title	Openenv Customer Support
emoji	🎯
colorFrom	pink
colorTo	red
sdk	docker
pinned	false

OpenEnv Customer Support Triage RL Environment

A customer support triage RL benchmark inspired by real-world support chaos. 🎯

---

Why This Exists—And What’s Actually Challenging Here

• Difficulty = not just more tickets: Higher levels mean genuinely ambiguous, conflicting, or spammy tickets—no keyword-only solution will suffice.
• No “trivial exploit” risk: Ground-truth labels for route/urgency/difficulty/priority never leave the backend. Agents only see reward and real observable metadata. (Judges: test /state, /step, /reset and see for yourself!)
• Reward shaping: You get partial credit when nearly correct; but full reward only if all outputs are spot-on, per ticket.
• Ambiguity + edge-cases: “Hard” mode includes multi-signal, multi-category, and spam tickets, inspired by messy real support data.
• Totally reproducible: Reset with a fixed seed allows full episode determinism via API and in our pipelined baseline script.

---

The Environment—How It Feels

At each step: You get a support ticket (subject, description, and metadata).
You must output a JSON action like:

{
  "action": {
    "route_category": "billing | technical | feature | feedback | spam",
    "urgency_assessment": "low | medium | high | critical",
    "resolution_difficulty": "easy | medium | hard",
    "priority_score": 0   // integer [0–100], higher = more urgent
  }
}

Modes:

• Easy: Clear intent, correct label, obvious answer.
• Medium: Sometimes ambiguous, possible overlap, initial mislabels.
• Hard: Real noise, mixed signals, “tricks” and realistic spam.

---

🛡️ Anti-Cheat Design

• No label leakage—agent never sees correct answers in any /reset, /step, or /state output.
• Only reward, correctness, and episode stats are exposed to the agent.
• (Seriously, try to break it!)

---

Quickstart

Build and run locally:

docker build -t openenv-customer-support .
docker run --rm -p 7860:7860 openenv-customer-support

Visit: http://localhost:7860/docs

Endpoint	Method	What it does
/reset	POST	Start new episode
/step	POST	Take agent action
/state	GET	Get current obs/ep
/docs	GET	API documentation
/health	GET	Health check

Remote Space:
https://johan45-openenv-customer-support.hf.space

---

📊 Baseline & Evaluation

Run the included baseline script:

python scripts/baseline_inference.py --difficulties easy medium hard --episodes 3 --seed 42

Scores (seed=42, episodes=3):

Difficulty	Score	Route acc.	Efficiency	Target	Pass
Easy	0.94	1.00	0.93	>0.85	✅
Medium	0.87	1.00	0.83	>0.70	✅
Hard	0.80	0.87	0.82	>0.50	✅

These are always reproducible with the included script and seed.

---

FAQ

• Can I “cheat” and find the true answer with an agent?
  No! All ground-truth is backend only.
• Why is hard mode so challenging?
  Realistic ambiguity, wrong categories, urgent spam, and blended requests—like a real support desk.
• Do I need a requirements.txt file?
  Nope—pyproject.toml and uv.lock are all you need for Docker and Python installs.

---

Project Structure

• server/ — FastAPI app and environment
• Dockerfile — container build
• pyproject.toml, uv.lock — dependency management
• scripts/baseline_inference.py — baseline agent + runner

To validate:

openenv validate

---