Teaching the filing cabinet to read

Among the things running on the little server by the router is Paperless-ngx, a self-hosted document system that the internet speaks of in reverent tones. You scan your post, it does the OCR, and — I told the household, with some confidence — it files everything itself. Titles, tags, who sent it. The filing cabinet thinks now. We live in the future.

I had scanned four documents. All four were sitting in the system under names like scan_2026-06-30.pdf, untagged, unfiled, and unthought-about. The future, it turned out, had not been switched on.

Worse: reading the documentation properly (a thing I do after making claims to the household, as is traditional), I learned that the future was never going to switch itself on. Paperless-ngx does have automatic tagging, but it’s a classifier that learns from your own manual filing. It can only ever apply tags you’ve already created and taught it by hand, across dozens of documents. It cannot invent a tag. It will never write a title. With four documents, it had precisely nothing to learn from and no intention of starting.

So the machine that was supposed to be doing the thinking was waiting for me to do the thinking, thoroughly, for several months, first.

There was, however, another option. I already had a small AI model running in the house on a mini PC — the one that answers questions and doesn’t send my prompts to California. What if it read the post instead?

$ cat teaching-the-filing-cabinet-to-read.md --full

cat: permission denied — the rest of this article is for subscribers

Subscribe to read the rest

Already a subscriber? Continue reading or sign in.

$ whoami

subscriber ✓

Continue reading