Borrowed Math, Original Niches: What's New When the Algorithms Aren't

TLDR

Building a browser-native PDF extraction engine involves minimal mathematical innovation: lattice intersection algorithms, Y-band clustering, and XY-cut partitioning have existed in academic literature since the 1980s. Originality lies in architectural constraints: executing the entire pipeline client-side inside Web Workers on top of PDF.js, enforcing non-overlapping region invariants, and enabling per-page streaming.

Architectural Dimension	Traditional Python Extraction Tools	Client-Side Web Worker Pipeline
Execution Environment	Server-side Python runtime	Client-side Web Worker threads
Data Privacy	Requires uploading PDF bytes to server	100% On-device processing (Zero server transit)
Memory Isolation	Single-threaded backend process	Nested Web Workers with page streaming

---

What is borrowed vs. What is novel

Standard industry primitives (borrowed)

• Lattice Line Intersections: Projection profile algorithms (Tabula, Camelot, pdfplumber).
• Y-Band Text Clustering: Font-proportional line grouping (pdfminer).
• Spatial Indexing: Standard 2D Spatial Hash and KD-Tree lookups.

Architectural innovations (novel)

• Nested Web Worker Execution: Running spatial extraction inside dedicated Web Workers on top of PDF.js operator streams.
• Non-Overlapping Region Invariants: Tracking assignedTextIndices to prevent multi-pass classifiers from claiming identical text twice.
• Viewport vs. Point Coordinate Discipline: Storing both PDF user points (fontSize, width) and viewport pixels (vFont, vWidth) explicitly on every metadata record.

---

Technical architecture: non-overlapping text indices

Prevent multi-stage classifiers from double-claiming text using a shared assignedTextIndices bitset:

export class DocumentExtractionContext {
  constructor(textItems) {
    this.textItems = textItems;
    this.assignedIndices = new Set();
  }
// Retrieve only unclaimed text items for downstream classifiers   getUnclaimedItems() {     return this.textItems.filter((_, idx) => !this.assignedIndices.has(idx));   }
// Claim items once a classifier claims a region   claimItems(claimedItemIndices) {     claimedItemIndices.forEach(idx => this.assignedIndices.add(idx));   } }

By maintaining this invariant across stages (Lattice $\rightarrow$ Stream Table $\rightarrow$ Column Split $\rightarrow$ Paragraph), downstream detectors never receive items already owned by earlier stages.

Rule of thumb: Define product novelty by your architectural constraints and negative space, what you refuse to build, rather than expecting core mathematical primitives to be unique.