Our approach
OCR.now documents the live recognition path so users can separate what the tool measures from what still requires source review.
Apply the review method with the OCR Quality Checker, then interpret the result with the OCR confidence guide.
Recognition path
Image uploads are validated with Pillow. PDF uploads are checked for a PDF signature, converted into page images with pdftoppm at 220 dpi, and processed with Tesseract using automatic page segmentation. Text is returned with explicit page labels for multi-page files.
Recognition path checks
- Accepted sources: PDF, JPG, JPEG, PNG, WebP, BMP, TIFF.
- Maximum upload size: 20 MB.
- PDF processing limit: first 25 pages per request.
- OCR process timeout: configured by the deployment.
- Available language choices come from installed Tesseract data rather than a fictional list.
Confidence and word counts
OCR.now runs a text pass and a TSV diagnostic pass. It averages confidence values only for recognized words with non-negative engine scores. The displayed value is an engine estimate, not a calibrated probability that the page or any specific field is correct.
Confidence and word counts checks
- Average confidence can conceal one consequential error.
- A word count describes recognized word tokens, not source completeness.
- Comparisons are most useful on the same passage with one controlled input change.
- Names, dates, totals, identifiers, and instructions require direct source review.
- No confidence threshold certifies fitness for legal, medical, financial, or safety decisions.
Temporary processing and file handling
The application creates a temporary working directory for each recognition request. Uploaded files and rendered PDF page images are removed when that request's processing context closes. The browser keeps a local preview object only for the active page session.
Temporary processing and file handling checks
- Raw upload paths are not exposed as public URLs.
- File extensions and underlying image readability are validated.
- Image pixel count is limited to reduce unsafe resource use.
- Subprocesses have time limits.
- High-risk documents should still use an approved offline or organizational workflow.
What the current tool does not do
Several pages document complete workflows that extend beyond the live browser output. Those pages state the boundary near the tool rather than implying a feature exists.
What the current tool does not do checks
- It does not automatically generate valid LaTeX from mathematical structure.
- It does not write a new searchable PDF text layer.
- It does not infer validated form fields, table cells, chart data, or accounting records.
- It does not compare two documents in one request.
- It does not verify document authenticity, meaning, arithmetic, authorship, or legal effect.
How to evaluate a workflow
Use a representative sample containing the hardest layouts and smallest consequential characters. Keep a trusted reference, document normalization rules, count exact corrections, and report results by document class rather than publishing one unexplained accuracy percentage.
How to evaluate a workflow checks
- Preserve the original source and processing settings.
- Measure character, word, field, layout, and reading-order errors separately when they matter.
- Include difficult and minority document classes in the sample.
- Investigate failed batches and draw a fresh sample after correction.
- Publish limitations beside any performance claim.