vertex_api_invoice_extractor.py

import base64
import json
import os
from google.oauth2 import service_account
import vertexai
from remittance_pdf_processing_utils import remittance_logger
from vertexai.generative_models import GenerativeModel, Part
import vertexai.preview.generative_models as generative_models
from remittance_pdf_processing_types import InvoiceNumbers,PaymentAmount
from remittance_pdf_processing_utils import remove_duplicate_lists

# Set up authentication
def initialize_vertexai():
    # Get the base64-encoded service account JSON from an environment variable
    encoded_sa_json = os.environ.get('VERTEX_AI_SERVICE_ACCOUNT_JSON')
    
    if not encoded_sa_json:
        raise ValueError("VERTEX_AI_SERVICE_ACCOUNT_JSON environment variable is not set")
    
    try:
        # Decode the base64 string to get the JSON content
        sa_json_str = base64.b64decode(encoded_sa_json).decode('utf-8')
        sa_info = json.loads(sa_json_str)
        
        # Create credentials object from the decoded JSON
        credentials = service_account.Credentials.from_service_account_info(
            sa_info,
            scopes=['https://www.googleapis.com/auth/cloud-platform']
        )
        
        # Initialize Vertex AI with the credentials
        vertexai.init(project="saltech-ai-sandbox", location="us-central1", credentials=credentials)
        
        print("Vertex AI initialized successfully.")
    except json.JSONDecodeError:
        raise ValueError("Invalid JSON format in the decoded service account information")
    except Exception as e:
        raise Exception(f"Error initializing Vertex AI: {str(e)}")


# Call this function at the start of your script or in your main function
initialize_vertexai()

def extract_invoice_numbers_with_vertex_ai(base64_image: str, multi_hop: bool = False) -> list[InvoiceNumbers]:
	"""
	Dispatches the invoice number extraction to either single-hop or multi-hop method based on the multi_hop parameter.
	
	Args:
	base64_image (str): The base64-encoded image string.
	multi_hop (bool): Whether to use multi-hop processing.
	
	Returns:
	list[InvoiceNumbers]: A list containing lists of extracted invoice numbers.
	"""
	if multi_hop:
		return extract_invoice_numbers_with_vertex_ai_multi_hop(base64_image)
	else:
		return extract_invoice_numbers_with_vertex_ai_single_hop(base64_image)

def extract_invoice_numbers_with_vertex_ai_single_hop(base64_image: str) -> list[InvoiceNumbers]:
	"""
	Extracts invoice numbers from a single base64-encoded image using Google's Gemini Flash model with single-hop processing.
	
	Args:
	base64_image (str): The base64-encoded image string.
	
	Returns:
	list[InvoiceNumbers]: A list containing lists of extracted invoice numbers.
	"""
	vertexai.init(project="saltech-ai-sandbox", location="us-central1")
	model = GenerativeModel("gemini-1.5-flash-001")

	image_part = Part.from_data(
		mime_type="image/png",
		data=base64.b64decode(base64_image),
	)

	text_prompt = """Given the remittance letter image, extract all invoice numbers. 
	Respond with a comma-separated list of invoice numbers only. 
	If no invoice numbers are found, respond with 'No invoice numbers found'."""

	generation_config = {
		"max_output_tokens": 8192,
		"temperature": 0.1,
		"top_p": 0.95,
	}

	safety_settings = {
		generative_models.HarmCategory.HARM_CATEGORY_HATE_SPEECH: generative_models.HarmBlockThreshold.BLOCK_MEDIUM_AND_ABOVE,
		generative_models.HarmCategory.HARM_CATEGORY_DANGEROUS_CONTENT: generative_models.HarmBlockThreshold.BLOCK_MEDIUM_AND_ABOVE,
		generative_models.HarmCategory.HARM_CATEGORY_SEXUALLY_EXPLICIT: generative_models.HarmBlockThreshold.BLOCK_MEDIUM_AND_ABOVE,
		generative_models.HarmCategory.HARM_CATEGORY_HARASSMENT: generative_models.HarmBlockThreshold.BLOCK_MEDIUM_AND_ABOVE,
	}

	responses = model.generate_content(
		[image_part, text_prompt],
		generation_config=generation_config,
		safety_settings=safety_settings,
		stream=True,
	)

	full_response = ""
	for response in responses:
		full_response += response.text

	remittance_logger.debug(f"Extracted invoice numbers (raw model response): {full_response}")

	extracted_numbers = parse_gemini_response(full_response)
	return [extracted_numbers]  # Wrap in a list to match the expected return type

def extract_column_headers(base64_image: str) -> list[str]:
	"""
	Extracts column header names that could contain invoice numbers from a base64-encoded image.

	Args:
	base64_image (str): The base64-encoded image string.

	Returns:
	list[str]: A list of column header names.
	"""
	vertexai.init(project="saltech-ai-sandbox", location="us-central1")
	model = GenerativeModel("gemini-1.5-flash-001")

	image_part = Part.from_data(
	  mime_type="image/png",
	  data=base64.b64decode(base64_image),
	)

	text_prompt = """Given the remittance letter image, extract all column header names that could contain invoice numbers. 
	Respond with a comma-separated list only."""

	generation_config = {
	  "max_output_tokens": 8192,
	  "temperature": 0.1,
	  "top_p": 0.95,
	}

	safety_settings = {
	  generative_models.HarmCategory.HARM_CATEGORY_HATE_SPEECH: generative_models.HarmBlockThreshold.BLOCK_MEDIUM_AND_ABOVE,
	  generative_models.HarmCategory.HARM_CATEGORY_DANGEROUS_CONTENT: generative_models.HarmBlockThreshold.BLOCK_MEDIUM_AND_ABOVE,
	  generative_models.HarmCategory.HARM_CATEGORY_SEXUALLY_EXPLICIT: generative_models.HarmBlockThreshold.BLOCK_MEDIUM_AND_ABOVE,
	  generative_models.HarmCategory.HARM_CATEGORY_HARASSMENT: generative_models.HarmBlockThreshold.BLOCK_MEDIUM_AND_ABOVE,
	}

	responses = model.generate_content(
	  [image_part, text_prompt],
	  generation_config=generation_config,
	  safety_settings=safety_settings,
	  stream=True,
	)

	full_response = ""
	for response in responses:
	  full_response += response.text

	remittance_logger.debug(f"Extracted column headers (raw model response): {full_response}")

	return [header.strip() for header in full_response.split(',')]

def extract_invoice_numbers_for_column(base64_image: str, column_name: str) -> InvoiceNumbers:
	"""
	Extracts invoice numbers from a specific column in a base64-encoded image.

	Args:
	base64_image (str): The base64-encoded image string.
	column_name (str): The name of the column to extract invoice numbers from.

	Returns:
	InvoiceNumbers: A list of extracted invoice numbers for the specified column.
	"""
	vertexai.init(project="saltech-ai-sandbox", location="us-central1")
	model = GenerativeModel("gemini-1.5-flash-001")

	image_part = Part.from_data(
	  mime_type="image/png",
	  data=base64.b64decode(base64_image),
	)

	text_prompt = f"""Given the remittance letter image, extract all invoice numbers from the column "{column_name}". 
	Respond with a comma-separated list only."""

	generation_config = {
	  "max_output_tokens": 8192,
	  "temperature": 0.1,
	  "top_p": 0.95,
	}

	safety_settings = {
	  generative_models.HarmCategory.HARM_CATEGORY_HATE_SPEECH: generative_models.HarmBlockThreshold.BLOCK_MEDIUM_AND_ABOVE,
	  generative_models.HarmCategory.HARM_CATEGORY_DANGEROUS_CONTENT: generative_models.HarmBlockThreshold.BLOCK_MEDIUM_AND_ABOVE,
	  generative_models.HarmCategory.HARM_CATEGORY_SEXUALLY_EXPLICIT: generative_models.HarmBlockThreshold.BLOCK_MEDIUM_AND_ABOVE,
	  generative_models.HarmCategory.HARM_CATEGORY_HARASSMENT: generative_models.HarmBlockThreshold.BLOCK_MEDIUM_AND_ABOVE,
	}

	responses = model.generate_content(
	  [image_part, text_prompt],
	  generation_config=generation_config,
	  safety_settings=safety_settings,
	  stream=True,
	)

	full_response = ""
	for response in responses:
	  full_response += response.text

	remittance_logger.debug(f"Extracted invoice numbers for column '{column_name}' (raw model response): {full_response}")

	return [number.strip() for number in full_response.split(',') if number.strip()]

def extract_invoice_numbers_with_vertex_ai_multi_hop(base64_image: str) -> list[InvoiceNumbers]:
	"""
	Extracts invoice numbers from a single base64-encoded image using Google's Gemini Flash model with multi-hop processing.

	Args:
	base64_image (str): The base64-encoded image string.

	Returns:
	list[InvoiceNumbers]: A list containing lists of extracted invoice numbers for each processed column.
	"""
	# First hop: Extract column headers
	column_headers = extract_column_headers(base64_image)
	remittance_logger.debug(f"Extracted column headers: {column_headers}")

	# Second hop: Extract invoice numbers for each column (up to 3 columns)
	all_invoice_numbers = []
	for column_name in column_headers[:3]:
		invoice_numbers = extract_invoice_numbers_for_column(base64_image, column_name)
		remittance_logger.debug(f"Extracted invoice numbers for column '{column_name}': {invoice_numbers}")
		if invoice_numbers:  # Only add non-empty lists
			all_invoice_numbers.append(invoice_numbers)

	# Remove duplicate lists using the utility function
	unique_invoice_numbers = remove_duplicate_lists(all_invoice_numbers)
	return unique_invoice_numbers

# def extract_invoice_numbers_from_text_with_vertex_ai(text: str, multi_hop: bool = False) -> list[InvoiceNumbers]:
# 	"""
# 	Extracts invoice numbers from text using Google's Gemini Flash model.
	
# 	Args:
# 	text (str): The text of the remittance letter.
# 	multi_hop (bool): Whether to use multi-hop processing (not implemented yet).
	
# 	Returns:
# 	list[InvoiceNumbers]: A list containing lists of extracted invoice numbers.
# 	"""
# 	vertexai.init(project="saltech-ai-sandbox", location="us-central1")
# 	model = GenerativeModel("gemini-1.5-flash-001")

# 	prompt = f"""Given the following remittance letter text, extract all invoice numbers. 
# 	Respond with a comma-separated list of invoice numbers only. 
# 	If no invoice numbers are found, respond with 'No invoice numbers found'.

# 	Remittance letter text:
# 	{text}
# 	"""

# 	generation_config = {
# 		"max_output_tokens": 8192,
# 		"temperature": 0.1,
# 		"top_p": 0.95,
# 	}

# 	safety_settings = {
# 		generative_models.HarmCategory.HARM_CATEGORY_HATE_SPEECH: generative_models.HarmBlockThreshold.BLOCK_MEDIUM_AND_ABOVE,
# 		generative_models.HarmCategory.HARM_CATEGORY_DANGEROUS_CONTENT: generative_models.HarmBlockThreshold.BLOCK_MEDIUM_AND_ABOVE,
# 		generative_models.HarmCategory.HARM_CATEGORY_SEXUALLY_EXPLICIT: generative_models.HarmBlockThreshold.BLOCK_MEDIUM_AND_ABOVE,
# 		generative_models.HarmCategory.HARM_CATEGORY_HARASSMENT: generative_models.HarmBlockThreshold.BLOCK_MEDIUM_AND_ABOVE,
# 	}

# 	responses = model.generate_content(
# 		prompt,
# 		generation_config=generation_config,
# 		safety_settings=safety_settings,
# 		stream=True,
# 	)

# 	full_response = ""
# 	for response in responses:
# 		full_response += response.text
	 
# 	remittance_logger.debug(f"Vertex AI invoice numbers full response: {full_response}")

# 	extracted_numbers = parse_gemini_response(full_response)
# 	return [extracted_numbers]  # Wrap in a list to match the expected return type

def parse_gemini_response(response: str) -> list[str]:
	"""
	Parses the response from Gemini Flash model and extracts invoice numbers.
	
	Args:
	response (str): The response string from Gemini Flash model.
	
	Returns:
	list[str]: A list of extracted invoice numbers.
	"""
	if response.strip().lower().startswith('no invoice numbers found'):
		return []
	
	# Split the comma-separated list and strip whitespace from each number
	invoice_numbers = [num.strip() for num in response.split(',')]
	return invoice_numbers

# Note: You'll need to set up authentication for Google Cloud.
# Typically, you'd set the GOOGLE_APPLICATION_CREDENTIALS environment variable
# to point to your service account key file.


def extract_invoice_numbers_from_text_with_vertex_ai(text: str, multi_hop: bool = False) -> list[InvoiceNumbers]:
	"""
	Dispatches the invoice number extraction to either single-hop or multi-hop method based on the multi_hop parameter.
	
	Args:
	text (str): The text of the remittance letter.
	multi_hop (bool): Whether to use multi-hop processing.
	
	Returns:
	list[InvoiceNumbers]: A list containing lists of extracted invoice numbers.
	"""
	if multi_hop:
		return extract_invoice_numbers_from_text_with_vertex_ai_multi_hop(text)
	else:
		return extract_invoice_numbers_from_text_with_vertex_ai_single_hop(text)

def extract_invoice_numbers_from_text_with_vertex_ai_single_hop(text: str) -> list[InvoiceNumbers]:
	"""
	Extracts invoice numbers from text using Google's Gemini Flash model with single-hop processing.
	
	Args:
	text (str): The text of the remittance letter.
	
	Returns:
	list[InvoiceNumbers]: A list containing lists of extracted invoice numbers.
	"""
	vertexai.init(project="saltech-ai-sandbox", location="us-central1")
	model = GenerativeModel("gemini-1.5-flash-001")

	prompt = f"""Given the following remittance letter text, extract all invoice numbers. 
	Respond with a comma-separated list of invoice numbers only. 
	If no invoice numbers are found, respond with 'No invoice numbers found'.

	Remittance letter text:
	{text}
	"""

	generation_config = {
		"max_output_tokens": 8192,
		"temperature": 0.1,
		"top_p": 0.95,
	}

	safety_settings = {
		generative_models.HarmCategory.HARM_CATEGORY_HATE_SPEECH: generative_models.HarmBlockThreshold.BLOCK_MEDIUM_AND_ABOVE,
		generative_models.HarmCategory.HARM_CATEGORY_DANGEROUS_CONTENT: generative_models.HarmBlockThreshold.BLOCK_MEDIUM_AND_ABOVE,
		generative_models.HarmCategory.HARM_CATEGORY_SEXUALLY_EXPLICIT: generative_models.HarmBlockThreshold.BLOCK_MEDIUM_AND_ABOVE,
		generative_models.HarmCategory.HARM_CATEGORY_HARASSMENT: generative_models.HarmBlockThreshold.BLOCK_MEDIUM_AND_ABOVE,
	}

	responses = model.generate_content(
		prompt,
		generation_config=generation_config,
		safety_settings=safety_settings,
		stream=True,
	)

	full_response = ""
	for response in responses:
		full_response += response.text
	 
	remittance_logger.debug(f"Vertex AI invoice numbers full response (single-hop): {full_response}")

	extracted_numbers = parse_gemini_response(full_response)
	return [extracted_numbers]  # Wrap in a list to match the expected return type

def extract_invoice_numbers_from_text_with_vertex_ai_multi_hop(text: str) -> list[InvoiceNumbers]:
	"""
	Extracts invoice numbers from text using Google's Gemini Flash model with multi-hop processing.

	Args:
	text (str): The text of the remittance letter.

	Returns:
	list[InvoiceNumbers]: A list containing lists of extracted invoice numbers for each processed column.
	"""
	# First hop: Extract column headers
	column_headers = extract_column_headers_from_text(text)
	remittance_logger.debug(f"Extracted column headers: {column_headers}")

	# Second hop: Extract invoice numbers for each column (up to 3 columns)
	all_invoice_numbers = []
	for column_name in column_headers[:3]:
		invoice_numbers = extract_invoice_numbers_for_column_from_text(text, column_name)
		remittance_logger.debug(f"Extracted invoice numbers for column '{column_name}': {invoice_numbers}")
		if invoice_numbers:  # Only add non-empty lists
			all_invoice_numbers.append(invoice_numbers)

	# Remove duplicate lists using the utility function
	unique_invoice_numbers = remove_duplicate_lists(all_invoice_numbers)
	return unique_invoice_numbers

def extract_column_headers_from_text(text: str) -> list[str]:
	"""
	Extracts column header names that could contain invoice numbers from the text.

	Args:
	text (str): The text of the remittance letter.

	Returns:
	list[str]: A list of column header names.
	"""
	vertexai.init(project="saltech-ai-sandbox", location="us-central1")
	model = GenerativeModel("gemini-1.5-flash-001")

	prompt = f"""Given the following remittance letter text, extract all column header names or section titles that could contain invoice numbers. 
	Respond with a comma-separated list only.

	Remittance letter text:
	{text}
	"""

	generation_config = {
		"max_output_tokens": 8192,
		"temperature": 0.1,
		"top_p": 0.95,
	}

	safety_settings = {
		generative_models.HarmCategory.HARM_CATEGORY_HATE_SPEECH: generative_models.HarmBlockThreshold.BLOCK_MEDIUM_AND_ABOVE,
		generative_models.HarmCategory.HARM_CATEGORY_DANGEROUS_CONTENT: generative_models.HarmBlockThreshold.BLOCK_MEDIUM_AND_ABOVE,
		generative_models.HarmCategory.HARM_CATEGORY_SEXUALLY_EXPLICIT: generative_models.HarmBlockThreshold.BLOCK_MEDIUM_AND_ABOVE,
		generative_models.HarmCategory.HARM_CATEGORY_HARASSMENT: generative_models.HarmBlockThreshold.BLOCK_MEDIUM_AND_ABOVE,
	}

	response = model.generate_content(
		prompt,
		generation_config=generation_config,
		safety_settings=safety_settings,
	)

	remittance_logger.debug(f"Extracted column headers (raw model response): {response.text}")

	return [header.strip() for header in response.text.split(',')]

def extract_invoice_numbers_for_column_from_text(text: str, column_name: str) -> InvoiceNumbers:
	"""
	Extracts invoice numbers from a specific column or section in the text.

	Args:
	text (str): The text of the remittance letter.
	column_name (str): The name of the column or section to extract invoice numbers from.

	Returns:
	InvoiceNumbers: A list of extracted invoice numbers for the specified column.
	"""
	vertexai.init(project="saltech-ai-sandbox", location="us-central1")
	model = GenerativeModel("gemini-1.5-flash-001")

	prompt = f"""Given the following remittance letter text, extract all invoice numbers from the column or section "{column_name}". 
	Respond with a comma-separated list only. If no invoice numbers are found, respond with 'No invoice numbers found'.

	Remittance letter text:
	{text}
	"""

	generation_config = {
		"max_output_tokens": 8192,
		"temperature": 0.1,
		"top_p": 0.95,
	}

	safety_settings = {
		generative_models.HarmCategory.HARM_CATEGORY_HATE_SPEECH: generative_models.HarmBlockThreshold.BLOCK_MEDIUM_AND_ABOVE,
		generative_models.HarmCategory.HARM_CATEGORY_DANGEROUS_CONTENT: generative_models.HarmBlockThreshold.BLOCK_MEDIUM_AND_ABOVE,
		generative_models.HarmCategory.HARM_CATEGORY_SEXUALLY_EXPLICIT: generative_models.HarmBlockThreshold.BLOCK_MEDIUM_AND_ABOVE,
		generative_models.HarmCategory.HARM_CATEGORY_HARASSMENT: generative_models.HarmBlockThreshold.BLOCK_MEDIUM_AND_ABOVE,
	}

	response = model.generate_content(
		prompt,
		generation_config=generation_config,
		safety_settings=safety_settings,
	)

	remittance_logger.debug(f"Extracted invoice numbers for column '{column_name}' (raw model response): {response.text}")

	return parse_gemini_response(response.text)

def extract_payment_amounts_with_vertex_ai(base64_image: str) -> list[PaymentAmount]:
	vertexai.init(project="saltech-ai-sandbox", location="us-central1")
	model = GenerativeModel("gemini-1.5-flash-001")

	image_part = Part.from_data(
		mime_type="image/png",
		data=base64.b64decode(base64_image),
	)

	text_prompt = """Given the remittance letter image, extract the total payment amount. 
	Respond with the payment amount only. 
	If no payment amounts are found, respond with 'No payment amounts found'."""

	generation_config = {
		"max_output_tokens": 256,
		"temperature": 0.1,
		"top_p": 0.95,
	}

	safety_settings = {
		generative_models.HarmCategory.HARM_CATEGORY_HATE_SPEECH: generative_models.HarmBlockThreshold.BLOCK_MEDIUM_AND_ABOVE,
		generative_models.HarmCategory.HARM_CATEGORY_DANGEROUS_CONTENT: generative_models.HarmBlockThreshold.BLOCK_MEDIUM_AND_ABOVE,
		generative_models.HarmCategory.HARM_CATEGORY_SEXUALLY_EXPLICIT: generative_models.HarmBlockThreshold.BLOCK_MEDIUM_AND_ABOVE,
		generative_models.HarmCategory.HARM_CATEGORY_HARASSMENT: generative_models.HarmBlockThreshold.BLOCK_MEDIUM_AND_ABOVE,
	}

	responses = model.generate_content(
		[image_part, text_prompt],
		generation_config=generation_config,
		safety_settings=safety_settings,
		stream=True,
	)

	full_response = ""
	for response in responses:
		full_response += response.text

	remittance_logger.debug(f"Vertex AI payment amount full response: {full_response}")

	extracted_amounts = parse_gemini_payment_response(full_response)
	return extracted_amounts

def extract_payment_amounts_from_text_with_vertex_ai(text: str) -> list[PaymentAmount]:
	"""
	Extracts payment amounts from text using Google's Gemini Flash model.
	
	Args:
	text (str): The text of the remittance letter.
	
	Returns:
	list[PaymentAmount]: A list of extracted payment amounts.
	"""
	vertexai.init(project="saltech-ai-sandbox", location="us-central1")
	model = GenerativeModel("gemini-1.5-flash-001")

	prompt = f"""Given the following remittance letter text, extract the total payment amount. 
	Respond with the payment amount only. 
	If no payment amounts are found, respond with 'No payment amounts found'.

	Remittance letter text:
	{text}
	"""

	generation_config = {
		"max_output_tokens": 256,
		"temperature": 0.1,
		"top_p": 0.95,
	}

	safety_settings = {
		generative_models.HarmCategory.HARM_CATEGORY_HATE_SPEECH: generative_models.HarmBlockThreshold.BLOCK_MEDIUM_AND_ABOVE,
		generative_models.HarmCategory.HARM_CATEGORY_DANGEROUS_CONTENT: generative_models.HarmBlockThreshold.BLOCK_MEDIUM_AND_ABOVE,
		generative_models.HarmCategory.HARM_CATEGORY_SEXUALLY_EXPLICIT: generative_models.HarmBlockThreshold.BLOCK_MEDIUM_AND_ABOVE,
		generative_models.HarmCategory.HARM_CATEGORY_HARASSMENT: generative_models.HarmBlockThreshold.BLOCK_MEDIUM_AND_ABOVE,
	}

	response = model.generate_content(
		prompt,
		generation_config=generation_config,
		safety_settings=safety_settings,
	)

	remittance_logger.debug(f"Vertex AI payment amount full response: {response.text}")

	extracted_amounts = parse_gemini_payment_response(response.text)
	return extracted_amounts

def parse_gemini_payment_response(response: str) -> list[PaymentAmount]:
	"""
	Parses the response from Gemini Flash model and extracts payment amounts.
	
	Args:
	response (str): The response string from Gemini Flash model.
	
	Returns:
	list[PaymentAmount]: A list of one extracted payment amount (or empty).
	"""
	if response.strip().lower() == 'no payment amounts found':
		return []
	
	payment_amounts = [response.strip()]
	return payment_amounts

def extract_payment_amounts_from_base64_images(base64_images: list[str]) -> list[PaymentAmount]:
	# Implementation similar to extract_invoice_numbers_with_vertex_ai
	# but focused on extracting payment amounts
	return []