OpenRouter Scoring Server (Local) — Rubric Integrated

This is a local FastAPI server that evaluates player counseling responses using an explicit rubric and OpenRouter LLM fallback. It performs a safety-first blacklist check, calls OpenRouter to get a structured JSON evaluation (score + sub-scores + feedback), post-processes results (deterministic formula), caches them, and returns them to the client (Unity or anything that can POST JSON).

Important:

• This server is designed to run locally (your machine). Do NOT store OpenRouter API keys in the Unity client.
• If you don't set an OpenRouter key, the server falls back to a heuristic evaluator so the game remains playable offline.

Files:

• server_openrouter.py — main FastAPI server (rubric + blacklist + OpenRouter integration)
• requirements.txt — Python deps
• .env.example — example env file

Quick start (Windows / macOS / Linux)

1. Create and activate a Python venv:

   • Windows (PowerShell): python -m venv .venv ..venv\Scripts\Activate.ps1
   • macOS / Linux: python -m venv .venv source .venv/bin/activate

2. Install dependencies: pip install -r requirements.txt

3. Add your OpenRouter API key:

   • Copy .env.example to .env and set OPENROUTER_KEY=sk_your_openrouter_key_here
   • If you do NOT want to use OpenRouter, leave it unset — server uses heuristics.

4. Run server: uvicorn server_openrouter:app --host 127.0.0.1 --port 8000 --reload

5. Test with curl (example): curl -X POST "http://127.0.0.1:8000/score" -H "Content-Type: application/json" -d "{"contextId":"c1","playerText":"Apologize sincerely, make sure she is safe and seek counseling."}"

5.1 Test with Powershell (example) $body = @{ contextId = "c1" playerText = "Apologize sincerely, make sure she is safe and seek counseling." } | ConvertTo-Json Invoke-RestMethod -Uri "http://127.0.0.1:8000/score" -Method Post -ContentType "application/json" -Body $body

Endpoints

• GET /health — basic status
• POST /score — body: { "contextId": "string", "playerText": "string" } → returns JSON evaluation: { "score": int, "empathy": int, "actionability": int, "safety": int, "accountability": int, "spiritual": int, "feedback": "string", "flagged": bool, "flag_reason": "string" }

Notes on rubric & safety

• Server performs a quick regex-based blacklist check for obviously harmful / violent instructions and immediately returns flagged response (score 0).
• Otherwise it constructs a compact prompt with few-shot examples and calls OpenRouter (temperature=0, deterministic).
• The server recomputes the final score using fixed weights to ensure consistency.

If you want, next I will:

• Add a small set of unit tests / example inputs (C step).
• Provide a Unity sample scene that posts answers and displays returned JSON.

Would you like me to prepare the Unity demo next or refine the few-shot examples further?