如何用 Docling Parse 建立版面感知的文件智慧解析流程

In this tutorial, we build a workflow for using Docling Parse to analyze PDF documents at a detailed structural level. We start by preparing a stable Python environment, handling common Colab dependency issues, and generating a custom multi-page PDF with text, columns, table-like content, vector shapes, and an embedded image. We then use Docling Parse to extract words, characters, and lines with page-level coordinates, render visual overlays, and save the results into structured JSON and CSV files. Through this workflow, we see how low-level PDF parsing can support document AI tasks such as layout analysis, reading-order reconstruction, table-aware processing, and retrieval-ready document preparation. Setting Up the Docling Parse Colab Environment and Dependencies Copy CodeCopiedUse a different Browserimport os, sys, subprocess, textwrap, json, time, shutil from pathlib import Path def run(cmd): print(f"\n$ {cmd}") return subprocess.run(cmd, shell=True, text=True, capture_output=False) run(f'{sys.executable} -m pip install -q --no-cache-dir -U "pillow>=10.4.0,<12" reportlab pandas matplotlib docling-core docling-parse') try: from PIL import Image, ImageDraw except ImportError: print("\nPillow import failed because Colab has a mixed PIL installation.") print("Reinstalling Pillow and restarting runtime. After restart, run this same cell again.") run(f'{sys.executable} -m pip uninstall -y pillow PIL') run(f'{sys.executable} -m pip install -q --no-cache-dir --force-reinstall "pillow>=10.4.0,<12"') os.kill(os.getpid(), 9) import pandas as pd import matplotlib.pyplot as plt from reportlab.lib.pagesizes import A4 from reportlab.lib import colors from reportlab.platypus import Table, TableStyle from reportlab.pdfgen import canvas from docling_core.types.doc.page import TextCellUnit from docling_parse.pdf_parser import DoclingPdfParser print("Environment ready.") print("Python:", sys.version.split()[0]) WORKDIR = Path("/content/docling_parse_advanced_tutorial") WORKDIR.mkdir(parents=True, exist_ok=True) PDF_PATH = WORKDIR / "advanced_docling_parse_demo.pdf" OUT_DIR = WORKDIR / "outputs" OUT_DIR.mkdir(exist_ok=True) DEMO_IMAGE_PATH = WORKDIR / "demo_bitmap.png" We set up the Colab environment by installing Docling Parse, Docling Core, Pillow, ReportLab, Pandas, and Matplotlib. We also handle the Pillow import issue safely so the notebook can recover if Colab has a broken or mixed PIL installation. We then define the working directory, output folder, PDF path, and image path that we use throughout the tutorial. Generating a Multi-Element Test PDF for Parser Evaluation Copy CodeCopiedUse a different Browserdef create_demo_image(path): img = Image.new("RGB", (320, 180), "white") draw = ImageDraw.Draw(img) draw.rectangle([20, 20, 300, 160], outline="black", width=3) draw.ellipse([55, 45, 145, 135], outline="black", width=4) draw.line([180, 140, 285, 45], fill="black", width=4) draw.text((45, 145), "Embedded bitmap image", fill="black") img.save(path) create_demo_image(DEMO_IMAGE_PATH) def build_pdf(pdf_path): c = canvas.Canvas(str(pdf_path), pagesize=A4) width, height = A4 c.setFont("Helvetica-Bold", 20) c.drawString(60, height - 70, "Docling Parse Advanced PDF Parsing Tutorial") c.setFont("Helvetica", 11) intro = ( "This generated document is designed for testing text extraction, coordinate parsing, " "line grouping, vector path detection, bitmap resources, and layout-aware reconstruction." ) text_obj = c.beginText(60, height - 105) text_obj.setLeading(15) for line in textwrap.wrap(intro, width=90): text_obj.textLine(line) c.drawText(text_obj) c.setFont("Helvetica-Bold", 14) c.drawString(60, height - 170, "1. Two-column text region") left_para = ( "The left column contains compact explanatory text. A parser should expose words, " "characters, and line-level cells along with coordinates. These coordinates allow us " "to reconstruct reading order and inspect the spatial structure of a page." ) right_para = ( "The right column contains a separate paragraph. In document AI pipelines, layout " "features are useful for retrieval, table extraction, chunking, and downstream RAG " "applications where page position can matter." ) y_start = height - 200 left_text = c.beginText(60, y_start) left_text.setFont("Helvetica", 10) left_text.setLeading(13) for line in textwrap.wrap(left_para, width=42): left_text.textLine(line) c.drawText(left_text) right_text = c.beginText(325, y_start) right_text.setFont("Helvetica", 10) right_text.setLeading(13) for line in textwrap.wrap(right_para, width=42): right_text.textLine(line) c.drawText(right_text) c.setStrokeColor(colors.darkblue) c.setLineWidth(2) c.rect(55, height - 315, 225, 130, stroke=1, fill=0) c.rect(320, height - 315, 225, 130, stroke=1, fill=0) c.setStrokeColor(colors.darkgreen) c.setLineWidth(3) c.circle(140, height - 390, 40, stroke=1, fill=0) c.line(220, height - 430, 310, height - 355) c.setFont("Helvetica-Bold", 14) c.setFillColor(colors.black) c.drawString(60, height - 470, "2. Simple table-like structure") data = [ ["Section", "Signal", "Expected parser behavior"], ["Text", "Words and lines", "Return text cells with coordinates"], ["Vector", "Boxes and lines", "Expose page path/vector resources"], ["Bitmap", "Embedded image", "Expose or render image resources"], ] table = Table(data, colWidths=[100, 130, 260]) table.setStyle(TableStyle([ ("BACKGROUND", (0, 0), (-1, 0), colors.lightgrey), ("GRID", (0, 0), (-1, -1), 0.7, colors.black), ("FONTNAME", (0, 0), (-1, 0), "Helvetica-Bold"), ("FONTSIZE", (0, 0), (-1, -1), 9), ("VALIGN", (0, 0), (-1, -1), "MIDDLE"), ])) table.wrapOn(c, width, height) table.drawOn(c, 60, height - 590) c.setFont("Helvetica", 9) c.drawString(60, 55, "Page 1: generated programmatic PDF with text, table-like layout, and vector paths.") c.showPage() c.setFont("Helvetica-Bold", 18) c.drawString(60, height - 70, "Page 2: Bitmap, Dense Text, and Reading Order") c.setFont("Helvetica", 10) dense = ( "This page includes an embedded bitmap image and several short blocks of text. " "We use it to test whether rendering works, whether the parser preserves page-level " "coordinates, and whether our own reconstruction logic can group words into lines." ) y = height - 105 for para_idx in range(4): tx = c.beginText(60, y) tx.setFont("Helvetica", 10) tx.setLeading(13) for line in textwrap.wrap(f"Block {para_idx + 1}: {dense}", width=92): tx.textLine(line) c.drawText(tx) y -= 70 c.drawImage(str(DEMO_IMAGE_PATH), 110, height - 510, width=320, height=180, preserveAspectRatio=True) c.setStrokeColor(colors.red) c.setLineWidth(2) c.roundRect(95, height - 525, 350, 210, 10, stroke=1, fill=0) c.setFillColor(colors.black) c.setFont("Helvetica-Bold", 12) c.drawString(60, height - 570, "Coordinate-aware extraction lets us keep page, text, and position together.") c.setFont("Helvetica", 9) c.drawString(60, 55, "Page 2: embedded bitmap image and multiple text blocks.") c.save() build_pdf(PDF_PATH) print("Created PDF:", PDF_PATH) We generate a small bitmap image and create a custom two-page PDF for testing Docling Parse. We add text blocks, two-column content, vector shapes, table-like content, and an embedded image so the parser has multiple document elements to process. We use the generated PDF as a controlled input to check text extraction, layout structure, rendering, and coordinate-aware parsing. Extracting Word, Character, and Line Cells with Docling Parse Copy CodeCopiedUse a different Browserdef safe_to_dict(obj, max_depth=2): if obj is None: return None if isinstance(obj, (str, int, float, bool)): return obj if isinstance(obj, (list, tuple)): return [safe_to_dict(x, max_depth=max_depth - 1) for x in obj[:50]] if isinstance(obj, dict): return { str(k): safe_to_dict(v, max_depth=max_depth - 1) for k, v in list(obj.items())[:50] } if hasattr(obj, "model_dump"): try: return obj.model_dump() except Exception: pass if hasattr(obj, "__dict__") and max_depth > 0: try: return { k: safe_to_dict(v, max_depth=max_depth - 1) for k,