Apppublic

pragyamandal/novamart-customer-service

sourceHugging Faceupdated 6mo agoView on Hugging Face

0likes

App README

NovaMart Customer Service Environment

An OpenEnv-compatible RL environment for training agents to resolve customer service issues under real policy constraints.

Overview

This environment simulates a customer service agent at NovaMart, a fictional electronics retailer. The novel mechanic is constraint negotiation — the agent must find the best possible resolution when the obvious answer is blocked by policy. Unlike existing environments, NovaMart features a ranked solution space where multiple valid resolutions exist with different quality scores. The agent must investigate, reason about policies, and navigate ambiguity to maximize customer satisfaction within the boundaries of what is actually allowed.

Environment Description

—Simulated company: NovaMart sells consumer electronics (phones, laptops, earbuds, smartwatches)
—8 available tools: lookup_order, check_policy, verify_defect, check_loyalty, process_refund, issue_store_credit, process_exchange, escalate_to_manager
—3 policy paths:
—Standard return — 30-day window, full refund, item must be unopened
—Defective item — 90-day window, full refund, defect must be verified
—Loyalty goodwill — 60-day window, store credit, silver or gold tier only

Tasks

Task	Difficulty	Description	Max Steps
`easy_refund`	Easy	Standard return within 30 days, all tools work perfectly	10
`defect_resolution`	Medium	Defect claim outside standard return window, agent must find the defect policy path	10
`loyalty_negotiation`	Hard	Gold tier customer outside all standard windows, ambiguous defect claim, multiple constrained resolution paths	10

Observation Space

Field	Type	Description
`customer_id`	string	Unique identifier for the customer
`customer_name`	string	Full name of the customer
`issue_description`	string	The customer's complaint or request
`order_id`	string	Order being discussed
`days_since_purchase`	integer	Number of days since the order was placed
`customer_tier`	string	Loyalty tier: standard, silver, or gold
`conversation_history`	list[string]	All messages exchanged so far
`available_actions`	list[string]	Actions the agent can take this step
`step_number`	integer	Current step in the episode
`max_steps`	integer	Maximum steps allowed for this task
`tools_used`	list[string]	Tools the agent has already called
`last_tool_result`	dict	Output from the most recent tool call

Action Space

Action Type	Description
`lookup_order`	Retrieve order details by order ID
`check_policy`	Check eligibility under a specific return policy
`verify_defect`	Verify whether a product defect is confirmed
`check_loyalty`	Check customer loyalty tier and goodwill eligibility
`process_refund`	Process a monetary refund for an order
`issue_store_credit`	Issue store credit to the customer's account
`process_exchange`	Exchange the product for a different item
`escalate_to_manager`	Escalate the case to a manager for review
`respond_to_customer`	Send a message to the customer without taking a tool action

Reward Structure

—step_reward: Given at every step, not just at the end of the episode
—resolution_score: How close the agent's resolution is to the optimal outcome (0.0–1.0)
—tool_use_score: Whether the right tool was called with correct parameters
—communication_score: Quality of the customer-facing message
—penalty: Deductions for incorrect actions
—−0.3 for choosing the wrong resolution type
—−0.4 for promising a refund the customer is not eligible for

Setup and Usage

Run with Docker

bash

docker build -t novamart-env .
docker run -p 7860:7860 novamart-env

Run inference

bash

cp .env.example .env
# Edit .env with your API credentials
python inference.py

Environment Variables

Variable	Description
`API_BASE_URL`	LLM API endpoint (OpenAI-compatible)
`MODEL_NAME`	Model identifier
`HF_TOKEN`	Your API key

Baseline Scores

Task	Expected Score Range
`easy_refund`	0.7 – 1.0
`defect_resolution`	0.5 – 0.8
`loyalty_negotiation`	0.3 – 0.6