Your document scanner finally hired an intern

What it does

paperless-gpt is a companion web service for paperless-ngx that hands your scanned documents off to large language models. It auto-suggests titles, tags, correspondents, created dates, and custom fields, and can rebuild PDFs with searchable text layers using either LLM vision or third-party OCR engines. A built-in web UI lets you approve suggestions manually or let the pipeline run unattended.

The interesting bit

Instead of treating OCR as a separate preprocessing step, the project feeds images directly to multimodal LLMs or specialized document-AI services, then overlays the extracted text as invisible layers on the original PDF. It also supports ad-hoc analysis across multiple documents with custom prompts, effectively turning your scan archive into a queryable corpus.

Key highlights

• Supports OpenAI, Ollama, Mistral, and Anthropic for metadata extraction and OCR.
• OCR backends include LLM vision, Google Document AI, Azure Document Intelligence, and Docling Server.
• Generates searchable, selectable PDFs by placing transparent text layers over the original scan.
• Three write modes for custom fields—append, update, replace—so you control how aggressively it overwrites data.
• Can run fully offline via Ollama if you prefer not to ship scans to cloud APIs.

Caveats

• The README flags a PDF_REPLACE upload option as dangerous, noting it deletes the original document.
• Troubleshooting sections explicitly warn about token-management headaches with local LLMs and call out separate issues with PDF processing and custom-field generation.
• It is strictly a paperless-ngx satellite; without that upstream system, the service has nothing to manage.

Verdict

If you already run paperless-ngx and your scanner produces more untagged PDFs than you care to name by hand, this is a sensible force multiplier. If your documents are already born-digital and neatly organized, skip it.