Spatial Extraction

The problem with standard PDF extraction

Standard PDF extractors read text in rendering order — the order in which the PDF renderer draws characters on screen. For a single-column document this matches reading order. For a two-column CV it does not.

Given this layout:

┌─────────────────┬─────────────────┐
│ Senior Engineer │ TypeScript      │
│ 2020 – Present  │ Node.js         │
│                 │ React           │
│ Junior Engineer │                 │
│ 2018 – 2020     │ PostgreSQL      │
└─────────────────┴─────────────────┘

A standard extractor produces:

Senior Engineer TypeScript
2020 – Present Node.js
Junior Engineer React
2018 – 2020 PostgreSQL

Dates and job titles are interleaved with skills. The LLM receives semantic chaos and hallucinates.

How resume-intel solves this

resume-intel extracts bounding box coordinates for every text block alongside the text content. It then:

1. Detects column boundaries — sorts blocks by X position and identifies significant horizontal gaps (> 8% of page width) that indicate column separators
2. Groups blocks into columns — assigns each block to a column based on its center X coordinate
3. Sorts within each column — sorts blocks top-to-bottom by Y coordinate within each column
4. Concatenates left-to-right — joins columns in order with a visual separator

The LLM receives:

Senior Engineer
2020 – Present

Junior Engineer
2018 – 2020

---

TypeScript
Node.js
React

PostgreSQL

Clean, ordered text that the LLM can parse correctly.

Configuration

const result = await parseResume(pdfBuffer, {
  model,
  layoutStrategy: 'spatial', // default — uses bounding box algorithm
  // layoutStrategy: 'linear', // faster but less accurate for multi-column
})

Limitations

• Rotated text — text rotated more than 45° may not be correctly ordered
• Overlapping columns — very narrow column gaps (< 8% of page width) may not be detected
• Complex tables — nested tables within CVs may require manual post-processing
• Scanned PDFs — no bounding boxes available from OCR; spatial reconstruction is skipped (text is returned as-is from Tesseract)

For scanned PDFs, see OCR Fallback.