TanujInsane commited on
Commit ·
e263f9c
1
Parent(s): 676fd8d
fix: add LLM classification via API_BASE_URL and API_KEY
Browse files- inference.py +70 -11
inference.py
CHANGED
|
@@ -3,17 +3,27 @@ import traceback
|
|
| 3 |
import sys
|
| 4 |
import os
|
| 5 |
import time
|
|
|
|
|
|
|
|
|
|
| 6 |
|
| 7 |
BASE_URL = os.environ.get("ENV_URL", "http://127.0.0.1:7860")
|
| 8 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 9 |
def wait_for_server():
|
| 10 |
"""Retries GET /reset to ensure server is ready before proceeding."""
|
| 11 |
print(f"Waiting for server at {BASE_URL}...")
|
| 12 |
for i in range(20):
|
| 13 |
try:
|
| 14 |
-
# Perform GET /reset as requested
|
| 15 |
resp = requests.get(f"{BASE_URL}/reset", timeout=5)
|
| 16 |
-
# If the server responds (even with 405 Method Not Allowed), it's online
|
| 17 |
if resp.status_code < 500:
|
| 18 |
print("Server online.")
|
| 19 |
return True
|
|
@@ -26,6 +36,43 @@ def wait_for_server():
|
|
| 26 |
print("Warning: Server did not respond after 20 retries. Proceeding anyway...")
|
| 27 |
return False
|
| 28 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 29 |
def main():
|
| 30 |
print("[START] task=DocumentClassification-v1", flush=True)
|
| 31 |
|
|
@@ -33,23 +80,37 @@ def main():
|
|
| 33 |
wait_for_server()
|
| 34 |
|
| 35 |
try:
|
| 36 |
-
# Reset the environment
|
| 37 |
reset_resp = requests.post(f"{BASE_URL}/reset", json={}, timeout=60)
|
| 38 |
reset_resp.raise_for_status()
|
| 39 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 40 |
|
| 41 |
total_reward = 0.0
|
| 42 |
num_steps = 5
|
| 43 |
|
| 44 |
for step in range(1, num_steps + 1):
|
| 45 |
-
#
|
| 46 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
| 47 |
step_resp.raise_for_status()
|
| 48 |
-
|
| 49 |
|
| 50 |
-
reward = float(
|
| 51 |
total_reward += reward
|
| 52 |
-
done =
|
|
|
|
|
|
|
|
|
|
| 53 |
|
| 54 |
print(f"[STEP] step={step} reward={reward}", flush=True)
|
| 55 |
|
|
@@ -60,10 +121,8 @@ def main():
|
|
| 60 |
print(f"[END] task=DocumentClassification-v1 score={score:.4f} steps={num_steps}", flush=True)
|
| 61 |
|
| 62 |
except Exception as e:
|
| 63 |
-
# On failure, ensure we print the END tag with zero score to satisfy the grader
|
| 64 |
print(f"[END] task=DocumentClassification-v1 score=0.0 steps=0", flush=True)
|
| 65 |
traceback.print_exc()
|
| 66 |
-
# sys.exit(1) removed - returning gracefully
|
| 67 |
|
| 68 |
if __name__ == "__main__":
|
| 69 |
main()
|
|
|
|
| 3 |
import sys
|
| 4 |
import os
|
| 5 |
import time
|
| 6 |
+
import re
|
| 7 |
+
from openai import OpenAI
|
| 8 |
+
from constants import CATEGORIES
|
| 9 |
|
| 10 |
BASE_URL = os.environ.get("ENV_URL", "http://127.0.0.1:7860")
|
| 11 |
|
| 12 |
+
# LLM API configuration from environment variables (Grader injected)
|
| 13 |
+
API_BASE_URL = os.environ.get("API_BASE_URL")
|
| 14 |
+
API_KEY = os.environ.get("API_KEY")
|
| 15 |
+
|
| 16 |
+
# Initialize LLM client
|
| 17 |
+
client = None
|
| 18 |
+
if API_BASE_URL and API_KEY:
|
| 19 |
+
client = OpenAI(base_url=API_BASE_URL, api_key=API_KEY)
|
| 20 |
+
|
| 21 |
def wait_for_server():
|
| 22 |
"""Retries GET /reset to ensure server is ready before proceeding."""
|
| 23 |
print(f"Waiting for server at {BASE_URL}...")
|
| 24 |
for i in range(20):
|
| 25 |
try:
|
|
|
|
| 26 |
resp = requests.get(f"{BASE_URL}/reset", timeout=5)
|
|
|
|
| 27 |
if resp.status_code < 500:
|
| 28 |
print("Server online.")
|
| 29 |
return True
|
|
|
|
| 36 |
print("Warning: Server did not respond after 20 retries. Proceeding anyway...")
|
| 37 |
return False
|
| 38 |
|
| 39 |
+
def get_llm_prediction(content, difficulty):
|
| 40 |
+
"""Uses LLM to classify document content based on available categories for the given difficulty."""
|
| 41 |
+
if not client:
|
| 42 |
+
return 0
|
| 43 |
+
|
| 44 |
+
# Get categories for the current task difficulty from constants
|
| 45 |
+
difficulty_cats = CATEGORIES.get(difficulty, CATEGORIES["easy"])
|
| 46 |
+
category_list = "\n".join([f"- {idx}: {name}" for idx, name in difficulty_cats.items()])
|
| 47 |
+
|
| 48 |
+
prompt = f"""You are a document classification expert.
|
| 49 |
+
Classify the following customer support document into exactly one of these categories:
|
| 50 |
+
{category_list}
|
| 51 |
+
|
| 52 |
+
Document Content:
|
| 53 |
+
{content}
|
| 54 |
+
|
| 55 |
+
Respond ONLY with the category index (integer). Do not include any other text."""
|
| 56 |
+
|
| 57 |
+
try:
|
| 58 |
+
# LiteLLM proxy typically handles the specific model routing
|
| 59 |
+
response = client.chat.completions.create(
|
| 60 |
+
model="gpt-4o",
|
| 61 |
+
messages=[{"role": "user", "content": prompt}],
|
| 62 |
+
temperature=0,
|
| 63 |
+
max_tokens=10
|
| 64 |
+
)
|
| 65 |
+
prediction = response.choices[0].message.content.strip()
|
| 66 |
+
|
| 67 |
+
# Robust parsing to extract the integer index
|
| 68 |
+
match = re.search(r'\d+', prediction)
|
| 69 |
+
if match:
|
| 70 |
+
return int(match.group())
|
| 71 |
+
except Exception as e:
|
| 72 |
+
print(f"LLM Reasoning Error: {e}")
|
| 73 |
+
|
| 74 |
+
return 0
|
| 75 |
+
|
| 76 |
def main():
|
| 77 |
print("[START] task=DocumentClassification-v1", flush=True)
|
| 78 |
|
|
|
|
| 80 |
wait_for_server()
|
| 81 |
|
| 82 |
try:
|
| 83 |
+
# Reset the environment and get initial observation
|
| 84 |
reset_resp = requests.post(f"{BASE_URL}/reset", json={}, timeout=60)
|
| 85 |
reset_resp.raise_for_status()
|
| 86 |
+
init_data = reset_resp.json()
|
| 87 |
+
current_obs = init_data.get("observation", {})
|
| 88 |
+
|
| 89 |
+
# Check environment state to determine difficulty/categories
|
| 90 |
+
state_resp = requests.get(f"{BASE_URL}/state", timeout=10)
|
| 91 |
+
state_resp.raise_for_status()
|
| 92 |
+
state_info = state_resp.json()
|
| 93 |
+
difficulty = state_info.get("difficulty", "easy")
|
| 94 |
|
| 95 |
total_reward = 0.0
|
| 96 |
num_steps = 5
|
| 97 |
|
| 98 |
for step in range(1, num_steps + 1):
|
| 99 |
+
# Use LLM to predict the action based on document content
|
| 100 |
+
content = current_obs.get("content", "")
|
| 101 |
+
action = get_llm_prediction(content, difficulty)
|
| 102 |
+
|
| 103 |
+
# Execute step with predicted action
|
| 104 |
+
step_resp = requests.post(f"{BASE_URL}/step", json={"action": action}, timeout=60)
|
| 105 |
step_resp.raise_for_status()
|
| 106 |
+
step_result = step_resp.json()
|
| 107 |
|
| 108 |
+
reward = float(step_result.get("reward", 0.0))
|
| 109 |
total_reward += reward
|
| 110 |
+
done = step_result.get("done", False)
|
| 111 |
+
|
| 112 |
+
# Update the observation for the next iteration
|
| 113 |
+
current_obs = step_result.get("observation", {})
|
| 114 |
|
| 115 |
print(f"[STEP] step={step} reward={reward}", flush=True)
|
| 116 |
|
|
|
|
| 121 |
print(f"[END] task=DocumentClassification-v1 score={score:.4f} steps={num_steps}", flush=True)
|
| 122 |
|
| 123 |
except Exception as e:
|
|
|
|
| 124 |
print(f"[END] task=DocumentClassification-v1 score=0.0 steps=0", flush=True)
|
| 125 |
traceback.print_exc()
|
|
|
|
| 126 |
|
| 127 |
if __name__ == "__main__":
|
| 128 |
main()
|