

WF3, the RAG reply drafter: retrieves from the knowledge base, drafts a grounded reply, and routes it to Telegram for approval.

The Problem

Support teams spend hours every day answering the same handful of questions: warranty length, order status, how to get a refund. Meanwhile, a genuinely upset customer can sit unnoticed in the same crowded inbox. SupportRAG removes the repetitive drafting work and makes sure the messages that need a human get one immediately.

How It Works

The system is six coordinated n8n workflows. An inbox-triage workflow acts as the orchestrator and calls the others as sub-workflows.

Knowledge ingestion chunks the company's help documents, embeds them with OpenAI, and stores the vectors in a Qdrant collection, the retrieval layer the drafter reads from.

WF1, knowledge ingestion: splits and embeds the knowledge base into Qdrant.

Inbox triage & routing is triggered on each new email. It classifies the message (topic, sentiment, urgency), logs it, and routes it one of three ways: straight to escalation for anything urgent or a complaint, to the RAG drafter for answerable questions, or to a human queue when it's unsure.

WF2, inbox triage: classifies each email, then routes it to draft, escalate, or a human.

The RAG reply drafter (shown above) is the showpiece. A retrieval-augmented agent searches the knowledge base, drafts a grounded reply that cites the sources it used, and passes a confidence gate. Confident drafts are sent to Telegram with Approve / Reject buttons; only on approval does the reply go out. Low-confidence cases are saved as a Gmail draft for a human instead of guessing.

Escalation handles the urgent path: it summarizes the situation with an LLM and posts an immediate Telegram alert to the team with a suggested next step.

WF4, escalation: summarizes urgent or angry mail and alerts the team instantly.

Two supporting workflows round it out: an error handler that posts a Telegram alert if any workflow fails, and a scheduled daily digest that aggregates the day's ticket stats from Google Sheets and posts a summary.

WF5a, the error handler: catches any workflow failure and alerts the team on Telegram.

WF5b, the daily digest: reads the day's tickets from Google Sheets, aggregates the stats, and posts a summary.

Guardrails

It won't make things up. Answers come only from the provided documents; if the answer isn't there, it says so and hands off to a person.
A human is always in control. No reply is ever sent automatically; someone approves each one.
Everything is logged. Every email, decision, and reply is recorded in Google Sheets for a full audit trail.

Tech Stack

n8n: the automation engine orchestrating all six workflows.
OpenAI (gpt-4o-mini): classification, summarization, and grounded reply drafting, plus embeddings for retrieval.
Qdrant: vector database powering the RAG knowledge base.
LangChain nodes (in n8n): the retrieval agent, LLM chains, and structured output parsing.
Gmail, Google Sheets, Telegram: inbox and sending, logging, and the human-in-the-loop approval / alert layer.
Docker Compose: runs n8n, Qdrant, and a Cloudflare tunnel together, so the public webhook needed for Telegram approvals works out of the box.