Player not loading? Watch on YouTube
This German-language tutorial adds Paperless-AI and Paperless-GPT to an existing Paperless-ngx installation through Docker Compose, with Ollama providing local AI models. Paperless-AI handles tags, correspondents, document types and titles. Paperless-GPT handles vision-based OCR. The presenter selects Llama 3.2 3B for text tasks and MiniCPM-V 8B for image recognition.
The setup covers host ports, storage folders, environment variables and the Paperless API token with its matching username. Models are downloaded through the web interface. A tag-based workflow passes documents from metadata processing to OCR, while a manual scan starts processing without waiting for the configured 30-minute interval. The presenter recommends backing up Paperless-ngx before allowing the add-ons to modify documents.
The German-document test produces mixed metadata and OCR results that the presenter judges worse than the standard Paperless-ngx output. Similar invoices receive inconsistent titles and correspondents, and one document gets an incorrect type. He likes the OCR's Markdown formatting but identifies invented text, an incorrect delivery date and altered financial identifiers.
These findings concern the demonstrated models, prompts and NAS, which has 16 GB of memory and no GPU. The presenter advises against this configuration for his German documents; the test does not establish how every model or hardware setup performs.
