README.md

metadata

title: Ecomm OpenEnv
emoji: 🤖
colorFrom: blue
colorTo: green
sdk: docker
app_port: 7860
pinned: false

E-Commerce Customer Support Decision Environment (OpenEnv)

Overview

This project is an OpenEnv environment for testing how an agent handles customer support decisions in an e-commerce setting.

The idea is to simulate cases where the decision isn’t obvious — things like missing proof, suspicious return patterns, or unclear customer messages. Instead of just picking the “correct” label, the agent has to make a reasonable call based on the situation.

---

Problem

Most customer support decisions involve trade-offs rather than clear rules.

Customer support isn’t clean or structured in real life. Some examples:

• product reported as damaged but no images
• customer has a lot of past returns (possible abuse)
• request is technically valid but context is messy
• message itself is vague

This environment tries to reflect that kind of ambiguity and see how an agent deals with it.

---

Actions

The agent selects one of the following actions:

approve_refund
approve_replacement
reject
request_info

---

Observations

Each case is turned into a support ticket (plain text). It includes:

• customer message
• product
• days since delivery
• whether images were provided
• return history

So the agent has to read and interpret instead of just using structured fields directly.

---

Reward logic

The reward isn’t just strict right/wrong.

• correct decision → 1.0
• reasonable alternative → 0.5
• wrong → 0.0
• bad decisions → negative penalties

There are also some overrides:

• high return history → lean towards reject
• damaged claim without images → ask for more info

Final score from /grader is always between 0 and 1.

---

Tasks

Dataset is split into:

easy → obvious cases
medium → needs some reasoning
hard → ambiguous, possible fraud, missing info

---

API

GET /reset (also supports POST)
POST /step
GET /state
GET /tasks
GET /grader
GET /baseline
GET /health

---

Running locally (Docker)

docker build -t openenv .
docker run -p 7860:7860 openenv

API will be at:

http://localhost:7860

---

Inference

inference.py is the required evaluation script.

It:

• uses OpenAI client via API_BASE_URL, MODEL_NAME, HF_TOKEN
• interacts with /reset and /step
• outputs logs in [START], [STEP], [END] format

---

Baseline

baseline.py uses OpenAI to choose actions from the ticket text.

Run it like this:

export OPENAI_API_KEY=your_key
export ENV_URL=http://localhost:7860

python baseline.py

Expected score is roughly:

Average score: 0.45 (20 episodes, fallback agent)
Note: OpenAI-based baseline expected to perform higher

---

Structure

data/
easy.json
medium.json
hard.json

env/
app.py

baseline.py
inference.py
openenv.yaml
Dockerfile
README.md

---

Possible extensions

• multi-step conversations
• stricter policy constraints
• larger dataset

---

Notes

This environment is designed to stay simple while still capturing realistic decision-making scenarios. The goal is not perfect accuracy, but reasonable behavior under uncertainty.