The Empty Quadrant: A Three-Part Series on the Design Space of Frontend PDF Extraction

TLDR

PDF extraction engines occupy four distinct quadrants along two primary architectural axes: Execution Environment (In-Browser vs. Server-Side) and Processing Strategy (Probabilistic ML vs. Deterministic Geometry). The Ginexys PDF Processor fills the previously unoccupied Deterministic-Structural-Frontend quadrant by combining raw operator vector parsing, baseline-proximity underline filtering, and off-thread Web Worker execution without shipping server dependencies or ML model weights.

Extraction Quadrant	Primary Technology Stack	Key Tradeoff	Ideal Use Case
Server-Side ML Models	Python, LayoutLM, LayoutParser	High server cost, slow turnaround	Unstructured scanned documents
Browser Canvas Viewers	pdfjs-dist Canvas Renderer	Rigid non-reflowable absolute text	Visual print-preview rendering
Browser Text-Only Parsers	Native PDF.js getTextContent()	Zero table/border vector awareness	Simple single-column text scraping
Deterministic Frontend	CTM Operators + Geometry Worker	Zero ML weights, $100\%$ Browser-native	Structured multi-column PDF export

---

The three-part architecture series

Part 1: mapping the 2x2 design matrix

• Cataloging Existing Camps: Analyzes pdf2htmlEX, tesseract.js, transformers.js, and text-only Y-clustering tools. Evaluates what each architecture trades away (server payload, canvas rigidity, or structural vector awareness).

Part 2: eight architectural invariants

• Filling the Empty Cell: Breaks down the eight specific architectural moves required to build a deterministic frontend layout engine: CTM-baked vector parsing, single-region text scoping, baseline underline filtering, topological cell merging, nearest-cell snapping, Web Worker isolation, per-page message streaming, and visual grid mapping.

Part 3: borrowed math, original composition

• Composition as Originality: Examines how textbook algorithms (lattice reconstruction, KD-trees, XY-cuts, Y-band clustering) are combined into a browser-native web processing pipeline.

Rule of thumb: Combine vector CTM operator parsing with Web Worker isolation to fill the deterministic-structural-frontend PDF extraction niche.