SoulkeyResearch

Research notes / version 1.0

Small tests. Explicit rules. Inspectable results.

By SoulkeyResearch · 30 September 2026

Our first edition asks a narrow question: how do two configured text-extraction workflows and one template-guided OCR workflow handle three clean, synthetic PDF tables? The files, outputs, scoring rules and reproduction code are available together.

1. What is in the sample?

  • Digital table: six invented purchasing records in six columns on one page, with embedded text and a ruled table.
  • Repeated headers: 12 invented records on two pages, with the same six column labels repeated on page two.
  • Image-only scan: the digital first page rendered at 200 DPI, converted to grayscale, and embedded as an image-only PDF. There is no searchable text.

Fields are RecordID, Date, Description, Qty, UnitUSD and TotalUSD. The sample includes leading zeros, ISO dates, two-word descriptions and decimal amounts. It contains no real customer, employer or personal data. We added no skew, blur, noise, handwriting or compression damage.

2. Which workflows did we run?

pdfplumber 0.11.8 + exact-header filter

We used default table extraction on each page, collapsed whitespace, concatenated tables, and removed only rows that exactly matched all six column labels. This is configured output, not untouched one-click output.

Poppler pdftotext 25.03.0 + six-column parser

We ran pdftotext -layout. The parser admitted rows starting with an R-plus-three-digits ID and containing exactly six fields separated by two or more spaces. Rejected candidate rows were logged. This parser is specific to the fixture format.

Tesseract 5.5.0 + known-cell crops

We tested OCR only on the scan. Each of the 36 data cells was cropped using the known table geometry and read with English language data and page-segmentation mode 7. Only whitespace was normalized. There was no character whitelist, O-to-zero substitution or manual correction.

Important advantage: the OCR workflow already knew every cell’s position. It did not discover a table or its structure automatically. Do not compare its score with a general document-OCR claim.

3. How were results scored?

Expected CSVs were generated from the original synthetic records independently of extraction. We compared exact strings at each original row and column position. Missing cells count as wrong; headers and page furniture are excluded.

We also report exact complete rows, matching IDs, unexpected IDs, duplicate IDs and extra rows. This makes it possible to see that correct numeric totals can coexist with broken record identities. No numeric or OCR values were repaired before scoring.

The scan’s 31/36 exact cells comprise 18 correct numeric cells, six dates, six descriptions and one correct ID. Five IDs were wrong. The six line totals still summed to $323.87.

4. Reproduce the run

Unzip the test kit into a new folder. Use the versions below, with Poppler’s pdftotext and pdftoppm, Tesseract English language data, and DejaVu Sans fonts available. Python dependencies are listed in requirements.txt.

python -m pip install -r requirements.txt
python build_sample.py

The script expects fonts in /usr/share/fonts/truetype/dejavu/. Adjust FONT_DIR if your fonts are elsewhere. The reproduction run uses local files and tools; it makes no API calls and needs no API key. Dependency installation requires access to the relevant package source.

Python 3.12.14
pdfplumber 0.11.8
Poppler 25.03.0
Tesseract 5.5.0
ReportLab 4.4.9
Pillow 12.3.0

Results may vary with tool versions, OCR language data and fonts. The supplied outputs are the recorded run, not a promise that a different environment will produce identical files.

5. Inspect the source material

The ZIP includes unedited text and table extraction, rejected rows, naive and cleaned CSVs, and raw OCR output. Keep these alongside the cleaned output when evaluating a different workflow.

6. What these results do not establish

  • No representative real-world accuracy rate: three clean fixtures are a small, deliberately controlled sample.
  • No commercial-tool comparison: PDF.co, Adobe Acrobat and other paid converters were not tested.
  • No speed, runtime-cost, security, retention or production-suitability conclusion.
  • No claim about mixed layouts, damaged scans, handwriting, other languages or automatic table recognition.

The next useful extension is a permission-cleared set with varied layouts, skew, blur and more difficult fields. A paid tool should receive the same fixtures and scoring rules before we report its performance.

Affiliate and editorial disclosure

This site is published by SoulkeyResearch. Some pages link to PDF.co using an affiliate referral. If a visitor signs up and makes a purchase through that link, Soulkey may receive a commission. Each referral link is labeled and accompanied by a disclosure.

PDF.co is an untested option in this edition. The affiliate relationship does not turn its product claims into test results. Free local tools, complete results and downloads are available without following an affiliate link.

Privacy and document handling

The PDF text checker reads a file you select in your browser. It does not upload that file, send its text or filename to a server, or save it in local storage. It keeps only temporary in-memory counts for the current result. Clearing the check removes those references. The checker and PDF.js assets are served from this site; demo buttons download our public synthetic fixtures.

This site has no account signup, payment form or visitor-data collection form. We have not added analytics scripts or marketing cookies. Downloads are available directly, without an email gate. The checker does not execute PDF scripts, render pages, perform OCR or convert tables.

The hosting provider may process standard request logs or use operational/security cookies. External websites, including an affiliate destination, have their own privacy and cookie practices. Emailing kim@soulkey.com shares the information you choose to send. Do not send confidential documents with a general inquiry.

Official tool references

Reference pages checked 30 September 2026: pdfplumber, Poppler, Tesseract, and PDF.co product information. Test claims are grounded in the downloadable run, not in vendor marketing.

For a correction, identify the guide or file and the specific issue in an email to kim@soulkey.com.