Make PDF searchable (OCR)

Use it whenever you have a PDF where the text isn't actually text — typical examples:

• Scanned contracts, deeds, transcripts, certificates
• Old archived reports digitised from paper
• Phone-camera "scans" of receipts and invoices
• PDFs exported from screenshot software where the source was an image

After processing, every PDF reader (Adobe Acrobat, Preview, Foxit, Chrome, Firefox) will let you select text, search with Ctrl/Cmd-F, and copy content into other apps — exactly as if the document had been typed natively.

This is the same "image-over-text" technique used by every commercial OCR pipeline (ABBYY FineReader, Acrobat Pro, Tesseract's own pdf output):

1. Rasterise. Each page is rendered to a PNG at the chosen DPI.
2. OCR. Tesseract recognises the text and remembers where each word landed on the rasterised page.
3. Build a page. Tesseract writes a one-page PDF where the PNG is the visible page content and the recognised text is drawn on top in text rendering mode 3 (invisible but still selectable and searchable).
4. Merge. pdf-lib copies all the per-page PDFs into a single document.

The visible result is pixel-identical to a clean scan, but the document is now a real searchable PDF.

DPI controls the resolution of the rasterised pages — and also affects how well OCR can read them:

• 150 DPI — smallest output, faster. Use for clean modern scans.
• 200 DPI — recommended default. Best speed / accuracy balance.
• 300 DPI — best accuracy on small / faded text. Doubles file size.

Language — pick the dominant script of the document. Tesseract tunes character recognition per language, so a mismatched pack will produce noticeably worse text. Mixed-language documents are still best served by picking the dominant language.

The output is almost always larger than the input — frequently 2×–4× depending on the source. Two costs add up:

• Re-rasterisation. Each page is rendered to a fresh PNG at your chosen DPI and embedded back into the PDF. A "small" scan compressed with JPEG can balloon when it's re-encoded losslessly.
• Invisible text layer. Tesseract embeds a glyphless font and one positioned text run per recognised word. On a dense page that's hundreds of objects.

Rules of thumb at the recommended 200 DPI:

• Clean printed scan: ~1.5×–2× the input size.
• Phone-camera "scan" with a photo background: ~2×–3×.
• Heavy colour or grayscale photo content: can exceed 4×.

Things this tool doesn't try to do:

• Reconstruct vector content. The output stores each page as a rasterised image — vector text, lines and shapes from the original become pixels. This is by design (it's how every "image-over-text" searchable PDF works) and is fine for scans; it's the wrong tool for clean vector PDFs.
• Preserve form fields, annotations or bookmarks. The output is a fresh PDF built from rendered pages — any AcroForm / annotation / outline metadata from the source is dropped. Run /tools/pdf-fill or similar separately if you need those features.
• Recognise handwriting. Tesseract is trained on printed text.
• Detect tables. Cells are searchable individually, but the table structure isn't surfaced anywhere in the PDF.

For the canonical "I have a scan and I want to search inside it" workflow this tool is exactly right. For anything beyond that, treat the searchable PDF as an intermediate format and run downstream tools on it.