import pdfplumber import re import json from typing import Dict, Optional from pathlib import Path # For LLM - supports both Anthropic and OpenAI from anthropic import Anthropic from openai import OpenAI OPENAI_URL = "https://api.deepinfra.com/v1/openai" OPENAI_MODEL = "Qwen/Qwen3-Next-80B-A3B-Instruct" class TenderLocationExtractor: """Extract delivery location from GeM tender PDFs""" def __init__(self, api_key: str, use_anthropic: bool = True): """ Initialize with API key for LLM service Args: api_key: API key for OpenAI use_anthropic: If True, use Anthropic Claude; if False, use OpenAI """ self.use_anthropic = use_anthropic if use_anthropic: self.client = Anthropic(api_key=api_key) else: self.client = OpenAI(api_key=api_key, base_url=OPENAI_URL) def extract_text_from_pdf(self, pdf_path: str) -> str: """ Extract text from PDF file using pdfplumber Args: pdf_path: Path to the PDF file Returns: Extracted text as string """ import os # Check if file exists and is not empty if not os.path.exists(pdf_path): raise Exception(f"PDF file does not exist at path: {pdf_path}") file_size = os.path.getsize(pdf_path) if file_size == 0: raise Exception(f"PDF file is empty (0 bytes): {pdf_path}") try: text = "" with pdfplumber.open(pdf_path) as pdf: for page in pdf.pages: page_text = page.extract_text() if page_text: text += page_text + "\n" return text.strip() except Exception as e: error_msg = str(e) if "No /Root object" in error_msg or "not a PDF" in error_msg: raise Exception(f"Invalid or corrupted PDF file (size: {file_size} bytes): {error_msg}") raise Exception(f"Error extracting text from PDF: {error_msg}") def extract_location_with_llm(self, text: str) -> Dict[str, Optional[str]]: """ Use LLM to extract location information from tender text Args: text: Extracted text from PDF Returns: Dictionary with city, state, and full_address """ prompt = f"""Analyze the following GeM (Government e-Marketplace) tender document text and extract the delivery location information. Focus on finding: 1. The consignee address (परेषती/Consignee address) 2. Office location (कार्यालय का नाम/Office Name) 3. Any delivery address mentioned From this information, identify: - City name - State name (if city is mentioned but not state, infer the state from the city) - Full delivery address Tender Document Text: {text} Please respond in JSON format with the following structure: {{ "city": "city name or null", "state": "full state name or null", "full_address": "complete delivery address or null", "confidence": "high/medium/low" }} Important: - For city names, map them to their respective states (e.g., Kanchipuram -> Tamil Nadu, Mumbai -> Maharashtra, Kadapa -> Andhra Pradesh) - Use full state names, not abbreviations (e.g., "Andhra Pradesh" not "AP") - Name of the state has to be one out of 28 states or 8 UTs. That is one out of: Andhra Pradesh, Arunachal Pradesh, Assam, Bihar, Chhattisgarh, Goa, Gujarat, Haryana, Himachal Pradesh, Jharkhand, Karnataka, Kerala, Madhya Pradesh, Maharashtra, Manipur, Meghalaya, Mizoram, Nagaland, Odisha, Punjab, Rajasthan, Sikkim, Tamil Nadu, Telangana, Tripura, Uttar Pradesh, Uttarakhand, and West Bengal. Andaman and Nicobar Islands, Chandigarh, Dadra and Nagar Haveli and Daman and Diu, Delhi, Jammu and Kashmir, Ladakh, Lakshadweep, and Puducherry. - If multiple addresses exist, prioritize the consignee/delivery address - Return null for fields you cannot confidently extract""" if self.use_anthropic: response = self.client.messages.create( model="claude-sonnet-4-20250514", max_tokens=1000, messages=[ {"role": "user", "content": prompt} ] ) result_text = response.content[0].text else: print("\nInvoking LLM with OpenAI model:", OPENAI_MODEL) response = self.client.chat.completions.create( model=OPENAI_MODEL, messages=[ {"role": "system", "content": "You are a helpful assistant that extracts location (Indian State) information from tender documents."}, {"role": "user", "content": prompt} ], response_format={"type": "json_object"} ) result_text = response.choices[0].message.content print("\nLLM response is:", result_text) # Parse JSON response try: # Clean the response - remove markdown code blocks if present result_text = re.sub(r'```json\s*', '', result_text) result_text = re.sub(r'```\s*$', '', result_text) location_data = json.loads(result_text.strip()) return location_data except json.JSONDecodeError as e: print(f"Error parsing LLM response: {e}") print(f"Raw response: {result_text}") return { "city": None, "state": None, "full_address": None, "confidence": "low", "error": "Failed to parse LLM response" } def extract_location_regex_fallback(self, text: str) -> Dict[str, Optional[str]]: """ Fallback method using regex patterns to extract location Args: text: Extracted text from PDF Returns: Dictionary with extracted location information """ # Common Indian states states = [ 'Andhra Pradesh', 'Arunachal Pradesh', 'Assam', 'Bihar', 'Chhattisgarh', 'Goa', 'Gujarat', 'Haryana', 'Himachal Pradesh', 'Jharkhand', 'Karnataka', 'Kerala', 'Madhya Pradesh', 'Maharashtra', 'Manipur', 'Meghalaya', 'Mizoram', 'Nagaland', 'Odisha', 'Punjab', 'Rajasthan', 'Sikkim', 'Tamil Nadu', 'Telangana', 'Tripura', 'Uttar Pradesh', 'Uttarakhand', 'West Bengal' ] # Look for consignee address pattern address_pattern = r'(?:पता|Address|परेषती)[:\s]*([^\n]{20,200})' address_matches = re.findall(address_pattern, text, re.IGNORECASE) full_address = address_matches[0] if address_matches else None # Find state state = None for s in states: if re.search(r'\b' + re.escape(s) + r'\b', text, re.IGNORECASE): state = s break return { "city": None, "state": state, "full_address": full_address, "confidence": "low", "method": "regex_fallback" } def process_tender_pdf(self, pdf_path: str, use_fallback: bool = False) -> Dict: """ Main method to process tender PDF and extract location Args: pdf_path: Path to the PDF file use_fallback: If True, use regex fallback instead of LLM Returns: Dictionary with extracted information """ print(f"Processing: {pdf_path}") # Extract text text = self.extract_text_from_pdf(pdf_path) print(f"Extracted {len(text)} characters from PDF") # Trim text to 400 chars after last "पता/Address" --- search_str = "पपतताा/Address" idx = text.rfind(search_str) if idx != -1: start_idx = idx + len(search_str) text = text[start_idx:start_idx + 400] print(f"Trimmed text to 400 characters after last '{search_str}'") else: print(f"'{search_str}' not found, using full text") # Extract location if use_fallback: location_info = self.extract_location_regex_fallback(text) else: location_info = self.extract_location_with_llm(text) # Add metadata result = { "file_name": Path(pdf_path).name, "file_path": pdf_path, **location_info } return result