OCR or native text: check the reference, not just the words
Does your PDF contain recoverable text or only an image? Choosing an extraction route avoids unnecessary work, but it does not replace verification. In our local test, a clean image produced readable text and three correct fields while changing the reference from OO10-I1 to 0010-I1. A similar-looking identifier does not necessarily identify the same document.
Waiting list; no immediate access or free plan guaranteed. The destination is in Spanish.
Situations to review
A one-page PDF contains five lines of native text.
For the synthetic nativo.pdf, PDF.js 4.10.38 returned five text items. We joined their strings with line breaks and compared them against a reference prepared before extraction, normalizing whitespace only.
All 81 positions in the normalized reference matched, and all four control-field lines were preserved literally. This establishes a result for that document and configuration only; it does not assess tables, columns, real documents or business accuracy.
Another synthetic PDF contains only an image, with no text layer.
The same extraction call returned zero items and an empty string for solo-imagen.pdf. To test OCR, we used our separate limpia.png and degradada.jpg images. We did not submit a PDF directly to Tesseract.js or test a converter or complete PDF-to-OCR pipeline.
This file returned no text items through PDF.js getTextContent. Recognizing its pixels requires a separate route. An empty result for another file would require investigating permissions, contents and the tool; it does not prove that every similar PDF lacks text.
OCR returns almost all the expected text but changes an identifier.
Tesseract.js 6.0.1, with the eng language model and single-block segmentation (PSM 6), read Reference: 0010-I1 from the clean image and Reference: 0010-11 from the degraded image. The prepared reference was Reference: OO10-I1: two capital letter Os at the start and a capital I followed by the digit one at the end.
Both outputs preserved Currency: USD, Subtotal: 180.00 and Approval: PENDING as exact lines, but the reference failed. Do not blindly replace zeros with Os: check the document or an authorized source before associating, updating or sending anything.
Waiting list; no immediate access or free plan guaranteed. The destination is in Spanish.
How to evaluate the work
- Define a minimal authorized sample, a checked reference transcription prepared before testing and the critical fields. Preserve originals and versions. Initially consider extraction for recoverable text and OCR for pixels that need recognition; handle mixed documents by page or region. Copying one word does not establish complete extraction of every page.
- Local test results from November 16, 2026, not customer outcomes: native PDF, edit distance 0/81 and four of four exact field lines; image-only PDF, empty extraction, 81/81 and zero of four; clean PNG, OCR 2/81 and three of four; degraded JPEG, OCR 3/81 and three of four. Character error rates for those two OCR outputs were 2.47% and 3.70%. The denominator is the full normalized 81-character reference, not the field count or an estimate of general accuracy. We do not aggregate these four tests into a score.
- Reproduction: create five synthetic lines containing TEST DOCUMENT; Reference: OO10-I1; Currency: USD; Subtotal: 180.00; Approval: PENDING. The supplied generator creates a native PDF, an image-only PDF and a 1800×900 PNG, then creates a variant reduced to 450×225, rotated by 2° and saved as JPEG quality 35. The execution script runs two PDF text extractions and two image OCR calls. Compare strings after trimming and collapsing whitespace, preserving case and punctuation; calculate Levenshtein distance/81 and exact matches for four field lines. Downloadable evidence records bytes, outputs, versions and hashes. The representations have different typography and layouts: this is not a controlled competition between engines or a test of real scans.
- Decision record: file/page/region; hash and version; route and configuration; control reference; literal output; mismatched field; review; decision; owner. In Think², collaborators, groups and context, compatible models and authorized tools can organize such an assignment; each collaborator can optionally have a computer. This test ran outside a Think² workflow and does not establish integrated OCR or autonomous correction. If your current extractor and review process already meet the need, you do not need another layer. A proposed use does not establish text fidelity. Contacting people, updating records or sharing documents requires the applicable explicit authorization; such authorization does not establish extraction accuracy.
Download the evidence
- Test outputs and metrics · JSON · SHA256: 734784e99ae3200c95dd353b7dd178fcff032c315a536fe826a888b0bf7ad2cd
- Synthetic samples and scripts — review before running · JSON · SHA256: 93133e137f934ce60a720d4d3e8938c983f1beba4e9a0cab4bb08691e5aff3de
Frequently asked questions
Can I choose OCR or native text extraction from the .pdf extension alone?
No. PDFs can contain text, images or combinations. Our call recovered text from a PDF constructed with text and nothing from the PDF constructed only with an image. Check the relevant document and pages rather than extending this result to every PDF.
Does a high OCR confidence score establish a correct reference?
No. The clean-image output reported confidence 93 while replacing both letter Os with zeros. We do not interpret that engine number as a calibrated probability that the field is correct. Identifiers require literal checking against a source.
Does this test establish that Think² reads invoices or that one engine is better?
No. It is a local test of four specific combinations using one synthetic English reference, without customer documents, timing, costs, tables, handwriting or Think² integration. Subtotal 180.00 and Approval: PENDING are test text, not an invoice, tax validation or actual permission. It does not establish results for other languages, versions or configurations.
Sources and scope
- Tesseract.js README — Image OCR and no direct PDF support; no accuracy guarantee or language list established by this consultation.
- PDF.js PDFPageProxy — getTextContent returns page text content; items/str details verified in the local execution, not established by the retrieved documentation excerpt.
Keep exploring
Waiting list; no immediate access or free plan guaranteed. The destination is in Spanish.