AMS Check

Single-project Node/TypeScript app.

All user-owned code is under src/:

• src/core - deterministic fetch/extract, LLM adapters, scoring
• src/api - Express API
• src/web - React + Vite frontend
• src/cli - CLI command

Architecture

5 stages:

URL → [1. Fetch, Extract & Parse] → [2. LLM Extraction] → [3. Deterministic Checks] → [4. LLM Assessment] → [5. Scoring]

Setup

pnpm install
cp .env.example .env

Recommended .env:

PORT=3001
LLM_PROVIDER=anthropic
LLM_API_KEY=your_provider_api_key_here
LLM_MODEL=
LLM_TIMEOUT_MS=120000
MAX_ARTICLE_CHARS=20000
CONTEXT_FILE_PATH=context.txt
VITE_API_URL=/api/analyse
VITE_API_START_URL=/api/analyse/start
VITE_API_STATUS_URL_BASE=/api/analyse/status

LLM_MODEL is optional. If empty, provider defaults are used:

• Anthropic: claude-haiku-4-5
• OpenAI: gpt-5-mini-2025-08-07
• Gemini: gemini-3-flash-lite-preview
• Grok: grok-4-1-fast-non-reasoning-latest

Performance tuning:

• LLM_TIMEOUT_MS: timeout for provider requests (default 90000)
• MAX_ARTICLE_CHARS: max article chars included in the LLM prompt (default 12000)
• CONTEXT_FILE_PATH: where clean extracted article text is saved (default context.txt)

Run

Run both the API and front end in separate terminals. I recommend using the front end, but it is possible to use the CLI (although not as nice)

API:

pnpm dev:api

Web:

pnpm dev:web

The web UI shows a single live STATUS: ... line that updates until completion. Each completed run overwrites context.txt with clean extracted article text.

• API: http://localhost:3001
• Web: http://localhost:5173

Health check:

curl http://localhost:3001/health

API

• GET /health
• POST /analyse
• POST /analyse/start (starts async analysis and returns requestId)
• GET /analyse/status/:requestId (returns latest status, result, or error)

Request body:

{
  "name": "John Smith",
  "dob": "1980-03-15",
  "url": "https://example.com/article"
}

CLI

Run:

pnpm analyse

Interactive behavior:

• CLI first checks .env for LLM_PROVIDER, LLM_MODEL, and LLM_API_KEY.
• If any of those are missing (or invalid provider), it prompts for them one-by-one.
• Once LLM settings are resolved, it prompts one-by-one for:
  • name
  • dob (YYYY-MM-DD)
  • url (http(s)://...)

You can still pass values as flags to skip prompts for those fields:

pnpm analyse --name "John Smith" --dob "1980-03-15" --url "https://..." --provider grok --api-key "xai-..." --model "grok-3-fast-latest"

Behavior:

• Calls core engine directly
• Prints live STATUS: ... progress lines to stderr while running
• Prints JSON to stdout
• Errors go to stderr and exit code 1

Build and Typecheck

pnpm typecheck
pnpm build