以 Arabic-first 設計支援分流,並設置安全發送邊界
Designing Arabic-first support triage with a safe send boundary
這個 n8n 示範專案為阿拉伯語優先、兼顧英語的支援訊息設計分流流程,讓系統分類訊息及擬寫回覆,同時避免生成內容未經確認便發送。專案分為 9 個工作流程,使用 4 個 n8n Data Tables,以 Gemini 為主要模型、Groq 為後備模型;遇上不確定或敏感個案時可交由人工處理。離線檢查有 72 項通過,但即時發送功能已停用,亦未經驗證。
Support automation should help classify and draft a response without quietly turning a generated answer into a sent message. This n8n portfolio demo explores that boundary for Arabic-first and English support requests.
What the workflow does
The project is split into nine workflows with four n8n Data Tables. Channel adapters cover a web form, Gmail, Telegram, WhatsApp, and a simulated multichannel API. A classifier routes the message into FAQ matching and a knowledge-grounded draft path. Gemini is the primary model with Groq as a configured fallback.
The workflow records ticket and event data, can prepare a response draft, and has an escalation path when the message needs human attention. The design goal is to give an operator context and a proposed next step while leaving uncertain or sensitive cases for review.
That separation matters for a support inbox: a plausible answer is not necessarily an authorized answer, and a successful model response is not the same as successful delivery to a customer. Channel boundaries and escalation rules should be explicit before connecting real accounts.
What is and is not verified
The repository reports 72 of 72 offline checks passing. An author report describes local tests using fictional messages, but I have not independently repeated those on the user's n8n runtime. Live Gmail, Telegram, and WhatsApp sending is off and not verified; the public demo sets ALLOW_LIVE_SEND=false. Some channel behavior is simulated. The walkthrough videos are screenshot-led edits, not direct screen recordings.
This is not a production support service, SLA, guaranteed-accuracy system, implemented RAG stack, or booking automation. To adapt it for a business, the message sources, knowledge ownership, escalation criteria, privacy requirements, retention, and approved sending channels would need to be defined and tested with that organization.
Source and setup notes: https://github.com/ahmednegm91/ai-customer-support-triage
This project demonstrates an architecture for assisted triage and review; it does not claim verified delivery or autonomous customer support.
來源:DEV Community · dev.to