Apppublic

aadhesh2025/ticket-triage-env

sourceHugging Faceupdated 6mo agoView on Hugging Face

0likes

App README

🎫 Ticket Triage OpenEnv

A Real-World AI Agent Training Environment

Train AI agents to triage customer support tickets — classify, prioritize, and respond.

![OpenEnv](https://huggingface.co/spaces/aadhesh2025/ticket-triage-env) ![HF Space](https://huggingface.co/spaces/aadhesh2025/ticket-triage-env) ![Docker](https://github.com/Crypto-Type/ticket-triage-env) ![Python](https://python.org) ![FastAPI](https://fastapi.tiangolo.com)

🚀 Live Demo · 📖 API Docs · 📋 Tasks

</div>

🌍 Why This Environment Exists

Every company receives hundreds of customer support tickets every day. Human agents must read each ticket, decide what type of problem it is, how urgent it is, and craft a helpful response. This process is:

—⏱️ Time-consuming — reading and triaging takes hours daily
—🤔 Judgment-heavy — requires experience and domain knowledge
—📈 Scalability bottleneck — hard to scale with growing user bases

This environment trains AI agents to automate that entire workflow.

🏗️ Environment Architecture

┌─────────────────────────────────────────────────────────┐
│                  AI Agent (My Model)                │
│                        │                            │
│              ┌─────────▼──────────┐                   │
│              │   OpenEnv API      │                 │
│              │  POST /reset       │  ← Get a ticket │
│              │  POST /step        │  ← Submit answer│
│              │  GET  /state       │  ← Check status │
│              └─────────┬──────────┘                   │
│                        │                            │
│              ┌─────────▼──────────┐                   │
│              │  TicketTriage Env  │                 │
│              │  environment.py    │                 │
│              └─────────┬──────────┘                  │
│                        │                            │
│         ┌──────────────┼──────────────┐               │
│         ▼              ▼              ▼             │
│    Task 1 Grader  Task 2 Grader  Task 3 Grader      │
│    (Easy)         (Medium)       (Hard)             │
└─────────────────────────────────────────────────────────┘

🗂️ Tasks Overview

🟢 Task 1 — Ticket Categorization `(Easy)`

Field	Details
Task ID	`task1_categorize`
Input	Ticket subject + body text
Output	Category label
Valid Values	`billing` · `technical` · `general`
Scoring	`1.0` if correct · `0.0` if wrong

Example:

json

Input:  "I was charged twice for my subscription"
Action: { "category": "billing" }
Reward: 1.0

🟡 Task 2 — Category + Priority Assignment `(Medium)`

Field	Details
Task ID	`task2_prioritize`
Input	Ticket subject + body text
Output	Category + priority label
Valid Priorities	`low` · `medium` · `high`
Scoring	`0.5` per correct field · max `1.0`

Example:

json

Action: { "category": "technical", "priority": "high" }
Reward: 1.0  (both correct) | 0.5  (one correct) | 0.0  (both wrong)

🔴 Task 3 — Full Ticket Triage `(Hard)`

Field	Details
Task ID	`task3_full_triage`
Input	Ticket subject + body text
Output	Category + priority + customer reply
Reply Requirement	Minimum 10 words
Scoring	`~0.34` per correct field · max `1.0`

Example:

json

Action: {
  "category": "billing",
  "priority": "high",
  "reply": "We sincerely apologize. Our billing team will process your refund within 24 hours."
}
Reward: 1.0

📐 Observation & Action Spaces

Observation Space

json

{
  "observation": {
    "ticket_id": "T001",
    "subject":   "I was charged twice for my subscription",
    "body":      "Hello, I noticed two charges of $29.99 on my credit card..."
  },
  "task_id": "task1_categorize",
  "task_description": "Classify the support ticket into: billing, technical, or general",
  "action_schema": {
    "category": "string — one of: billing | technical | general"
  }
}

Action Space

Task 1: { "category": "billing" }

Task 2: { "category": "billing", "priority": "high" }

Task 3:

json

{
  "category": "billing",
  "priority": "high",
  "reply": "Thank you for contacting us. We will resolve your issue within 24 hours."
}

🏆 Reward Function

Task 1:  R = 1.0 if category correct, else 0.0

Task 2:  R = (0.5 × category_correct) + (0.5 × priority_correct)

Task 3:  R = (0.34 × category_correct)
           + (0.33 × priority_correct)
           + (0.33 × reply_length >= 10 words)

🔌 API Reference

Method	Endpoint	Description
`GET`	`/`	Live dashboard UI
`POST`	`/reset`	Start new episode
`POST`	`/step`	Submit action, get reward
`GET`	`/state`	Current environment state
`GET`	`/tasks`	List all tasks + schemas
`GET`	`/grader`	Last episode score
`GET`	`/baseline`	Run baseline agent

Python Example

python

import requests

BASE = "https://aadhesh2025-ticket-triage-env.hf.space"

# Start episode
obs = requests.post(f"{BASE}/reset", json={"task_id": "task1_categorize"}).json()
print(obs["observation"]["subject"])

# Submit action
result = requests.post(f"{BASE}/step", json={"action": {"category": "billing"}}).json()
print(f"Reward: {result['reward']}")

📊 Baseline Scores

Task	Avg Reward	Agent Type
`task1_categorize`	~0.33	Random agent
`task2_prioritize`	~0.28	Random agent
`task3_full_triage`	~0.56	Random agent

🚀 Local Setup

bash

# Clone
git clone https://github.com/Crypto-Type/ticket-triage-env
cd ticket-triage-env

# Install
pip install -r requirements.txt

# Run
uvicorn main:app --host 0.0.0.0 --port 7860 --reload

# Open: http://localhost:7860

Run Inference Script

bash

export API_BASE_URL="https://api-inference.huggingface.co/v1"
export MODEL_NAME="Qwen/Qwen2.5-72B-Instruct"
export HF_TOKEN="your_hf_token_here"
export ENV_BASE_URL="http://localhost:7860"

python inference.py

🐳 Docker

bash

docker build -t ticket-triage-env .
docker run -p 7860:7860 ticket-triage-env

📁 Project Structure

ticket-triage-env/
├── main.py            # FastAPI server + UI dashboard
├── environment.py     # Core env: reset(), step(), state()
├── tasks.py           # Tasks, tickets, grader functions
├── inference.py       # LLM inference script (OpenAI client)
├── baseline.py        # Random baseline agent
├── openenv.yaml       # OpenEnv specification
├── pyproject.toml     # Python project config
├── requirements.txt   # Python dependencies
├── Dockerfile         # Container definition
└── README.md          # This file

🧪 Environment Variables

Variable	Description	Default
`API_BASE_URL`	LLM API endpoint	`https://api-inference.huggingface.co/v1`
`MODEL_NAME`	Model identifier	`Qwen/Qwen2.5-72B-Instruct`
`HF_TOKEN`	Hugging Face API key	`""`
`ENV_BASE_URL`	Environment server URL	`http://localhost:7860`

📋 OpenEnv Compliance

—✅ Typed Pydantic models for Observation, Action, Reward
—✅ POST /reset — returns clean initial observation
—✅ POST /step — returns observation, reward, done, info
—✅ GET /state — returns current environment state
—✅ openenv.yaml with full metadata
—✅ Deployed on Hugging Face Spaces with Docker
—✅ inference.py using OpenAI-compatible client
—✅ Scores reproducible between runs

👤 Author

Aravind S — Built for OpenEnv Round 1 · Hugging Face × Meta AI Competition

Built with ❤️ using FastAPI · Docker · Hugging Face Spaces

🚀 Try it Live · 📖 API Docs

</div>

🎫 Ticket Triage OpenEnv

A Real-World AI Agent Training Environment

🌍 Why This Environment Exists

🏗️ Environment Architecture

🗂️ Tasks Overview

🟢 Task 1 — Ticket Categorization (Easy)

🟡 Task 2 — Category + Priority Assignment (Medium)

🔴 Task 3 — Full Ticket Triage (Hard)

📐 Observation & Action Spaces

Observation Space

Action Space

🏆 Reward Function

🔌 API Reference

Python Example

📊 Baseline Scores

🚀 Local Setup

Run Inference Script

🐳 Docker

📁 Project Structure

🧪 Environment Variables

📋 OpenEnv Compliance

👤 Author

🟢 Task 1 — Ticket Categorization `(Easy)`

🟡 Task 2 — Category + Priority Assignment `(Medium)`

🔴 Task 3 — Full Ticket Triage `(Hard)`