inference.py

import os
import json
from openai import OpenAI
from email_triage_env.env import EmailTriageEnv 
from email_triage_env.models import Action

# --- 1. MANDATORY ENVIRONMENT VARIABLES ---
API_BASE_URL = os.getenv("API_BASE_URL", "https://api.openai.com/v1")
MODEL_NAME = os.getenv("MODEL_NAME", "gpt-4-turbo-preview")
HF_TOKEN = os.getenv("HF_TOKEN")
LOCAL_IMAGE_NAME = os.getenv("LOCAL_IMAGE_NAME")

# --- 2. CONFIGURE CLIENT ---
client = OpenAI(
    base_url=API_BASE_URL,
    api_key=HF_TOKEN or os.getenv("OPENAI_API_KEY", "dummy-key")
)

def run_baseline():
    env = EmailTriageEnv()
    
    for task_idx in range(len(env.tasks)):
        # Safely get the task name (e.g., 'easy', 'medium')
        task_name = env.tasks[task_idx].get('id', f'task_{task_idx}')
        
        # --- REQUIRED LOG: START ---
        print(f"[START] task={task_name}", flush=True)
        
        obs = env.reset(task_idx)
        done = False
        step_counter = 0
        current_reward = 0.0
        
        messages = [
            {"role": "system", "content": "You are a customer support agent. Goal: " + env.tasks[task_idx]['goal']},
            {"role": "user", "content": f"Initial state: {obs.model_dump_json()}"}
        ]
        
        while not done:
            step_counter += 1
            
            try:
                response = client.chat.completions.create(
                    model=MODEL_NAME, 
                    messages=messages,
                    tools=[{
                        "type": "function",
                        "function": {
                            "name": "take_action",
                            "description": "Perform an action on an email.",
                            "parameters": Action.model_json_schema()
                        }
                    }],
                    tool_choice={"type": "function", "function": {"name": "take_action"}}
                )
                
                tool_call = response.choices[0].message.tool_calls[0]
                action_args = json.loads(tool_call.function.arguments)
                action = Action(**action_args)
                
                obs, reward, done, info = env.step(action)
                current_reward = reward.value
                
                # --- REQUIRED LOG: STEP ---
                print(f"[STEP] step={step_counter} reward={current_reward}", flush=True)
                
                messages.append({"role": "assistant", "tool_calls": [tool_call]})
                messages.append({
                    "role": "tool",
                    "tool_call_id": tool_call.id,
                    "name": "take_action",
                    "content": f"Observation: {obs.model_dump_json()} | Reward: {current_reward} | Done: {done}"
                })
                
            except Exception as e:
                print(f"Caught an unhandled API or parsing exception: {e}", flush=True)
                # If the grader forces an error, break the loop safely
                break 
                
            # Failsafe to prevent infinite loops locally
            if step_counter >= 10:
                break
        
        # --- THE FIX: Clamp the score strictly between 0 and 1 ---
        # If it's 0.0, it becomes 0.01. If it's 1.0, it becomes 0.99.
        final_score = max(0.01, min(0.99, float(current_reward)))
        
        # --- REQUIRED LOG: END ---
        print(f"[END] task={task_name} score={final_score} steps={step_counter}", flush=True)

if __name__ == "__main__":
    run_baseline()