@unerr-ai/unerr-linux-x64

unerr-docs

Documentation site for unerr, served at docs.unerr.dev. It is a Next.js 16 App Router app that renders MDX docs with Fumadocs. Pages live in content/docs/**; adding a page means dropping one .mdx file.

Stack

Part	Version / choice
Framework	Next.js 16.2.9 (App Router, Turbopack, output: "standalone")
UI	React 19.2.7, Tailwind v4
Docs engine	fumadocs-ui 16.10.2, fumadocs-core 16.10.2, fumadocs-mdx 14.2.11
Language	TypeScript 5.9
Package manager	pnpm@10.26.2 (Corepack), Node >=20.9.0

Quick start

corepack enable
pnpm install
pnpm dev

Open http://localhost:3000. predev runs fumadocs-mdx to generate .source/ before the dev server starts.

Other commands:

Command	Does
pnpm build	Production build
pnpm start	Run the built app
pnpm lint	ESLint
pnpm typecheck	fumadocs-mdx && next typegen && tsc --noEmit
pnpm format / pnpm format:check	Prettier write / check

Project layout

Path	What
content/docs/**	MDX doc pages; folder order via meta.json
app/	App Router: (home), docs/[[...slug]], api/search, api/chat, api/health, llms.txt/llms-full.txt/llms.mdx, og/
lib/source.ts	Fumadocs source loader
lib/site-url.ts	Runtime base URL (from SITE_URL)
lib/layout.shared.tsx	Nav and branding
components/	Shared components
proxy.ts	Markdown content negotiation
source.config.ts	Fumadocs MDX config
.source/	Generated by fumadocs-mdx (git-ignored, regenerated on install and build)

Adding a doc page

See CONTRIBUTING.md. Short version: add an .mdx file under content/docs/, set its order in the section's meta.json, run pnpm dev to preview.

Search and Ask-AI

• Search: built-in Orama at /api/search. No external service.
• Ask-AI: /api/chat streams from ChatGPT via the Vercel AI SDK + @ai-sdk/openai, using OPENAI_API_KEY. Optional — when the key is unset, the Ask-AI trigger is hidden and keyword search still works.
• Ask-AI abuse/cost guards (lib/ai/guards.ts): same-origin check, per-IP rate limit (12 req/60s, in-memory), input caps (≤24 messages, ≤12k chars), and an 800-token reply cap. These are the first layer only — they are per-task, so durable cluster-wide limits belong at the ALB / AWS WAF (aws-infra).
• Agent-readable docs: /llms.txt, /llms-full.txt, and per-page .md.

Environment variables

No NEXT_PUBLIC_* vars. They are inlined at build time, so one Docker image could not serve multiple environments. All per-environment config is read at runtime on the server.

Variable	When	Purpose
SITE_URL	Runtime	Base URL (read via lib/site-url.ts)
OPENAI_API_KEY	Runtime	Ask-AI key; optional. Unset = Ask-AI disabled
OPENAI_MODEL	Runtime	ChatGPT model for Ask-AI; injected by ECS

Deployment

Docker image (multi-stage Alpine, standalone, non-root, tini) → AWS ECR → ECS/Fargate behind an ALB, at docs.unerr.dev. Health check at /api/health.

CI/CD: .github/workflows/ci-cd.yml — test → verify-image on PR → build-push on main → manual deploy-qa / deploy-prod. AWS infra (ECR/ECS/ALB/DNS) lives in the separate aws-infra Terraform repo.

See .internal/IMPLEMENTATION_PLAN.md for the full setup plan and status.