fixed API calls
Browse files- inference.py +36 -24
inference.py
CHANGED
|
@@ -2,13 +2,13 @@ import asyncio
|
|
| 2 |
import os
|
| 3 |
import requests
|
| 4 |
import time
|
|
|
|
| 5 |
|
| 6 |
-
#
|
| 7 |
-
API_BASE_URL = os.
|
| 8 |
-
|
| 9 |
-
|
| 10 |
|
| 11 |
-
# CONFIG
|
| 12 |
SPACE_URL = "https://arrowman123-customer-supp-env.hf.space"
|
| 13 |
|
| 14 |
TASK_NAME = "customer-support"
|
|
@@ -19,21 +19,30 @@ MAX_TOTAL_REWARD = 20
|
|
| 19 |
SUCCESS_SCORE_THRESHOLD = 0.6
|
| 20 |
|
| 21 |
|
| 22 |
-
#
|
| 23 |
-
def get_model_message(state):
|
| 24 |
-
|
|
|
|
| 25 |
|
| 26 |
-
|
| 27 |
-
|
| 28 |
-
elif "wrong item" in state:
|
| 29 |
-
return "replace"
|
| 30 |
-
elif "not sure" in state or "seems" in state:
|
| 31 |
-
return "ask_proof"
|
| 32 |
-
else:
|
| 33 |
-
return "reject"
|
| 34 |
|
|
|
|
|
|
|
|
|
|
| 35 |
|
| 36 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 37 |
def safe_post(url, json=None, retries=3):
|
| 38 |
for i in range(retries):
|
| 39 |
try:
|
|
@@ -44,29 +53,32 @@ def safe_post(url, json=None, retries=3):
|
|
| 44 |
raise Exception("Failed after retries")
|
| 45 |
|
| 46 |
|
| 47 |
-
# LOGGING
|
| 48 |
def log_start(**kwargs):
|
| 49 |
print("[START]", kwargs, flush=True)
|
| 50 |
|
| 51 |
-
|
| 52 |
def log_step(**kwargs):
|
| 53 |
print("[STEP]", kwargs, flush=True)
|
| 54 |
|
| 55 |
-
|
| 56 |
def log_end(**kwargs):
|
| 57 |
print("[END]", kwargs, flush=True)
|
| 58 |
|
| 59 |
|
| 60 |
-
#
|
| 61 |
async def main():
|
| 62 |
rewards = []
|
| 63 |
steps_taken = 0
|
| 64 |
success = False
|
| 65 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 66 |
log_start(task=TASK_NAME, env=BENCHMARK, model=MODEL_NAME)
|
| 67 |
|
| 68 |
try:
|
| 69 |
-
# RESET
|
| 70 |
res = safe_post(f"{SPACE_URL}/reset")
|
| 71 |
result = res.json()
|
| 72 |
|
|
@@ -77,7 +89,8 @@ async def main():
|
|
| 77 |
if done:
|
| 78 |
break
|
| 79 |
|
| 80 |
-
|
|
|
|
| 81 |
|
| 82 |
res = safe_post(
|
| 83 |
f"{SPACE_URL}/step",
|
|
@@ -116,6 +129,5 @@ async def main():
|
|
| 116 |
)
|
| 117 |
|
| 118 |
|
| 119 |
-
# ENTRY
|
| 120 |
if __name__ == "__main__":
|
| 121 |
asyncio.run(main())
|
|
|
|
| 2 |
import os
|
| 3 |
import requests
|
| 4 |
import time
|
| 5 |
+
from openai import OpenAI
|
| 6 |
|
| 7 |
+
# REQUIRED VARIABLES
|
| 8 |
+
API_BASE_URL = os.environ["API_BASE_URL"]
|
| 9 |
+
API_KEY = os.environ["API_KEY"]
|
| 10 |
+
MODEL_NAME = os.getenv("MODEL_NAME", "gpt-4o-mini")
|
| 11 |
|
|
|
|
| 12 |
SPACE_URL = "https://arrowman123-customer-supp-env.hf.space"
|
| 13 |
|
| 14 |
TASK_NAME = "customer-support"
|
|
|
|
| 19 |
SUCCESS_SCORE_THRESHOLD = 0.6
|
| 20 |
|
| 21 |
|
| 22 |
+
# LLM AGENT
|
| 23 |
+
def get_model_message(client, state):
|
| 24 |
+
prompt = f"""
|
| 25 |
+
You are a customer support agent.
|
| 26 |
|
| 27 |
+
State:
|
| 28 |
+
{state}
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 29 |
|
| 30 |
+
Choose ONLY ONE word:
|
| 31 |
+
refund / replace / reject / ask_proof
|
| 32 |
+
"""
|
| 33 |
|
| 34 |
+
try:
|
| 35 |
+
response = client.chat.completions.create(
|
| 36 |
+
model=MODEL_NAME,
|
| 37 |
+
messages=[{"role": "user", "content": prompt}]
|
| 38 |
+
)
|
| 39 |
+
return response.choices[0].message.content.strip().lower()
|
| 40 |
+
except Exception as e:
|
| 41 |
+
print(f"[ERROR] LLM call failed: {e}")
|
| 42 |
+
return "reject" # fallback
|
| 43 |
+
|
| 44 |
+
|
| 45 |
+
# SAFE REQUEST
|
| 46 |
def safe_post(url, json=None, retries=3):
|
| 47 |
for i in range(retries):
|
| 48 |
try:
|
|
|
|
| 53 |
raise Exception("Failed after retries")
|
| 54 |
|
| 55 |
|
| 56 |
+
# LOGGING
|
| 57 |
def log_start(**kwargs):
|
| 58 |
print("[START]", kwargs, flush=True)
|
| 59 |
|
|
|
|
| 60 |
def log_step(**kwargs):
|
| 61 |
print("[STEP]", kwargs, flush=True)
|
| 62 |
|
|
|
|
| 63 |
def log_end(**kwargs):
|
| 64 |
print("[END]", kwargs, flush=True)
|
| 65 |
|
| 66 |
|
| 67 |
+
# MAIN
|
| 68 |
async def main():
|
| 69 |
rewards = []
|
| 70 |
steps_taken = 0
|
| 71 |
success = False
|
| 72 |
|
| 73 |
+
# Initialize client using THEIR proxy
|
| 74 |
+
client = OpenAI(
|
| 75 |
+
base_url=API_BASE_URL,
|
| 76 |
+
api_key=API_KEY
|
| 77 |
+
)
|
| 78 |
+
|
| 79 |
log_start(task=TASK_NAME, env=BENCHMARK, model=MODEL_NAME)
|
| 80 |
|
| 81 |
try:
|
|
|
|
| 82 |
res = safe_post(f"{SPACE_URL}/reset")
|
| 83 |
result = res.json()
|
| 84 |
|
|
|
|
| 89 |
if done:
|
| 90 |
break
|
| 91 |
|
| 92 |
+
# USE LLM
|
| 93 |
+
message = get_model_message(client, state)
|
| 94 |
|
| 95 |
res = safe_post(
|
| 96 |
f"{SPACE_URL}/step",
|
|
|
|
| 129 |
)
|
| 130 |
|
| 131 |
|
|
|
|
| 132 |
if __name__ == "__main__":
|
| 133 |
asyncio.run(main())
|