CÉDRIC HECKLY CHÂTEL-ST-DENIS // 46.529° N, 6.901° E
FR

Infrastructure/Case study 03

Paperless-ngx and local AI

Filing of my personal documents with Paperless-ngx, complemented by language models running on the server.

  1. Document
  2. Paperless-ngx
  3. n8n
  4. Local model
  5. Larger model
  6. Tags

Goal

Find a document without filing it by hand: a consistent title, a type and tags assigned automatically, with as few errors as possible.

Setup

Paperless-ngx runs on the server and holds my personal documents. Ollama is installed next to it, without a GPU, with at most 8 GB of memory.

Results

Automatic renaming of documents by a model works fairly well when the prompt is carefully written. Results on tags are more uneven: in October 2026 I rate them as average, which is a polite way of saying I still fix a lot by hand.

Target workflow

I installed n8n locally to build a multi-step workflow. A small local model does a first pass, then hands over to larger models that analyse the document in detail. Tags are applied only if the result is fully consistent.