4.2 KiB
4.2 KiB
LinkedIn Draft & Consultant Review: Architecting a 24/7 AI Support & Entitlement Engine Without Human-in-the-Loop
文档性质：从
blog.hoelee.com改写并在 NocoDB 排期的 LinkedIn 生产级动态。 发布状态：已通过外部顾问审阅 / NocoDB 已排期待发布排期波次：Wave 2 (纵深期) - 槽位 5 (拟发布日期: 2026-11-12 Thu) 源文章：https://blog.hoelee.com/posts/how-i-built-the-digikedai-telegram-bot/ 系统标识：how-i-built-the-digikedai-telegram-bot| 分类：case-studies字数：2101 chars (严格控制在 3,000 字符内)
一、 核心背景与铁律约束提示（供复核对照）
- 求职画像契合（Senior 全栈/DevOps/AI 架构师）：彻底脱敏“做生意的老板/小公司创始人”倾向；一切商业实战均转换为“生产系统架构、高可用交付与自动化运维”视角。
- 纯洁人设与零禁用词：绝对禁止
homelab,home lab,家用,小主机,玩具,折腾,玩票等词汇。必须使用the infrastructure I manage,my self-hosted production nodes,production host。 - 第一人称单数：统一使用单数
I/我，杜绝虚张声势的复数we。 - 数据绝对保真：所有数字、性能比率、吞吐提升、故障现象严禁凭空捏造。
- OPSEC 安全脱敏：无内部绝对路径（如
D:/dev），无真实私有内网 IP，公网 IP 必须转为文档保留地址。
二、 LinkedIn 拟发布动态文案（已按顾问意见修订）
Deploying an LLM as a customer-facing bot is easy. Building an autonomous system that provisions production account entitlements without hallucinating or granting unauthorized access is an entirely different engineering challenge.
I architected a 24/7 support and automated trial-provisioning engine running across Telegram and WhatsApp. The stack: TypeScript + Node 20, grammY in webhook mode behind a Hono HTTP server, a self-hosted LiteLLM gateway, and PostgreSQL for persistent state.
Three architectural boundaries keep it reliable in production:
1. Three-Tier Retrieval Over Context Stuffing
Dumping an entire service catalog into system prompts guarantees hallucinated specifications and prompt injection risks. Instead, the engine resolves intent through a strict hierarchy:
• Deterministic plan/SKU regex token matching (`/\b[A-Z]{2,}\d{2,}\b/gi`)
• Normalized keyword metadata filtering
• Targeted semantic retrieval with bounded context injection
2. State Machines Over Prompt Engineering
An LLM is fantastic at natural language disambiguation and terrible at transactional state. Actions that mutate state — tenant trial provisioning, entitlement issuance, verification — run on deterministic TypeScript state machines. The LLM only parses intent; it never touches the business transaction directly.
3. The Webhook Timeout Trap
Telegram aggressively re-delivers webhook updates if your server does not respond within a tight window. If an LLM generation takes 4 seconds, Telegram re-fires the webhook, spawning duplicate model calls and cascading into an infinite reply storm. The fix: immediately acknowledge the incoming payload with HTTP 200, dispatch generation to an async worker queue, and enforce idempotency locks on the `update_id`.
The system runs autonomously in production, backed by ~175 unit tests verifying state transitions, retrieval precision, and payload contracts.
How do you separate deterministic business logic from probabilistic model output in your production agent workflows?
Write-up: https://blog.hoelee.com/posts/how-i-built-the-digikedai-telegram-bot/
三、 外部顾问审阅评估反馈单与归档记录
- 审阅结论：需微调（已更新修订稿）
- 1. 诊断与修改批注：求职红线防范：已按顾问修订稿消除电商副业色彩，强化 B2B 多租户权限开通与高可靠网关架构定位。
- 2. NocoDB 归档记录：
- Candidate Key:
cand:how-i-built-the-digikedai-telegram-bot(Status:APPROVED, Priority:HIGH) - Draft Key:
cand:how-i-built-the-digikedai-telegram-bot:v1(Status:APPROVED) - 拟定发布节奏：每周 2 篇（周二/周四），平稳推进 5 周。
- Candidate Key: