Ollama and Paperless-NGX: local AI tagging and OCR setup

This tutorial adds document classification and vision OCR to an existing Paperless-NGX installation on Unraid. It builds on earlier parts of the series, so it assumes the document management stack is already running. The speaker uses Ollama to run models locally and Open WebUI to download them, with the GPU services on a separate server from Paperless.

Paperless AI uses qwen3:14b to assign tags, detect correspondents, set document types and generate titles. The setup covers API authentication and the choice between reusing existing tags and allowing new ones. Document chat and RAG search demonstrations include retrieving a breakdown service's phone number after indexing the collection.

Paperless GPT handles vision OCR with minicpm-v. The speaker disables its metadata generation to avoid overlap with Paperless AI, then tests an energy drink image, a server screenshot and a handwritten note. Automatic OCR uses a consume subfolder named after the processing tag, with PAPERLESS_CONSUMER_RECURSIVE and PAPERLESS_CONSUMER_SUBDIRS_AS_TAGS set to true.

The speaker recommends at least 12GB of VRAM and describes CPU processing as impractically slow for this workflow. He reports crashes with a Qwen 2.5 vision model on the Ollama versions he tested. His chosen setup keeps document processing on his own hardware; the examples do not establish accuracy for every document.