Upload 7 files
Browse files- Dockerfile +11 -0
- api.py +54 -0
- dataset.py +111 -0
- environment.py +42 -0
- inference.py +102 -0
- openenv.yaml +16 -0
- test_env.py +17 -0
Dockerfile
ADDED
|
@@ -0,0 +1,11 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
FROM python:3.10
|
| 2 |
+
|
| 3 |
+
WORKDIR /app
|
| 4 |
+
|
| 5 |
+
COPY . .
|
| 6 |
+
|
| 7 |
+
RUN pip install fastapi uvicorn pydantic
|
| 8 |
+
|
| 9 |
+
EXPOSE 7860
|
| 10 |
+
|
| 11 |
+
CMD ["python", "-m", "uvicorn", "api:app", "--host", "0.0.0.0", "--port", "7860"]
|
api.py
ADDED
|
@@ -0,0 +1,54 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
from fastapi import FastAPI
|
| 2 |
+
from pydantic import BaseModel
|
| 3 |
+
from environment import SupportEnv
|
| 4 |
+
from dataset import cases
|
| 5 |
+
|
| 6 |
+
app = FastAPI()
|
| 7 |
+
|
| 8 |
+
env = SupportEnv(cases)
|
| 9 |
+
|
| 10 |
+
# -------- Models --------
|
| 11 |
+
class Action(BaseModel):
|
| 12 |
+
message: str
|
| 13 |
+
|
| 14 |
+
# -------- Parser --------
|
| 15 |
+
def parse_action(message):
|
| 16 |
+
message = message.lower()
|
| 17 |
+
|
| 18 |
+
if "refund" in message:
|
| 19 |
+
return "refund"
|
| 20 |
+
elif "replace" in message:
|
| 21 |
+
return "replace"
|
| 22 |
+
elif "reject" in message:
|
| 23 |
+
return "reject"
|
| 24 |
+
elif "proof" in message:
|
| 25 |
+
return "ask_proof"
|
| 26 |
+
else:
|
| 27 |
+
return "unknown"
|
| 28 |
+
|
| 29 |
+
# -------- Routes --------
|
| 30 |
+
@app.post("/reset")
|
| 31 |
+
def reset():
|
| 32 |
+
obs = env.reset()
|
| 33 |
+
|
| 34 |
+
return {
|
| 35 |
+
"observation": {
|
| 36 |
+
"echoed_message": obs
|
| 37 |
+
},
|
| 38 |
+
"reward": 0.0,
|
| 39 |
+
"done": False
|
| 40 |
+
}
|
| 41 |
+
|
| 42 |
+
@app.post("/step")
|
| 43 |
+
def step(action: Action):
|
| 44 |
+
parsed_action = parse_action(action.message)
|
| 45 |
+
|
| 46 |
+
obs, reward, done = env.step(parsed_action)
|
| 47 |
+
|
| 48 |
+
return {
|
| 49 |
+
"observation": {
|
| 50 |
+
"echoed_message": obs
|
| 51 |
+
},
|
| 52 |
+
"reward": reward,
|
| 53 |
+
"done": done
|
| 54 |
+
}
|
dataset.py
ADDED
|
@@ -0,0 +1,111 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
cases = [
|
| 2 |
+
|
| 3 |
+
# =========================
|
| 4 |
+
# 🟢 1. GENUINE REQUESTS
|
| 5 |
+
# =========================
|
| 6 |
+
|
| 7 |
+
{
|
| 8 |
+
"query": "I received a damaged product, want a refund",
|
| 9 |
+
"days": 2,
|
| 10 |
+
"has_proof": True,
|
| 11 |
+
"correct_action": "refund"
|
| 12 |
+
},
|
| 13 |
+
{
|
| 14 |
+
"query": "The item is broken, can I get a replacement?",
|
| 15 |
+
"days": 1,
|
| 16 |
+
"has_proof": True,
|
| 17 |
+
"correct_action": "replace"
|
| 18 |
+
},
|
| 19 |
+
{
|
| 20 |
+
"query": "Wrong item delivered, I ordered something else",
|
| 21 |
+
"days": 3,
|
| 22 |
+
"has_proof": True,
|
| 23 |
+
"correct_action": "replace"
|
| 24 |
+
},
|
| 25 |
+
{
|
| 26 |
+
"query": "Product stopped working within 5 days",
|
| 27 |
+
"days": 5,
|
| 28 |
+
"has_proof": True,
|
| 29 |
+
"correct_action": "refund"
|
| 30 |
+
},
|
| 31 |
+
{
|
| 32 |
+
"query": "I want to return the product, it is defective",
|
| 33 |
+
"days": 6,
|
| 34 |
+
"has_proof": True,
|
| 35 |
+
"correct_action": "refund"
|
| 36 |
+
},
|
| 37 |
+
|
| 38 |
+
# =========================
|
| 39 |
+
# 🔴 2. CUSTOMER LYING
|
| 40 |
+
# =========================
|
| 41 |
+
|
| 42 |
+
{
|
| 43 |
+
"query": "I never received my order but tracking says delivered",
|
| 44 |
+
"days": 4,
|
| 45 |
+
"has_proof": False,
|
| 46 |
+
"delivery_status": "delivered",
|
| 47 |
+
"correct_action": "reject"
|
| 48 |
+
},
|
| 49 |
+
{
|
| 50 |
+
"query": "Product is damaged (no images attached)",
|
| 51 |
+
"days": 3,
|
| 52 |
+
"has_proof": False,
|
| 53 |
+
"correct_action": "ask_proof"
|
| 54 |
+
},
|
| 55 |
+
{
|
| 56 |
+
"query": "I want refund but I already used the product",
|
| 57 |
+
"days": 5,
|
| 58 |
+
"has_proof": False,
|
| 59 |
+
"used_product": True,
|
| 60 |
+
"correct_action": "reject"
|
| 61 |
+
},
|
| 62 |
+
{
|
| 63 |
+
"query": "Claiming wrong item but order history matches delivery",
|
| 64 |
+
"days": 2,
|
| 65 |
+
"has_proof": False,
|
| 66 |
+
"correct_action": "reject"
|
| 67 |
+
},
|
| 68 |
+
{
|
| 69 |
+
"query": "Requesting refund after clearly using the product",
|
| 70 |
+
"days": 4,
|
| 71 |
+
"has_proof": False,
|
| 72 |
+
"used_product": True,
|
| 73 |
+
"correct_action": "reject"
|
| 74 |
+
},
|
| 75 |
+
|
| 76 |
+
# =========================
|
| 77 |
+
# 🟡 3. SUSPICIOUS REQUESTS
|
| 78 |
+
# =========================
|
| 79 |
+
|
| 80 |
+
{
|
| 81 |
+
"query": "Product seems damaged but not sure",
|
| 82 |
+
"days": 2,
|
| 83 |
+
"has_proof": False,
|
| 84 |
+
"correct_action": "ask_proof"
|
| 85 |
+
},
|
| 86 |
+
{
|
| 87 |
+
"query": "I think I got a defective item",
|
| 88 |
+
"days": 3,
|
| 89 |
+
"has_proof": False,
|
| 90 |
+
"correct_action": "ask_proof"
|
| 91 |
+
},
|
| 92 |
+
{
|
| 93 |
+
"query": "Item not working properly, can you help?",
|
| 94 |
+
"days": 4,
|
| 95 |
+
"has_proof": False,
|
| 96 |
+
"correct_action": "ask_proof"
|
| 97 |
+
},
|
| 98 |
+
{
|
| 99 |
+
"query": "Received product but facing issues",
|
| 100 |
+
"days": 5,
|
| 101 |
+
"has_proof": False,
|
| 102 |
+
"correct_action": "ask_proof"
|
| 103 |
+
},
|
| 104 |
+
{
|
| 105 |
+
"query": "The product quality is not good",
|
| 106 |
+
"days": 6,
|
| 107 |
+
"has_proof": False,
|
| 108 |
+
"correct_action": "ask_proof"
|
| 109 |
+
}
|
| 110 |
+
|
| 111 |
+
]
|
environment.py
ADDED
|
@@ -0,0 +1,42 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
import random
|
| 2 |
+
|
| 3 |
+
class SupportEnv:
|
| 4 |
+
def __init__(self, cases):
|
| 5 |
+
self.cases = cases
|
| 6 |
+
self.index = 0
|
| 7 |
+
|
| 8 |
+
def reset(self):
|
| 9 |
+
self.index = 0
|
| 10 |
+
return self._get_obs()
|
| 11 |
+
|
| 12 |
+
def step(self, action):
|
| 13 |
+
case = self.cases[self.index]
|
| 14 |
+
correct = case["correct_action"]
|
| 15 |
+
|
| 16 |
+
# reward logic
|
| 17 |
+
if action == correct:
|
| 18 |
+
reward = 1.0
|
| 19 |
+
elif action == "ask_proof" and correct == "ask_proof":
|
| 20 |
+
reward = 0.8
|
| 21 |
+
else:
|
| 22 |
+
reward = -0.5
|
| 23 |
+
|
| 24 |
+
# move to next case
|
| 25 |
+
self.index += 1
|
| 26 |
+
done = self.index >= len(self.cases)
|
| 27 |
+
|
| 28 |
+
obs = self._get_obs() if not done else "All cases completed."
|
| 29 |
+
|
| 30 |
+
return obs, reward, done
|
| 31 |
+
|
| 32 |
+
def _get_obs(self):
|
| 33 |
+
case = self.cases[self.index]
|
| 34 |
+
|
| 35 |
+
return f"""
|
| 36 |
+
Customer Query: {case['query']}
|
| 37 |
+
Days since delivery: {case['days']}
|
| 38 |
+
Proof provided: {case.get('has_proof', False)}
|
| 39 |
+
|
| 40 |
+
Choose one action:
|
| 41 |
+
refund / replace / reject / ask_proof
|
| 42 |
+
"""
|
inference.py
ADDED
|
@@ -0,0 +1,102 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
import asyncio
|
| 2 |
+
from openai import OpenAI
|
| 3 |
+
|
| 4 |
+
API_BASE_URL = "https://api.openai.com/v1"
|
| 5 |
+
API_KEY = "your_api_key_here"
|
| 6 |
+
MODEL_NAME = "gpt-4o-mini"
|
| 7 |
+
|
| 8 |
+
IMAGE_NAME = "support-env"
|
| 9 |
+
TASK_NAME = "customer-support"
|
| 10 |
+
BENCHMARK = "openenv"
|
| 11 |
+
|
| 12 |
+
MAX_STEPS = 20
|
| 13 |
+
MAX_TOTAL_REWARD = 20
|
| 14 |
+
SUCCESS_SCORE_THRESHOLD = 0.6
|
| 15 |
+
|
| 16 |
+
|
| 17 |
+
def get_model_message(client, step, state, last_reward, history):
|
| 18 |
+
prompt = f"""
|
| 19 |
+
You are a customer support agent.
|
| 20 |
+
|
| 21 |
+
State:
|
| 22 |
+
{state}
|
| 23 |
+
|
| 24 |
+
Choose one action:
|
| 25 |
+
refund / replace / reject / ask_proof
|
| 26 |
+
|
| 27 |
+
Respond with your decision.
|
| 28 |
+
"""
|
| 29 |
+
response = client.chat.completions.create(
|
| 30 |
+
model=MODEL_NAME,
|
| 31 |
+
messages=[{"role": "user", "content": prompt}]
|
| 32 |
+
)
|
| 33 |
+
|
| 34 |
+
return response.choices[0].message.content
|
| 35 |
+
|
| 36 |
+
|
| 37 |
+
def log_start(**kwargs):
|
| 38 |
+
print("[START]", kwargs, flush=True)
|
| 39 |
+
|
| 40 |
+
|
| 41 |
+
def log_step(**kwargs):
|
| 42 |
+
print("[STEP]", kwargs, flush=True)
|
| 43 |
+
|
| 44 |
+
|
| 45 |
+
def log_end(**kwargs):
|
| 46 |
+
print("[END]", kwargs, flush=True)
|
| 47 |
+
|
| 48 |
+
|
| 49 |
+
async def main():
|
| 50 |
+
client = OpenAI(base_url=API_BASE_URL, api_key=API_KEY)
|
| 51 |
+
|
| 52 |
+
from openenv import Env
|
| 53 |
+
env = await Env.from_docker_image(IMAGE_NAME)
|
| 54 |
+
|
| 55 |
+
history = []
|
| 56 |
+
rewards = []
|
| 57 |
+
steps_taken = 0
|
| 58 |
+
score = 0.0
|
| 59 |
+
success = False
|
| 60 |
+
|
| 61 |
+
log_start(task=TASK_NAME, env=BENCHMARK, model=MODEL_NAME)
|
| 62 |
+
|
| 63 |
+
try:
|
| 64 |
+
result = await env.reset()
|
| 65 |
+
last_state = result.observation.echoed_message
|
| 66 |
+
last_reward = 0.0
|
| 67 |
+
|
| 68 |
+
for step in range(1, MAX_STEPS + 1):
|
| 69 |
+
if result.done:
|
| 70 |
+
break
|
| 71 |
+
|
| 72 |
+
message = get_model_message(client, step, last_state, last_reward, history)
|
| 73 |
+
|
| 74 |
+
result = await env.step({"message": message})
|
| 75 |
+
|
| 76 |
+
reward = result.reward or 0.0
|
| 77 |
+
done = result.done
|
| 78 |
+
|
| 79 |
+
rewards.append(reward)
|
| 80 |
+
steps_taken = step
|
| 81 |
+
last_state = result.observation.echoed_message
|
| 82 |
+
last_reward = reward
|
| 83 |
+
|
| 84 |
+
log_step(step=step, action=message, reward=reward, done=done, error=None)
|
| 85 |
+
|
| 86 |
+
history.append(f"{message} -> {reward}")
|
| 87 |
+
|
| 88 |
+
if done:
|
| 89 |
+
break
|
| 90 |
+
|
| 91 |
+
score = sum(rewards) / MAX_TOTAL_REWARD
|
| 92 |
+
score = max(0.0, min(1.0, score))
|
| 93 |
+
success = score >= SUCCESS_SCORE_THRESHOLD
|
| 94 |
+
|
| 95 |
+
finally:
|
| 96 |
+
await env.close()
|
| 97 |
+
log_end(success=success, steps=steps_taken, score=score, rewards=rewards)
|
| 98 |
+
|
| 99 |
+
|
| 100 |
+
if __name__ == "__main__":
|
| 101 |
+
asyncio.run(main())
|
| 102 |
+
|
openenv.yaml
ADDED
|
@@ -0,0 +1,16 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
name: customer-support-env
|
| 2 |
+
description: Simulated e-commerce customer support environment for evaluating AI agents on handling refund, replacement, and fraud detection scenarios.
|
| 3 |
+
|
| 4 |
+
version: 1.0
|
| 5 |
+
|
| 6 |
+
entrypoint:
|
| 7 |
+
type: http
|
| 8 |
+
url: http://localhost:7860
|
| 9 |
+
|
| 10 |
+
tasks:
|
| 11 |
+
- name: easy
|
| 12 |
+
description: Handle straightforward genuine customer requests
|
| 13 |
+
- name: medium
|
| 14 |
+
description: Handle policy-based decisions and edge cases
|
| 15 |
+
- name: hard
|
| 16 |
+
description: Handle suspicious or ambiguous requests requiring clarification
|
test_env.py
ADDED
|
@@ -0,0 +1,17 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
from environment import SupportEnv
|
| 2 |
+
from dataset import cases # your dataset file
|
| 3 |
+
|
| 4 |
+
env = SupportEnv(cases)
|
| 5 |
+
|
| 6 |
+
obs = env.reset()
|
| 7 |
+
done = False
|
| 8 |
+
|
| 9 |
+
while not done:
|
| 10 |
+
print(obs)
|
| 11 |
+
|
| 12 |
+
action = input("Enter action: ") # simulate agent
|
| 13 |
+
|
| 14 |
+
obs, reward, done = env.step(action)
|
| 15 |
+
|
| 16 |
+
print(f"Reward: {reward}")
|
| 17 |
+
print("-" * 40)
|