Document scanning

How it works

When you upload a document, ScanLedger runs it through a multi-step pipeline:

1. Preprocessing — OpenCV and Pillow normalize the image: deskew, denoise, contrast correction.
2. Table detection — if the document contains tabular data, the table structure is detected separately so rows and columns map cleanly.
3. OCR extraction — the active engine (GPT-4 Vision or Gemini 2.5 Flash) reads the document and returns structured fields.
4. Confidence scoring — each field is scored; anything below 0.85 is flagged for review.

OCR engines

Two engines are supported, configurable via the OCR_ENGINE environment variable:

Gemini includes an automatic fallback to GPT-4 Vision for geo-restricted regions, so you never see a request fail because of provider availability.

Supported formats

• Images: JPG, JPEG, PNG, GIF, BMP, TIFF, WebP, HEIC
• PDFs: single- and multi-page (each page is processed individually)

Max file size is 20 MB per image and 50 MB per upload.

Document types

When scanning free-form (not through a template), you pick a document type so the model knows what to expect:

• invoice — supplier invoices with line items, totals, and parties.
• receipt — point-of-sale receipts, often with line items and taxes.
• inventory_log — stock logs with product, quantity, and movement direction.
• attendance_sheet — staff attendance or sign-in sheets.
• payment_slip — deposit slips or payment acknowledgements.
• general_table — any tabular document.
• general_note — free-form handwritten or typed notes.
• auto_detect — let the AI pick the best category.

Confidence scores

Every extracted field gets a number between 0 and 1 indicating how confident the model is. The UI highlights anything below 0.85. Click the field to edit — editing flips it to verified and updates the score.

Tips for better accuracy

• Use good, even lighting. Side shadows are the biggest cause of misread digits.
• Keep the camera square to the page and avoid perspective distortion.
• For handwriting, use dark ink on light paper.
• For glossy thermal receipts, tilt slightly to kill the glare before you shoot.
• If a document is critical, scan it as a PDF from a flatbed scanner rather than a phone.

Batch scanning

You can upload multiple files at once. Each file becomes an independent document. Pro and Enterprise plans process batches in parallel through the background worker; the Free trial is capped at 5 scans per day.

Plan limits

Next steps

• Document templates — map every scan to a fixed schema.
• Datasets — turn scanned data into searchable tables.