main.py

from fastapi import FastAPI, Request, BackgroundTasks
import json
import io
from openai import Client
from supabase import create_client
from typing import List, Dict, Any
import asyncio
import logging
from datetime import datetime
import os
import tiktoken

# Initialize logging
logging.basicConfig(level=logging.INFO)
logger = logging.getLogger(__name__)

app = FastAPI()
client = Client(api_key=os.getenv('OPENAI_API_KEY'),organization=os.getenv('ORG_ID'))
url: str = os.getenv('SUPABASE_URL')
key: str = os.getenv('SUPABASE_KEY')
supabase: Client = create_client(url, key)


@app.post("/send/batch_processing")
async def testv1(request: Request, background_tasks: BackgroundTasks):
    try:
        body_data = await request.json()

        print(body_data)
        # Create initial batch job record
        save_data = {
            'batch_job_id': f"batch_{datetime.utcnow().strftime('%Y%m%d_%H%M%S')}",
            "batch_job_status": False,
            "created_at": datetime.utcnow().isoformat()
        }
        
        response = (
            supabase.table("batch_processing_details")
            .insert(save_data)
            .execute()
        )
        
        # Add processing to background tasks
        background_tasks.add_task(process_batch_job, body_data, save_data['batch_job_id'])

        return {'data': 'Batch job is scheduled!'}
        
        # return {'data': 'Batch job is scheduled!', 'batch_job_id': save_data['batch_job_id']}
           
        
    except Exception as e:
        return {'error': str(e)}


def receipt_radar_prompt(raw_text:str)->str:
    insurance_response_structure = """
    {
      "insurance_type": "Classify it into 8 categories travel , health , term , vehicle, property,liability, life , buisness only .Try to find the closest possible based on the receipt text, if you don't understand the type classify it as others.",
      "policy_details": {
        "policyholder_name": "",
        "policy_number": "",
        "insurance_start_date": "",
        "insurance_end_date": "",
        "premium_amount": "",
        "payment_frequency": ""
      },
      "coverage_details": {
        "covered_items": {
          "item_type": "",
          "product_company": "",
          "product_model": "",
          "product_manufacturing_year": ""
        },
        "comprehensive_coverage_type_policy": "yes/no"
      }
    }
    """
    travel_response_structure = """
        travel_type(bus,train,airplane,taxi,bike,rickshaw classify in these categories only strictly),travel_company_name , departure_destination , arrival_destination , arrival_city(if you are not able to find the arrival city add the arrival destination into this field strictly. ), departure_date,arrival_date .If the arrival and departure dates are the same from receipt text given to you analyse it properly to check that, then only use the same date in both the fields .if you don't find any field mark it as null.
    """
    hotel_data_points = """ hotel_type(hotel_stay , dine_in , dine_in + stay(use both keyword strictly)), hotel_brand_name , hotel_location , hotel_checkin_date , hotel_checkout_date. if you don't find any field mark it as null """
    system_prompt = f"""Extract information from the following receipt OCR text and return a JSON object with these exact keys: brand, total_cost, location, purchase_category, brand_category, Date, currency, filename, payment_method, metadata.
    Rules:
    1. For total_cost, use the highest monetary value in the text.
    2. For brand_category, choose the closest match from: ["Fashion and Apparel", "Jewelry and Watches", "Beauty and Personal Care", "Automobiles", "Real Estate", "Travel(it may contain reciepts of airlines , trains , taxi ,cruise ,etc)", "Hospitality(it will include reciepts of Hotels (stays) , restaurants , cafe's , bar's , Accommodation Services , Beverages Services (don't include food delivery service in hospitality))","Food Delivery Services(like swiggy , zomato,eatsure and any other you can analyse from receipt text)", "Home and Lifestyle", "Technology and Electronics", "Sports and Leisure", "Art and Collectibles", "Health and Wellness", "Stationery and Writing Instruments", "Children and Baby", "Pet Accessories", "Financial Services", "Insurance"]
    3. Format Date as dd-mm-yyyy.Strictly return the date in the format dd-mm-yyyy.
    4. metadata: For insurance receipts extract the data points given in the JSON and return the JSON with structure: \n """ + insurance_response_structure + """
    5.metadata : For travel receipts(flight ,bus,train) extract these data points as a JSON object exactly""" + travel_response_structure + """
    6. metadata : For hotel receipts extract these data points as a JSON object exactly""" + hotel_data_points + f"""
    For non-insurance and non-travel , non-hotel receipts, return metadata as null.
    4. Use currency codes (e.g., USD, EUR) instead of symbols.
    5. Generate filename as 'PURCHASE_TYPE_BRAND_DATE' (e.g., 'clothing_gucci_20230715').
    6. If a value is not found, return null.
    7. If all values are null, return null.
    Ensure the strictly that output is a valid JSON object containing strictly the above keys, without any explanations.
    Here's the OCR text below analyse it and convert into json using keys provided in first line and using the rules provided in rules section:
    Generate a JSON response in the following format without using the ```json block. Ensure the output is properly formatted as plain text JSON.
    {raw_text}
    """
    return system_prompt

def adjust_prompt_tokens_v1(prompt: str) -> str:
    max_tokens = 127500
    encoding = tiktoken.encoding_for_model(LLM_MODEL)
    tokenized_prompt = encoding.encode(prompt)
    
    # If token count exceeds max_tokens, trim it from the end while keeping full words
    if len(tokenized_prompt) > max_tokens:
        # Find the maximum index for the tokens that keeps the length within max_tokens
        trimmed_tokens = tokenized_prompt[:max_tokens]
        
        # Decode the trimmed tokens back to text
        trimmed_text = encoding.decode(trimmed_tokens)
        
        # Ensure we don't end up with a partial word; trim back to the last full word
        last_space = trimmed_text.rfind(' ')
        if last_space != -1:
            trimmed_text = trimmed_text[:last_space]

    else:
        # If within the limit, no trimming needed
        trimmed_text = prompt

    return trimmed_text


async def process_batch_job(dataset: Dict[str, Any], batch_job_id: str):
    """
    Background task to process the batch job
    """
    try:
        logger.info(f"Starting batch processing for job {batch_job_id}")
        
        openai_tasks = []
        for ds in dataset.get('data'):
            message_id = ds.get('message_id')
            user_id = ds.get('user_id')
            receipt_text = ds.get('receipt_text')
            email = ds.get('email')

            text = adjust_prompt_tokens_v1(receipt_radar_prompt(receipt_text))
            
            task = {
                "custom_id": f"{message_id}-{user_id}-{email}",
                "method": "POST",
                "url": "/v1/chat/completions",
                "body": {
                    "model": "gpt-4o-mini",
                    "temperature": 0.1,
                    "response_format": { 
                        "type": "json_object"
                    },
                    "messages": [
                        {
                            "role": "user",
                            "content": text
                        }
                    ]
                }
            }
            openai_tasks.append(task)
        
        # Create batch file
        json_obj = io.BytesIO()
        for obj in openai_tasks:
            json_obj.write((json.dumps(obj) + '\n').encode('utf-8'))
        
        batch_file = client.files.create(
            file=json_obj,
            purpose="batch"
        )
        
        # Create batch job
        batch_job = client.batches.create(
            input_file_id=batch_file.id,
            endpoint="/v1/chat/completions",
            completion_window="24h"
        )
        
        # Update status in Supabase
        supabase.table("batch_processing_details").insert({
            "batch_job_status": True,
            "completed_at": datetime.utcnow().isoformat()
        }).execute()
        
        logger.info(f"Batch job {batch_job_id} processed successfully")
        
    except Exception as e:
        logger.error(f"Error processing batch job {batch_job_id}: {str(e)}")
        # Update status with error
        supabase.table("batch_processing_details").insert({
            "batch_job_status": False,
            "error": str(e),
            "completed_at": datetime.utcnow().isoformat()
        }).execute()