Q&A

How does text extraction and OCR work in Paperwise?

Paperwise first tries to extract text directly from the document when possible. This works well for machine-readable PDFs and avoids unnecessary OCR.

When the file is a scan, image-heavy PDF, or low-quality document, Paperwise falls back to OCR based on the mode you choose in Settings > Model Config.

OCR paths

• Local Tesseract keeps OCR on your machine using tesseract and pdftoppm.
• LLM OCR using your main connection sends rendered page images to the same provider family you use elsewhere.
• Separate OCR model routing lets you choose a dedicated OCR model independently from metadata extraction and grounded Q&A.

For OpenAI-based vision OCR, you can also tune image detail to auto, low, or high.

Which models should I use?

Paperwise works well with both GPT and Gemini setups. Use task-specific models so OCR, extraction, and grounded Q&A are tuned separately.

Task	GPT example	Gemini example	Notes
OCR	gpt-5-mini	gemini-2.5-flash	Good fast multimodal starting points for scanned PDFs and forms
Metadata extraction	gpt-5-mini	gemini-2.5-flash	Balanced choices for structured fields
Grounded Q&A	gpt-5.1	gemini-2.5-pro	Better picks for more demanding cross-document questions
Budget / bulk work	gpt-5-nano	gemini-2.5-flash-lite	Useful for lighter classification and triage

If your documents are mostly clean text PDFs, start with the faster models and only move up when quality is not good enough.

Why does OCR sometimes time out on dense PDFs?

Large page images can take longer to process. Paperwise logs per-page OCR progress and retries timed-out pages. Longer OCR request timeouts are enabled for vision calls.

How are duplicate files handled?

Documents are deduplicated by SHA256 checksum to prevent repeated ingestion of the same file content.

Where can I get more help?

Open an issue in the repository:

• github.com/zellux/paperwise