content_moderation_models/openai_moderation.py

## Expects a csv passed as command line argument with columns 'prompt' and 'response' and outputs csv with ratings appended. 
## Define Open AI Key in key.env file as cgpt_key=" " or hardcode in script.
import openai
import pandas as pd
import os
import time
from datetime import datetime
from dotenv import load_dotenv
import sys

# Load environment variables (Ensure your API key is in key.env)
load_dotenv("key.env")
api_key = os.getenv("cgpt_key")  # Get API key

# Initialize OpenAI client
client = openai.OpenAI(api_key=api_key)

# Input & output file paths >> Set desired output name here
INPUT_CSV = sys.argv[1]
OUTPUT_CSV = "openai_librai_moderated_output.csv"

# Function to check moderation
def moderate_text(text):
    try:
        start_time = time.time()  # Start timing
        response = client.moderations.create(input=text, model="omni-moderation-latest") 
        end_time = time.time()  # End timing

        moderation_result = response.results[0]

        return {
            "Flagged": moderation_result.flagged,
            "Hate": moderation_result.categories.hate,
            "Hate/Threatening": moderation_result.categories.hate_threatening,
            "Harassment": moderation_result.categories.harassment,
            "Harassment/Threatening": moderation_result.categories.harassment_threatening,
            "Self-Harm": moderation_result.categories.self_harm,
            "Self-Harm/Intent": moderation_result.categories.self_harm_intent,
            "Self-Harm/Instructions": moderation_result.categories.self_harm_instructions,
            "Violence": moderation_result.categories.violence,
            "Violence/Graphic": moderation_result.categories.violence_graphic,
            "Sexual": moderation_result.categories.sexual,
            "Sexual/Minors": moderation_result.categories.sexual_minors,
            "Illicit": moderation_result.categories.illicit,  
            "Illicit/Violent": moderation_result.categories.illicit_violent,
            "Processing_Time (s)": round(end_time - start_time, 4),  
            "Timestamp": datetime.utcnow().strftime("%Y-%m-%d %H:%M:%S"), 
            "Model": response.model,  
            "Response_ID": response.id  
        }
    except Exception as e:
        print(f"Error processing text: {text} -> {e}")
        return None  # Return None if an error occurs

# Load dataset
df = pd.read_csv(INPUT_CSV)

# Create an output CSV
if not os.path.exists(OUTPUT_CSV):
    columns = ["prompt", "response", "Flagged", "Hate", "Hate/Threatening", 
               "Harassment", "Harassment/Threatening", "Self-Harm", "Self-Harm/Intent", 
               "Self-Harm/Instructions", "Violence", "Violence/Graphic", "Sexual", 
               "Sexual/Minors", "Illicit", "Illicit/Violent", 
               "Processing_Time (s)", "Timestamp", "Model", "Response_ID"]
    pd.DataFrame(columns=columns).to_csv(OUTPUT_CSV, index=False)

# Process each row
for index, row in df.iterrows():
    prompt = row["prompt"]
    response = row["response"]

    # Moderate the response
    moderation_results = moderate_text(response)

    if moderation_results:
        row_data = {
            "prompt": prompt,
            "response": response,
            **moderation_results  # Expands the dictionary into individual columns
        }
        
        # Append to CSV
        pd.DataFrame([row_data]).to_csv(OUTPUT_CSV, mode="a", header=False, index=False)

        # Print progress
        print(f"Processed row {index+1}/{len(df)} - Flagged: {moderation_results['Flagged']} - Time: {moderation_results['Processing_Time (s)']}s")

print(f"✅ Moderation complete. Results saved to {OUTPUT_CSV}")