Alejandro-STC's picture
Update comments
6f94dc6 verified
raw
history blame
11.9 kB
import os
#DSPY
import dspy
from dspy import Prediction
from dspy.evaluate import Evaluate
from dspy import Prediction
from dspy.teleprompt import BootstrapFewShot
from dspy.teleprompt import BootstrapFewShotWithRandomSearch
# Data handling
# import pandas as pd
# Calculations and formatting
import re
from decimal import Decimal
# UI
import gradio as gr
from gradio_pdf import PDF
# PDF handling
import pdfplumber
pdf_examples_dir = './pdfexamples/'
model = dspy.OpenAI(
model='gpt-3.5-turbo-0125',
api_key=os.getenv('OPENAI_PROJECT_KEY'),
max_tokens=2000,
temperature=0.01)
dspy.settings.configure(lm=model)
# Utils
def parse_CSV_string(csv_string):
# Parses a CSV string into a list
return list(map(str.strip, csv_string.split(',')))
def parse_CSV_string_to_unique(csv_string):
# Parses a CSV string into a unique list
return list(set(map(str.lower, map(str.strip, csv_string.split(',')))))
def parse_list_of_CSV_strings(list_of_csv_strings):
# Parses a list of CSV strings with invoice numbers into a list of lists
parsed_csv_list = []
for csv_string in list_of_csv_strings:
parsed_csv_list.append(parse_CSV_string_to_unique(csv_string))
return parsed_csv_list
def parse_column_names(s):
"""
Parse a comma-separated list of column names from a string.
Removes the prefix string before splitting the string.
Args:
s: raw response from the model, comma-separated list of column names (string)
Returns:
list of column names (list of strings)
"""
prefix = 'Column Header Names:'
prefix_length = len(prefix)
# r_index = s.rfind(prefix)
# s = s[r_index+prefix_length:] if r_index != -1 else s
if s.strip().lower().startswith(prefix.lower()):
s = s[prefix_length:]
return list(map(str.strip,s.split(',')))
def remove_duplicate_lists(lists):
"""
Remove duplicate lists from a list of lists.
Args:
lists:
a list of lists of strings
Returns:
a list of lists of strings, where each list is unique
"""
seen = set()
unique_lists = []
for lst in lists:
sorted_list = tuple(sorted(lst))
if sorted_list not in seen:
seen.add(sorted_list)
unique_lists.append(lst)
return unique_lists
def parse_invoice_number(s):
# Return the invoice number in a specific format if found, otherwise just return the input string
rp = r'^\s*?([\S\d]+\d{6})'
m = re.search(rp, s)
return m.group(1) if m else s
def standardize_number(s):
# Find the last occurrence of a comma or period
last_separator_index = max(s.rfind(','), s.rfind('.'))
if last_separator_index != -1:
# Split the string into two parts
before_separator = s[:last_separator_index]
after_separator = s[last_separator_index+1:]
# Clean the first part of any commas, periods, or whitespace
before_separator_cleaned = re.sub(r'[.,\s]', '', before_separator)
# Ensure the decimal part starts with a period, even if it was a comma
standardized_s = before_separator_cleaned + '.' + after_separator
else:
# If there's no separator, just remove commas, periods, or whitespace
standardized_s = re.sub(r'[.,\s]', '', s)
return standardized_s
def remove_chars_after_last_digit(s):
# Remove any non-digit characters following the last digit in the string
return re.sub(r'(?<=\d)[^\d]*$', '', s)
def clean_text(s):
# This pattern looks for:
# - Optional non-digit or non-negative sign characters followed by whitespace (if any)
# - Followed by any characters until a digit is found in the word
# It then replaces this matched portion with the remaining part of the word from the first digit
# cleaned_s = re.sub(r'\S*?\s*?(\S*\d\S*)', r'\1', s)
cleaned_s = re.sub(r'[^\d-]*\s?(\S*\d\S*)', r'\1', s)
return cleaned_s
def format_text_decimal(text_decimal):
# Run functions to format a text decimal
return clean_text(remove_chars_after_last_digit(standardize_number(text_decimal.strip().lower())))
# PDF handling
def extract_text_using_pdfplumber(file_path):
# TODO: add check for text vs image PDF
with pdfplumber.open(file_path) as pdf:
extracted_text = ''
for i, page in enumerate(pdf.pages):
# Remove duplicate characters from the page
deduped_page = page.dedupe_chars(tolerance=1)
extracted_text += deduped_page.extract_text()
return extracted_text
def get_PDF_examples(directory):
example_pdf_files = []
for filename in os.listdir(directory):
if filename.endswith('.pdf'):
example_pdf_files.append(os.path.join(directory, filename))
return example_pdf_files
# Signatures and Models
class FindInvoiceNumberColumns(dspy.Signature):
"""Given an input remittance letter, return a list of column header names that may contain invoice numbers."""
content = dspy.InputField(desc="remittance letter", format=lambda s:s) # s:s so it doesn't skip the new lines
column_header_names = dspy.OutputField(desc="comma-separated list of column header names that may contain "\
"invoice numbers")
class InvoiceColumnHeaders(dspy.Module):
"""
Predict the column headers containing invoice numbers from the remittance letter.
Attributes:
response_parser: a function that takes a string and returns a list of strings.
"""
def __init__(self, response_parser=parse_CSV_string):
super().__init__()
self.response_parser = response_parser
self.potential_invoice_column_headers = dspy.Predict(FindInvoiceNumberColumns)
def forward(self, file_content):
prediction = self.potential_invoice_column_headers(content=file_content)
# Remove duplicates from the prediction
unique_headers = list(set(self.response_parser(prediction.column_header_names)))
# Create a new Prediction object with the unique headers
return Prediction(column_header_names=unique_headers)
class FindInvoiceList(dspy.Signature):
"""Given an input remittance letter and a column header name output a comma-separated list of all invoice numbers """\
"""that belong to that column."""
content = dspy.InputField(desc="remittance letter", format=lambda s:s) # s:s so it doesn't skip the new lines
invoice_column_header = dspy.InputField(desc="invoice column header name")
candidate_invoice_numbers = dspy.OutputField(desc="comma-separated list of invoice numbers")
class InvoiceList(dspy.Module):
"""
Retrieves a list of list of potential invoice numbers from a remittance letter.
Attributes:
response_parser: A function that takes a string and returns a list of invoice numbers.
Returns:
A Prediction object with the following fields:
candidate_invoice_numbers: A list of lists of invoice numbers.
"""
def __init__(self, response_parser=parse_CSV_string_to_unique):
super().__init__()
self.response_parser = response_parser
self.find_invoice_headers = InvoiceColumnHeaders(response_parser=parse_column_names) # here we could load a compiled program also
self.find_invoice_numbers = dspy.Predict(FindInvoiceList)
def forward(self, file_content):
predict_column_headers = self.find_invoice_headers(file_content=file_content)
potential_invoice_column_headers = predict_column_headers.column_header_names
candidates = []
for header in potential_invoice_column_headers:
prediction = self.find_invoice_numbers(content=file_content, invoice_column_header=header)
invoice_number_list = self.response_parser(prediction.candidate_invoice_numbers)
candidates.append(invoice_number_list)
# Remove duplicates
candidates = remove_duplicate_lists(candidates)
return Prediction(candidate_invoice_numbers=candidates)
class FindTotalAmountColumns(dspy.Signature):
"""Given an input remittance letter, return a list of column header names that may contain the total payment amount."""
content = dspy.InputField(desc="remittance letter", format=lambda s:s) # s:s so it doesn't skip the new lines
total_column_header_names = dspy.OutputField(desc="comma-separated list of column header names that may contain "\
"the remittance letter total payment amount")
class TotalAmountColumnHeaders(dspy.Module):
def __init__(self):
super().__init__()
self.potential_total_amount_column_headers = dspy.Predict(FindTotalAmountColumns)
def forward(self, file_content):
prediction = self.potential_total_amount_column_headers(content=file_content)
return prediction
class FindTotalAmount(dspy.Signature):
"""Given an input remittance letter and a column header name output the total payment amount """\
"""that belongs to that column."""
content = dspy.InputField(desc="remittance letter", format=lambda s:s) # s:s so it doesn't skip the new lines
total_amount_column_header = dspy.InputField(desc="total amount header name")
total_amount = dspy.OutputField(desc="total payment amount")
class RemittanceLetterTotalAmount(dspy.Module):
def __init__(self):
super().__init__()
self.find_total_amount_header = TotalAmountColumnHeaders()
self.find_total_amount = dspy.Predict(FindTotalAmount)
def forward(self, file_content):
# Predict column headers (returns a Prediction with a CSV string in "column_header_names")
predict_column_headers = self.find_total_amount_header(file_content=file_content)
# Parse CSV into a list
potential_total_amount_column_headers = parse_CSV_string_to_unique(predict_column_headers.total_column_header_names)
potential_total_amounts = []
for header in potential_total_amount_column_headers:
prediction = self.find_total_amount(content=file_content, total_amount_column_header=header)
potential_total_amounts.append(prediction.total_amount)
# Remove duplicates
potential_total_amounts = list(set(potential_total_amounts))
return Prediction(candidate_total_amounts=potential_total_amounts)
# Pipeline
def poc_production_pipeline_without_verification(file_content):
# Get invoice candidates
invoice_list_baseline = InvoiceList()
candidate_invoices = invoice_list_baseline(file_content=file_content).candidate_invoice_numbers
candidate_invoices = [",".join(lst) for lst in candidate_invoices]
# Get total amount candidates
total_amount_baseline = RemittanceLetterTotalAmount()
# Format all decimals
candidate_total_amounts = list(map(format_text_decimal,
total_amount_baseline(file_content=file_content).candidate_total_amounts))
# Only keep unique amounts
candidate_total_amounts = list(set(candidate_total_amounts))
# For UI visualisation purposes, create a list of tuples where the second tuple value is empty
candidate_invoices_for_UI = []
candidate_total_amounts_for_UI = []
for candidate in candidate_invoices:
candidate_invoices_for_UI.append((candidate,))
for candidate in candidate_total_amounts:
candidate_total_amounts_for_UI.append((candidate,))
return candidate_invoices_for_UI, candidate_total_amounts_for_UI
def poc_production_pipeline_without_verification_from_PDF(file_path):
file_content = extract_text_using_pdfplumber(file_path)
return poc_production_pipeline_without_verification(file_content)
# Main app
fake_PDF_examples = get_PDF_examples(pdf_examples_dir)
remittance_letter_demo_without_verification_from_PDF = gr.Interface(
poc_production_pipeline_without_verification_from_PDF,
[PDF(label="Remittance advice", height=1000)],
[
gr.Dataframe(col_count=(1, 'fixed'), label="", headers=["Retrieved invoice proposals"], wrap=True),
gr.Dataframe(col_count=(1, 'fixed'), label="", headers=["Retrieved total amount proposals"], wrap=True)
],
examples=fake_PDF_examples,
allow_flagging='never'
)
remittance_letter_demo_without_verification_from_PDF.launch()