vivekvish2004/openenv-customer-support

🎫 OpenEnv Customer Support Environment

A high-fidelity, real-world OpenEnv simulation environment designed to train and benchmark AI agents in enterprise customer support decision-making.

Implements the full OpenEnv step() / reset() / state() API with standard Pydantic models.

---

💡 Motivation & Real-world Relevance

In modern enterprise operations, customer support is not just about answering questions—it's about complex multi-step decision-making under SLA (Service Level Agreement) pressure. Handling a support queue requires consistent logic, empathetic communication, and accurate technical classification.

This environment provides a structured sandbox for AI agents to master:

• Triage: Accurately classifying issues to route to the correct engineering teams.
• Prioritization: Balancing customer sentiment with business urgency.
• Empathy: Nuanced response generation for frustrated or panicked users.
• Workflow Integrity: Ensuring all steps (category, priority, response) are completed before resolution.

---

🛠️ Environment Specification

Action Space

The agent interacts with the environment using a typed Action model.

Action Type	Payload Description	Allowed Values
classify_ticket	Categorize the issue	refund, technical_issue, login_issue, general_inquiry, feedback, security
assign_priority	Set business priority	low, medium, high
generate_response	Draft a text response	Any string (e.g., "I'm sorry for the inconvenience...")
search_kb	Query internal policy	Returns technical/billing policy facts
ask_clarification	Request missing info	Used for vague tickets to unlock resolution
resolve	Close the ticket	{} (Requires classification, priority, and response)
escalate	Direct to senior level	{} (Appropriate for high-sentiment/emergency)

Observation Space

The environment returns a comprehensive state dictionary in every step.

Key	Type	Description
ticket_text	string	The raw customer inquiry text.
sentiment	string	Customer mood: angry, panicked, curious, happy, concerned, neutral.
status	string	Lifecycle state: open, closed, session_complete.
priority	string	The currently assigned priority.
classification	string	The currently assigned category.
steps_taken	int	Number of actions performed on the current ticket.
sla_limit	int	Maximum steps allowed for this ticket type.
total_reward	float	Cumulative reward across the entire 3-ticket session.
last_step_status	string	Result of the previous action: success, failed, neutral.
kb_context	string	Contains the most recent Knowledge Base search result.
is_clarified	bool	True if the agent has asked for clarification.

---

📈 Reward Function

The environment utilizes a dense reward function to provide guidance throughout the trajectory:

• Correct Classification: +0.35 (Penalty for wrong: -0.20)
• Correct Priority: +0.25 (Penalty for wrong: -0.15)
• Professional Response: +0.20
  • Empathy Requirement: Responses to upset/panicked customers must contain empathy keywords.
• Successful Resolution: +0.40
  • SLA Penalty: -0.25 if resolved after the SLA step limit.
• Efficiency Penalty: -0.02 per step to encourage direct, non-redundant behavior.

---

🏁 Baseline Benchmarks

Verified scores from the consolidated validation suite.

Agent Type	Avg. Total Reward	Queue Completion Rate	Evaluation
Random Agent	-0.85	0%	Failed
Simple Heuristic	1.45	45%	Moderate
Perfect Baseline	3.36	100%	Excellent

---

🚀 Getting Started

Installation

# Clone and install dependencies
git clone <repo_url>
pip install -r backend/requirements.txt

Running Locally

1. Start the Backend:

   python3 backend/main.py
   

2. Launch the Dashboard:

   cd frontend && npm install && npm run dev
   

Running Inference

Use the standard OpenEnv inference script to run your model (requires OPENAI_API_KEY):

python scripts/inference.py

---

🧪 Evaluation & Grading

The environment includes 10 deterministic graders spanning Easy, Medium, Hard, and Extreme difficulties.

• EASY Tasks: Single-attribute checks (e.g., correct classification).
• MEDIUM Tasks: Partial workflow checks (e.g., Priority + Response empathy).
• HARD Tasks: Full end-to-end lifecycle resolution under SLA constraints.
• EXTREME Tasks: Multi-turn workflows requiring Knowledge Base (KB) lookups, cross-referencing policies, and clarification of vague customer inputs.

---

🏆 Judging Criteria & Technical Compliance

Criterion	Implementation in Customer Support Env
Real-world Relevance	Simulates complex enterprise support workflows (SLA, Triage, Policy).
Complexity	10 hand-crafted tasks with tiered difficulty (Easy to Extreme), multi-step trajectories (up to 12 steps), and state-dependent logic.
Standard Compliance	100% adherence to OpenEnv step()/reset()/state() API. Verified via openenv validate.
Robustness	Session-based isolation for multi-agent evaluation and error-resilient inference pipeline.
Realism	Uses 12 high-fidelity customer scenarios with dynamic sentiment decay (customers get angrier over time).

---

📋 Detailed Task Matrix

Task ID	Name	Difficulty	Core Objective	Evaluation Metric
task_easy_1	Classification	EASY	Categorize issue correctly	Classification match (0.99)
task_easy_2	Priority	EASY	Set correct priority	Priority match (0.99)
task_medium_1	Empathy Check	MEDIUM	Classify + Respond	Classification + Empathy keywords
task_medium_2	Professionalism	MEDIUM	Classify + Solve	Classification + Solution keywords
task_hard_1	Full Lifecycle	HARD	End-to-end resolution	Success on all 4 steps (0.25 each)
task_hard_2	De-escalation	HARD	High-priority anger	High Priority + Empathy keywords
task_hard_3	SLA Challenge	HARD	Efficiency under pressure	Bonus for fewer steps (≤4)
task_extreme_1	Policy Lookup	EXTREME	Use KB for decision	KB match + Policy citation
task_extreme_2	Vague Input	EXTREME	Clarify before resolve	ask_clarification call requirement
task_extreme_3	Security P0	EXTREME	Handle breach	Security KB + High Priority + Escalate

---

🛡️ Reliability & Concurrency

Session Isolation

The backend supports concurrent evaluation of multiple agents. By using the session_id query parameter, each evaluator gets a dedicated, isolated environment instance to prevent state crosstalk.

Robust Inference

The provided inference.py includes built-in retry logic (max 3 attempts) and multi-pass JSON validation. This ensures the evaluation pipeline is resilient to transient LLM failures or malformed model outputs.

---

📄 License

MIT © 2024