Extraction begins with the file’s real content
A PDF may contain selectable text, page images, or both. A DOCX is structured XML. A spreadsheet stores values, formulas, and formatting in cells. HTML combines visible content with navigation and hidden elements. Images need OCR.
The first question is whether machine-readable text already exists. Parsing it is usually faster and more accurate than OCR.
| Source | Primary method | Main limitation |
|---|---|---|
| Selectable PDF | Text-layer extraction | Reading order and columns |
| Scanned PDF/image | Render, preprocess, OCR | Resolution, language, skew |
| DOCX/PPTX | Parse package content | Text boxes, tables, notes |
| XLSX/CSV | Read cells/records | Formulas, merged cells, quoting |
| HTML | Extract main/visible text | Navigation and repeated chrome |
Use the Jivaro extractor in four passes
Open OCR & Document Text Extractor.
Identify native text versus scans.
Choose pages, language, and processing options.
Remove artifacts and compare samples.
Chunk only after the text is correct.
Use native PDF text first
If text can be selected, direct extraction avoids character-recognition errors. Check columns, line-end hyphenation, repeated headers, ligatures, and tables. For mixed PDFs, extract native pages directly and OCR only image-only pages.
OCR scans and images
- Use enough resolution.
- Correct rotation and skew.
- Improve contrast without deleting thin strokes.
- Crop borders and background clutter.
- Select the correct language.
- Treat handwriting as a separate, less reliable problem.
Aggressive thresholding can erase punctuation and accents. Compare difficult pages with the source.
Office files and presentations
DOCX and presentation files can contain tables, headers, footers, comments, footnotes, floating text boxes, hyperlinks, and notes. Extraction recovers content; it does not reproduce layout. Compare complex pages and retain the original.
Spreadsheets and CSV
Decide whether you need displayed values, formulas, or both. Preserve headers, sheet names, and record boundaries. CSV requires delimiter and quoting awareness: a comma inside a quoted field is not a new column.
A number without its row label, column header, sheet, or cell reference may be useless or misleading.
HTML extraction
Raw text can repeat menus, cookie notices, footers, and related links. Prefer the main content and preserve links that carry meaning. Process only pages you own or have permission to use.
Clean without rewriting meaning
- Remove repeated headers and footers.
- Join words broken by line-end hyphens.
- Normalize spaces and line breaks.
- Remove page-number-only lines.
- Preserve paragraphs before chunking.
Global replacements can delete legitimate dates, quantities, identifiers, or formatting.
Verify names, numbers, and citations
- Search critical names, numbers, dates, and references.
- Compare a clean page, a difficult page, and a table-heavy page.
- Inspect unusual words and symbols.
- Confirm page/section coverage.
- Retain page numbers for formal quotations.
Chunk and export last
Clean first, then split by words or characters. Use semantic boundaries, overlap only when needed, number chunks, and preserve document/page context. Export TXT for simple text, DOCX for editing, CSV for records, HTML for structure, and ZIP for multiple outputs.
Measure OCR quality instead of trusting a clean-looking page
For consequential extraction, select a representative sample and calculate a simple error rate: compare a fixed number of words or fields against the source, count substitutions, omissions, and insertions, and record the conditions. Include small text, a table, a low-contrast area, and at least one page with names or numbers. A 99% character result can still contain a critical error in every account number or medication name.
When confidence is insufficient, preserve page images beside the extracted text and require a human to verify high-risk fields. OCR is best used to accelerate review, not to erase the source.
Privacy and file handling
Local browser processing reduces the need to upload documents, but the files still exist on the device, in browser memory, and in any downloaded outputs. Use a trusted computer, close unrelated extensions when the material is sensitive, avoid shared download folders, and delete temporary exports according to the project’s retention rules. Do not assume browser storage is encrypted backup.
Frequently asked questions
Try selecting and copying text. If the page behaves like one image or copied text is unusable, OCR may be needed.
PDFs and presentations may store positioned text rather than a logical reading sequence.
The workflow is primarily for printed text; handwriting accuracy varies widely.
After cleanup and verification, with semantic boundaries and retained source context.
Related Jivaro apps
Extract and review text from PDFs, scans, images, Office files, spreadsheets, HTML, CSV, Markdown, and plain text, then export searchable PDFs or structured text locally.
Open appSources and references
- Tesseract User ManualTesseract OCR · reference
- File APIMDN Web Docs · reference
- OCR & Document Text ExtractorJivaro · first-party