Skavio Agent Tools: PDF text extraction or OCR? Check the page first

Skavio Agent Tools provides authenticated, company-scoped document processing through asynchronous API jobs. For business paperwork, the first useful decision comes before submission: does the PDF contain usable text, or only a picture of it? Inspect the document page by page, select the appropriate documented operation, then compare the result with the source. Readable output is not, by itself, a validated business record.

1. Inspect the text—not just the file extension

A PDF exported from business software and a PDF containing scanned paper may look similar on screen, but they can require different reading methods. Ordinary text extraction reads text stored inside the PDF. Optical character recognition, or OCR, recognizes characters in page images. Open the document and copy a representative passage into a plain-text editor. Compare it with the visible page: are the reference number, punctuation and amounts intact? Repeat the check across the document, particularly where attachments or scanned pages begin. Selectable text is only a diagnostic clue. A scan may already contain an invisible OCR text layer, including recognition errors. Conversely, a page that looks perfectly readable to a person may contain no extractable text.

2. Choose the documented operation, not OCR for everything

The published Skavio Agent Tools catalogue lists `pdf_to_text` and `document_ocr` as distinct operations. That supports an explicit choice; it does not establish automatic scan detection or an automatic OCR fallback. Use `pdf_to_text` when the stored text is usable. Do not routinely turn a digital PDF into images just to recognize it again: native extraction can use font and encoding information that image recognition does not have. Where the content exists only as an image, check the documented requirements for `document_ocr` before submitting it. For a mixed PDF, note which pages have usable text and which need image recognition. Check the current documentation for supported inputs and processing scope rather than assuming per-page routing is built in. Neither method guarantees correct reading order or reconstructed tables.

3. Resolve obvious scan problems before processing

Before submitting scanned paperwork, inspect orientation, skew, small print, shadows and cropped edges. Tesseract’s external OCR guidance documents how skew, noise and unsuitable resolution can affect recognition. This is general input-quality guidance, not evidence that Skavio uses Tesseract or exposes its settings. Consider a fictional supplier invoice whose scan cuts through the totals column. Do not treat another OCR attempt as evidence for the missing digits. Obtain a complete scan instead. If you crop or straighten a working copy, preserve the original and check that labels, negative signs and footnotes remain visible.

4. Submit a traceable job and distinguish acceptance from completion

Skavio Agent Tools uses company-scoped Bearer authentication, a shared prepaid credit wallet and asynchronous jobs. Its documentation demonstrates a PDF request using `tool: pdf_to_text`, an `idempotency_key` and `arguments.data_base64`. Do not reuse that example as an assumed OCR request schema; check the current documentation for `document_ocr` requirements. The documented submission response includes a credit quote. Inspect that quote rather than budgeting from an old price table. This does not establish a separate quote-and-confirm step before submission. A queued response means the job has been accepted for asynchronous handling—not that the document has been read successfully.

5. Review business meaning, not merely readable words

Compare the downloaded result with the original pages before passing it into another business process. Check coverage as well as individual characters: a readable first page does not establish that every attachment or table was captured. On fictional supplier paperwork, a quantity and unit price might both appear correctly while their relationship is lost when columns become plain text. Likewise, a plausible amount is still wrong if its decimal separator or credit-note sign has changed. Record uncertain values as exceptions; do not silently fill them from context. Keep approval separate from extraction. The objective is not simply to obtain text, but to establish which parts of that text are supported by the source.

Explore Skavio Agent Tools →