import base64 import re from tempfile import TemporaryDirectory from math import atan, cos, sin from typing import Dict, Optional, Tuple from xml.etree import ElementTree as ET from xml.etree.ElementTree import Element import numpy as np import PyPDF2 from PyPDF2 import PdfFileMerger from doctr.io import DocumentFile from doctr.models import ocr_predictor from PIL import Image from reportlab.lib.colors import black from reportlab.lib.units import inch from reportlab.lib.utils import ImageReader from reportlab.pdfgen.canvas import Canvas class HocrParser(): def __init__(self): self.box_pattern = re.compile(r'bbox((\s+\d+){4})') self.baseline_pattern = re.compile(r'baseline((\s+[\d\.\-]+){2})') def _element_coordinates(self, element: Element) -> Dict: """ Returns a tuple containing the coordinates of the bounding box around an element """ out = out = {'x1': 0, 'y1': 0, 'x2': 0, 'y2': 0} if 'title' in element.attrib: matches = self.box_pattern.search(element.attrib['title']) if matches: coords = matches.group(1).split() out = {'x1': int(coords[0]), 'y1': int( coords[1]), 'x2': int(coords[2]), 'y2': int(coords[3])} return out def _get_baseline(self, element: Element) -> Tuple[float, float]: """ Returns a tuple containing the baseline slope and intercept. """ if 'title' in element.attrib: matches = self.baseline_pattern.search( element.attrib['title']).group(1).split() if matches: return float(matches[0]), float(matches[1]) return (0.0, 0.0) def _pt_from_pixel(self, pxl: Dict, dpi: int) -> Dict: """ Returns the quantity in PDF units (pt) given quantity in pixels """ pt = [(c / dpi * inch) for c in pxl.values()] return {'x1': pt[0], 'y1': pt[1], 'x2': pt[2], 'y2': pt[3]} def _get_element_text(self, element: Element) -> str: """ Return the textual content of the element and its children """ text = '' if element.text is not None: text += element.text for child in element: text += self._get_element_text(child) if element.tail is not None: text += element.tail return text def export_pdfa(self, out_filename: str, hocr: ET.ElementTree, image: Optional[np.ndarray] = None, fontname: str = "Times-Roman", fontsize: int = 12, invisible_text: bool = True, add_spaces: bool = True, dpi: int = 300): """ Generates a PDF/A document from a hOCR document. """ width, height = None, None # Get the image dimensions for div in hocr.findall(".//div[@class='ocr_page']"): coords = self._element_coordinates(div) pt_coords = self._pt_from_pixel(coords, dpi) width, height = pt_coords['x2'] - \ pt_coords['x1'], pt_coords['y2'] - pt_coords['y1'] # after catch break loop break if width is None or height is None: raise ValueError("Could not determine page size") pdf = Canvas(out_filename, pagesize=(width, height), pageCompression=1) span_elements = [element for element in hocr.iterfind(".//span")] for line in span_elements: if 'class' in line.attrib and line.attrib['class'] == 'ocr_line' and line is not None: # get information from xml pxl_line_coords = self._element_coordinates(line) line_box = self._pt_from_pixel(pxl_line_coords, dpi) # compute baseline slope, pxl_intercept = self._get_baseline(line) if abs(slope) < 0.005: slope = 0.0 angle = atan(slope) cos_a, sin_a = cos(angle), sin(angle) intercept = pxl_intercept / dpi * inch baseline_y2 = height - (line_box['y2'] + intercept) # configure options text = pdf.beginText() text.setFont(fontname, fontsize) pdf.setFillColor(black) if invisible_text: text.setTextRenderMode(3) # invisible text # transform overlayed text text.setTextTransform( cos_a, -sin_a, sin_a, cos_a, line_box['x1'], baseline_y2) elements = line.findall(".//span[@class='ocrx_word']") for elem in elements: elemtxt = self._get_element_text(elem).strip() # replace unsupported characters elemtxt = elemtxt.translate(str.maketrans( {'ff': 'ff', 'ffi': 'f‌f‌i', 'ffl': 'f‌f‌l', 'fi': 'fi', 'fl': 'fl'})) if not elemtxt: continue # compute string width pxl_coords = self._element_coordinates(elem) box = self._pt_from_pixel(pxl_coords, dpi) if add_spaces: elemtxt += ' ' box_width = box['x2'] + pdf.stringWidth(elemtxt, fontname, fontsize) - box['x1'] else: box_width = box['x2'] - box['x1'] font_width = pdf.stringWidth(elemtxt, fontname, fontsize) # Adjust relative position of cursor cursor = text.getStartOfLine() dx = box['x1'] - cursor[0] dy = baseline_y2 - cursor[1] text.moveCursor(dx, dy) # suppress text if it is 0 units wide if font_width > 0: text.setHorizScale(100 * box_width / font_width) text.textOut(elemtxt) pdf.drawText(text) # overlay image if provided if image is not None: pdf.drawImage(ImageReader(Image.fromarray(image)), 0, 0, width=width, height=height) pdf.save() from langchain_huggingface import HuggingFaceEmbeddings from transformers import AutoModel, AutoTokenizer, AutoModelForCausalLM from langchain_community.vectorstores import Chroma from langchain.schema import Document from langchain.text_splitter import RecursiveCharacterTextSplitter from langchain_community.llms.huggingface_pipeline import HuggingFacePipeline import torch embedding_model_name = 'l3cube-pune/punjabi-sentence-similarity-sbert' model_kwargs = {'device':'cpu',"trust_remote_code": True} embeddings = HuggingFaceEmbeddings( model_name=embedding_model_name, model_kwargs=model_kwargs ) vectorstore = None def read_file(data: str) -> Document: f = open(data,'r') content = f.read() f.close() doc = Document(page_content=content, metadata={"name": data.split('/')[-1]}) return doc text_splitter = RecursiveCharacterTextSplitter(chunk_size=2000, chunk_overlap=400) def add_doc(data,vectorstore): doc = read_file(data) splits = text_splitter.split_documents([doc]) vectorstore = Chroma.from_documents(documents=splits, embedding=embeddings) retriever = vectorstore.as_retriever(search_kwargs={'k':1}) return retriever, vectorstore def delete_doc(delete_name,vectorstore): delete_doc_ids = [] for idx,name in enumerate(vectorstore.get()['metadatas']): if name['name'] == delete_name: delete_doc_ids.append(vectorstore.get()['ids'][idx]) for id in delete_doc_ids: vectorstore.delete(ids = id) # vectorstore.persist() retriever = vectorstore.as_retriever(search_kwargs={'k':1}) return retriever, vectorstore def delete_all_doc(vectorstore): delete_doc_ids = vectorstore.get()['ids'] for id in delete_doc_ids: vectorstore.delete(ids = id) # vectorstore.persist() retriever = vectorstore.as_retriever(search_kwargs={'k':1}) return retriever, vectorstore