Apppublic

Harshvardhan-M/customer-support-triage

sourceHugging Facemitupdated 6mo agoView on Hugging Face

0likes

App README

🎧 Customer Support Triage — OpenEnv Environment

An OpenEnv-compliant RL environment where AI agents learn to triage real-world customer support tickets: classify urgency, route to departments, escalate sensitive cases, and draft responses.

👉 [Interactive Demo →](/demo)

Why This Environment?

Customer support triage is a high-volume, high-stakes real-world task:

—Misrouting or mis-prioritizing tickets costs revenue and customer trust
—Requires multi-step reasoning: inferring urgency from tone, applying business rules, recognising legal risk
—Evaluation is deterministic — no human labellers needed
—The RL/agent community can immediately use this for evaluation and training

Tasks

Task	Difficulty	Max Steps	Description
`urgency_classification`	🟢 Easy	10	Classify 5 tickets as low\	medium\	high\	critical. Partial credit for adjacent levels.
`department_routing`	🟡 Medium	18	Route 6 tickets to billing\	technical\	returns\	general\	manager. Optional urgency bonus.
`full_triage`	🔴 Hard	40	Handle 5 complex tickets: classify → route → escalate → draft response → resolve.

API

bash

# Reset
curl -X POST http://localhost:7860/reset \
  -H "Content-Type: application/json" \
  -d '{"task_name": "full_triage"}'

# Step
curl -X POST http://localhost:7860/step \
  -H "Content-Type: application/json" \
  -d '{"action_type": "classify", "urgency": "critical"}'

# State
curl http://localhost:7860/state

# Validate spec compliance
curl http://localhost:7860/validate

Endpoint	Method	Description
`/`	GET	Health check
`/reset`	POST	Start episode: `{"task_name": "..."}`
`/step`	POST	Take action
`/state`	GET	Full internal state
`/tasks`	GET	Task metadata
`/validate`	GET	OpenEnv spec compliance check
`/docs`	GET	Swagger UI
`/demo`	GET	Interactive Gradio demo

Observation Space

json

{
  "current_ticket": {
    "ticket_id": "TKT-005",
    "subject": "URGENT: Production database is DOWN",
    "body": "...",
    "product": "CloudDB Enterprise",
    "channel": "phone",
    "created_at": "2024-03-15T02:47:00Z"
  },
  "customer": {
    "email": "sre@megacorp.com",
    "tier": "enterprise",
    "account_age_days": 1825,
    "total_tickets": 42,
    "previous_tickets": [...]
  },
  "queue_position": 3,
  "total_in_queue": 5,
  "task_name": "full_triage",
  "available_actions": ["classify","route","escalate","draft_response","resolve"],
  "step_count": 7,
  "episode_reward": 0.85,
  "done": false,
  "last_action_feedback": "dept='technical' vs 'billing' → 0.00",
  "last_action_error": null
}

Action Space

json

{
  "action_type": "draft_response",
  "response_draft": "We are treating this as a P0 incident and escalating immediately.",
  "reasoning": "Enterprise customer, $3k/min revenue loss — needs immediate human escalation."
}

Field	Type	Values
`action_type`	string (required)	classify \	route \	escalate \	draft_response \	resolve \	request_info
`urgency`	string\	null	low \	medium \	high \	critical
`department`	string\	null	billing \	technical \	returns \	general \	manager
`response_draft`	string\	null	Customer-facing text
`escalate_reason`	string\	null	Reason for escalation
`reasoning`	string\	null	Chain of thought (not scored)

Reward Function

Shaped across the full trajectory — not sparse:

Component	Description
`urgency_accuracy`	Partial credit: 1.0 exact, 0.5 off-by-1, 0.2 off-by-2
`routing_accuracy`	Binary: 1.0 correct department, 0.0 wrong
`escalation_correctness`	Correct=+1.0, wrong decision=−0.3
`response_quality`	Keyword coverage of expected response terms
`efficiency_bonus`	Fast resolution of critical tickets
`error_penalty`	Invalid actions, approaching step limit
Terminal	+0.3 if final score ≥ 0.9, −0.2 if < 0.4

Reward clamped to [−1.0, 1.0] per step.

Setup

bash

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

# Docker
docker build -t cst-env .
docker run -p 7860:7860 cst-env

# Run baseline agent
export API_BASE_URL="https://api.openai.com/v1"
export MODEL_NAME="gpt-4o-mini"
export HF_TOKEN="your-key"
export ENV_BASE_URL="http://localhost:7860"
python inference.py

# Pre-submission validation
python validate.py

Baseline Scores

Measured with gpt-4o-mini at temperature 0:

Task	Difficulty	Grader Score
`urgency_classification`	Easy	~0.72
`department_routing`	Medium	~0.61
`full_triage`	Hard	~0.38

The hard task genuinely challenges frontier models — escalation decisions and response quality both require multi-step business reasoning.

Project Structure

├── app.py              FastAPI app (HF Space entry point)
├── demo.py             Gradio interactive demo (mounted at /demo)
├── inference.py        Baseline agent script
├── validate.py         Pre-submission validator
├── openenv.yaml        OpenEnv spec metadata
├── Dockerfile
├── requirements.txt
├── README.md
└── env/
    ├── __init__.py
    ├── environment.py  Core OpenEnv class: reset() / step() / state()
    ├── models.py       Pydantic models: Observation, Action, Reward
    ├── ticket_data.py  13 synthetic tickets with ground-truth labels
    ├── graders.py      Deterministic graders for all 3 tasks
    └── reward.py       Shaped reward function

OpenEnv Compliance

—[x] Typed Pydantic models: TriageObservation, TriageAction, TriageReward
—[x] reset(task_name) → TriageObservation
—[x] step(action) → (observation, reward, done, info)
—[x] state() → full internal state dict
—[x] openenv.yaml with complete metadata
—[x] 3 tasks: easy → medium → hard
—[x] All grader scores in [0.0, 1.0], deterministic
—[x] Non-sparse shaped reward
—[x] inference.py using OpenAI client
—[x] Working Dockerfile
—[x] 22/22 unit tests passing