Text recognition
Purpose
A filed invoice is worth only as much as it is findable. This module makes files searchable by their content rather than by file name only — choosing per file type the route that actually works.
Scope
- PDF through the text layer (
pdftotext). The vast majority of PDFs in the document store —
invoices, quotes, exports — carry a text layer. Reading it takes milliseconds, whereas real image recognition takes minutes and reads worse.
- Scans through image recognition (
tesseract), but only once the text layer yielded almost
nothing — then it really is a scan.
- Images directly through image recognition.
- Office files as plain text: docx, xlsx, pptx and their OpenDocument counterparts are ZIP
archives of XML; the text is already plain in there. That is not a half result like guessed image recognition but exact reading.
- Recognition runs in the background and does not hold up the upload; when the file is deleted,
the recognised text goes with it.
Interplay
Recognition hangs off the document store and feeds full-text search. It is therefore also the precondition for a scanned document to appear in AI access at all.
Limits
- No Composer dependency — the tools are system programs. Without
pdftotextortesseracton
the server nothing happens; the module does not invent a result.
- Not supported: videos and the old binary formats
.docand.xls. Those would need further
programs, and a half result would be worse than an honest „not supported".
- Image recognition is estimation. With poor scans the text says what the program read — not
necessarily what is on the paper.