OCR engines vary widely. A simple screenshot can work with many engines, but invoices, tables, formulas, and scanned PDFs need stronger document parsing.
| Engine | Strength | Tradeoff |
|---|---|---|
| Tesseract | Open source, self-hostable, good for plain text | Weak layout and table structure |
| PaddleOCR-VL | Strong document layout parsing | Requires provider integration and tuning |
| Google Vision | Mature cloud OCR | Cost and cloud dependency |
| Azure OCR | Enterprise cloud workflow | Vendor-specific setup |
Example decision
If the source is a clean screenshot, Image to Text in Simple OCR mode is enough. If the source is an invoice with line items, use Invoice OCR or Formatted OCR.
Tables and formulas
For tables, use Table OCR. For equations, use Formula OCR and export Markdown so LaTeX-style notation is easier to keep.
Product integration
If your product needs a stable contract, wrap OCR behind POST /api/v1/ocr. That keeps authentication, credits, request logs, and response format consistent.
Cost planning
Use Pricing to estimate pages. Simple OCR is 1 credit per page, while Formatted OCR is 10 credits per page.
