Sujith2005/ticket-triage-v3

OpenEnv: Customer Support Ticket Triage

Environment Description

Customer Support Ticket Triage is a complete, real-world OpenEnv environment where an AI agent acts as a customer support specialist. It must read incoming tickets, ask customers for more information when tickets are ambiguous, and route them to the correct department with the proper category and priority.

This is not a toy problem—it simulates real helpdesk routing logic often handled by humans or automated systems. The environment uses the standard step() / reset() / state() API through FastAPI.

Action Space

The agent can perform the following actions via POST /step:

• —read: Read the full content of a ticket by providing ticket_id.
• —ask_customer: Ask the customer for more information by providing ticket_id and question.
• —route: Route a ticket using ticket_id, category, priority, and department.
• —Categories: billing, technical, sales, general
• —Priorities: low, medium, high
• —Departments: finance, engineering, sales, support
• —submit: Finish the episode once triage is complete.

Observation Space

Returns JSON containing:

• —feedback: String containing the result of the last action (e.g. ticket content, customer reply, or success/error message).
• —inbox_summary: List of tickets currently pending.
• —done: Boolean indicating episode completion.

Rewards

The environment provides dense, meaningful rewards pointing towards partial progress:

• —+0.2 for successfully routing a ticket.
• —+0.1 for asking a necessary clarifying question.
• —+0.01 for reading a ticket (to encourage exploration).
• —-0.05 for invalid actions, bad routing attempts, or unnecessary questions.

Setup Instructions

1. Docker (Recommended for HF Spaces)

docker build -t openenv-triage .
docker run -p 7860:7860 -e OPENAI_API_KEY=\"your_api_key_here\" openenv-triage

2. Local Python Setup

pip install -r requirements.txt
uvicorn main:app --host 0.0.0.0 --port 7860

Running the Baseline

The baseline uses the OpenAI API client (gpt-4o-mini) to infer against the environment tasks. Make sure you have your open API key set in your terminal:

export OPENAI_API_KEY=\"sk-your-key\"

Run via endpoint: POST http://localhost:7860/baseline

Or run locally via script:

python baseline.py

Baseline Scores (Typical)

• —task_1_easy (1 ticket): 1.0
• —task_2_medium (3 tickets): 1.0
• —task_3_hard (3 tickets + 1 ambiguous requiring question): 1.0

Endpoints

This environment fully complies with OpenEnv specs and standard HF Space requirements.

• —POST /reset - Resets and returns initial observation.
• —POST /step - Takes Pydantic Action, returns StepResponse.
• —GET /state - Dumps full state dictionary.
• —GET /tasks - Lists tasks (task_1_easy, task_2_medium, task_3_hard) and action schema.
• —GET /grader - Evaluates the run (0.0 - 1.0).
• —POST /baseline - Triggers the background inference script baseline. "# openenv-ticket-triage"