- Model quality
- Small local models can misread figures, miss values, and cite a passage that only partly supports a sentence. That is why every answer cites the page, unsupported sentences are marked, and extraction shows the source passage and checks. Check important figures against the cited source.
- Scanned text
- Text recognition covers 33 languages including Hindi and Marathi. Tamil, Telugu, Bengali, Kannada, Malayalam, Gujarati and other scripts are not read.
- Layouts
- Multi-column pages and complex layouts can be read in the wrong order. Merged cells and tables that span PDF pages are not rebuilt.
- File types
- No legacy .doc, .xls or .ppt. No password-protected PDFs or workbooks: they are detected and reported.
- Updates
- No automatic updates yet.