OCR Fallback

What is a scanned PDF?

A scanned PDF is a PDF that contains images of pages rather than embedded text. It is created by scanning a physical document and saving it as a PDF. Unlike text-native PDFs (created by Word, LaTeX, or a PDF printer), scanned PDFs have no text layer — they are just images.

When pdfjs-dist extracts text from a scanned PDF, it returns an empty string. Without OCR, the LLM receives nothing to work with.

Automatic detection

resume-intel automatically detects scanned PDFs by measuring text density:

• If the average extracted text is < 50 characters per page → classified as scanned (95% confidence)
• If the average is < 150 characters per page → classified as likely scanned (70% confidence)
• Otherwise → classified as text-native

When a scanned PDF is detected, the OCR pipeline triggers automatically. No configuration needed.

The OCR pipeline

PDF buffer
    ↓
pdfjs-dist renders each page to PNG
    ↓
@napi-rs/canvas provides the canvas context
    ↓
Tesseract.js (WASM) runs OCR on each PNG
    ↓
OCR text cleaner strips artefacts
    ↓
Clean text → LLM extraction

Step 1 — Rasterization

Each PDF page is rendered to a PNG image at 150 DPI using pdfjs-dist with @napi-rs/canvas as the rendering backend. 150 DPI provides a good balance between OCR accuracy and processing speed.

Step 2 — OCR

Tesseract.js runs locally in WASM mode. A single worker is created and reused across all pages to avoid re-initializing the WASM engine for each page.

Step 3 — Text cleaning

OCR output contains artefacts that LLMs interpret literally. The cleaner removes:

• Block-fill characters from skill meters (████░░░, ■■■□□)
• Pipe/dash separator lines (| | | |, ----------)
• Isolated single special characters (|, •, ·)
• Page number patterns (Page 1 of 3, - 2 -)
• Repeated header/footer lines (same line appearing 2+ times)

This eliminates corrupted field values like "fluency": "|" that were common in v0.1.0.

Checking if OCR was used

const result = await parseResume(pdfBuffer, { model })
 
if (result.meta.ocrFallback) {
  console.log('Scanned PDF detected — OCR was used')
}

Performance

OCR is significantly slower than text extraction:

The extra time comes from Tesseract.js initialization and per-page image processing. This is a one-time cost per document.

Limitations

• OCR accuracy depends on scan quality. Low-resolution or blurry scans may produce poor results.
• Bounding box coordinates are not available from OCR output — spatial reconstruction is skipped.
• Languages other than English require specifying the language option (see below).

Language support

// Default is English
const result = await parseResume(pdfBuffer, {
  model,
  // OCR language is configured internally — English by default
})

For multilingual CVs, the OCR language is currently fixed to English (eng). Support for additional Tesseract language packs is planned for a future release.