Plan: Transitioning to Forward Deployed Engineer / AI Engineer
一份分階段路線圖,供後端工程師轉向 FDE 與 AI Engineer 角色——解讀頭銜混亂、應用型 LLM 技術棧與面試循環。
AI 翻譯
尚未偵測到更新。GitHub 發布版本後會顯示在這裡。
個人 GitHub 帳號下公開儲存庫的擁有者可以登入並驗證所有權,自行管理介紹,無需管理員核准。
登入並產生驗證檔案,在 30 分鐘內將其提交至此儲存庫的預設分支。其他分支或分叉儲存庫中的檔案無效。然後在下方檢查檔案,即可無需管理員核准直接管理專案。請保留檔案以維持管理權限。
如需登入,將前往登入頁面。登入後請回到此處,再次點擊所選的驗證按鈕。
在專案資料夾中執行 /nomorevibe verify,即可驗證所有權並自行管理介紹。
一分鐘完成驗證SocratEase – a Socratic RAG course assistant for fullstack TypeScript
翻譯準備中
取自儲存庫 README
正在載入完整 README。
檢視完整 README ↗A Socratic course assistant for a fullstack JavaScript/TypeScript program. It answers students only from the course material in Notion and never solves assignments.
GLOSSARY.mddocs/adr/docs/spec-v1.md and the issues on the project boardRequirements: Node 22+, pnpm and Docker.
pnpm install
cp .env.example .env.local # local database URL
pnpm db:up # start Postgres with pgvector (Docker Compose)
pnpm db:migrate # apply migrations
pnpm dev # http://localhost:3100
The start page shows whether the database is reachable, how many migrations are applied and the pgvector version.
| Script | What it does |
|---|---|
pnpm db:up | Starts local Postgres (port 54329) and waits until it's healthy |
pnpm db:migrate | Applies migrations. On Vercel it runs as part of the build, the same way as locally. Preview deployments skip it unless ALLOW_PREVIEW_MIGRATIONS=1, so a PR never migrates a database shared with production |
pnpm db:generate | Generates a new migration from the schema |
pnpm typecheck | Generates Next.js route types and runs tsc |
pnpm lint | ESLint |
pnpm test | Vitest, against its own database (<name>_test), with the fake AI provider |
pnpm evaluate | Runs the Evaluation Set against the real AI provider (see below) |
The Evaluation Set is a fixed list of questions about the fictional test Course, with the behaviour expected of each answer. The questions live in evaluation/cases.ts, grouped as in the v1 spec:
Run it before every release, and on every PR that changes how the assistant answers (prompts, retrieval, chunking or the chat model):
pnpm db:up
AI_PROVIDER=local pnpm evaluate # or AI_PROVIDER=vertex once it is set up
It indexes the test Course into its own database (<name>_eval), asks every question once through the real provider, prints each failed question with its answer, and a pass/fail count per group. Any failure in group 4 blocks the release: the command then exits with code 1. Failures in the other groups are for the Teacher to judge, since a real model's answers vary between runs.
Only the fictional test Course is ever sent, so any provider may be used, including local (ADR 0008).
The app runs on Vercel (region fra1), with a managed Postgres with pgvector in the EU (ADR 0004). DATABASE_URL is set in the Vercel project.