Paperless-ngx vs a Folder of PDFs: When OCR Jobs Fill the Disk

A folder of PDFs on the NAS is a filing cabinet. Paperless-ngx is a factory. You drop a scan in consume, Tesseract wakes up, a 40-page statement becomes a fat derivative, the search index grows, and the 256 GB SSD you thought was “plenty for documents” hits 92 percent because OCR scratch and thumbnails are not the same bytes as the original PDF. The folder would have taken the original and shut up. Paperless took a job.

I like Paperless. I still keep a boring folder for things I will never search. The comparison is not “digital vs primitive.” It is when the factory’s jobs become the outage. Manuals for a toaster can live as manuals/toaster.pdf forever. The mortgage packet can live in Paperless and in 2024/house/. Two homes for the few documents that matter is cheaper than a clever tag I will not remember during a refinance.

Search is a feeling. Disk is a number. I watch the number weekly. I do not watch the feeling. The feeling will tell me to scan more. The number will tell me to stop.

What the folder still wins

A dated tree — 2026/taxes/w2.pdf — is understandable to a partner, a lawyer, and a future you without a container. rsync and restic treat it as files. Restore is copy. There is no migrate, no consume queue, no Redis, no Postgres. If the document set is hundreds, not tens of thousands, and you remember where you put things, the folder is the whole product.

The folder fails at search across bad scans, at tags that are not filenames, and at “what did the insurance company send in March.” That is why Paperless exists. It also fails safely: disk full means the copy failed. Paperless disk full can mean a wedged worker, a half-imported document, and a consume directory you are afraid to touch.

Where Paperless earns the RAM

Inbox to tagged archive. Correspondent and document type. Full text on scans that were photos of letters. Multi-user with a real login. Consume from a scanner share. That workflow is why people migrate. It is excellent until the hardware is a Pi that also runs Immich.

OCR is CPU and temporary disk. A duplex 20-page color scan is a spike. A weekend of “I’ll just scan the whole filing cabinet” is a fill-the-volume event. Paperless stores originals plus processed files plus archive files depending on settings. Read those settings before you ingest a cabinet. “Archive original and archive PDF” is a multiplier. The folder had one file. You now have a family.

I put consume, media, and data on a volume I can grow. I do not put them on the OS SSD next to Docker images. I limit parallel OCR workers. I do not scan 400 pages on a Friday on a 128 GB disk. The software is fine. The job is a batch process. Batch processes fill disks. That sentence is the article title.

The jobs that look like “Paperless is down”

Worker crash on a poisoned PDF. Queue backup. Consume folder with a file that never finishes. Index rebuild after an upgrade. All of these look like the web UI hanging. The folder never hung. It also never searched.

Upgrades with migrations want disk headroom. If you are at 90 percent, you cannot upgrade. You also cannot OCR. You are in a corner. The folder would have rsynced to a bigger disk in an hour. Paperless wants a procedure. Have the procedure before 90 percent, not after.

Backup is not the consume folder

restic the media and data volumes, and the database. A copy of consume is not the library. Export/document dump if you need an escape hatch back to a folder. I keep an occasional export of originals so I can leave Paperless if I must. The folder people already live in that escape hatch.

If Paperless is the only copy of a tax PDF, you made the factory the archive. The factory can burn. Dual-write originals to a dated folder if the document would hurt to lose. Paperless can consume a copy. The folder is the legal-feeling tree. Search lives in Paperless. This is boring and it survives a bad migration.

Filename versus tags, and the partner test

A folder forces a naming scheme. That is a gift. Paperless lets tags drift until “insurance” and “Insurance-home” are two worlds. Pick a small controlled vocabulary or you will search like you still have a pile. The partner test is: can they find last year’s car title without you. If the answer is only via your brain, you built a personal database, not a household archive. A folder with spoken rules sometimes wins that test.

Permissions are the other partner test. A share everyone can read is simple. Paperless users and permissions are better and easier to get wrong after an invite. Do not give a kid the consume directory as a toy. They will OCR memes onto the tax volume.

Scanner hardware is part of the disk story

A cheap ADF that produces 600 dpi color of a utility bill will generate monsters. Dial the scanner down. Black and white for text. The folder would have stored the monster too, but you would have noticed the file size in Explorer. Paperless hides the size in a thumbnail until df screams. Look at media volume growth after ten scans. Extrapolate before the cabinet weekend.

Network scanners that dump 50 files at once will stampede workers. Stagger. Or scan to the holding folder and drop ten at a time. The factory likes a conveyor, not a truck backing into the hopper.

When to stay a folder

Fewer than a few hundred PDFs. One human. Good filenames. No desire to run Postgres. A partner who will not learn a web UI. Stay a folder. Add a search tool later if it hurts. Do not install Paperless because a blog made your cabinet feel shameful.

When to run Paperless

You scan weekly. You search monthly. You have a disk that can take 3–5x the originals if you keep derivatives. You will monitor free space like Immich. You will not colocate OCR with a photo ML queue. You have a restore drill for the database plus files.

A small x86 box with rust for media is happier than a Pi 5 sharing an SD card with HA. OCR on a Pi is a space heater and a queue. It works for a letter. It does not work for a cabinet weekend. If the Pi is what you have, ingest ten pages a night and live with the calendar. The cabinet will still be there. The SD card may not if you fight it.

Docker logs for the worker are part of operations. A silent failed consume is a document you think you have. Glance at the queue after each drop. The folder at least showed a new file. Paperless can swallow a file into an error you did not open. Uptime on port 8000 is not a document count.

A consume policy that does not fill the disk

Scan to a holding folder you review. Delete junk before consume. Turn off extra archive copies you do not need. Cap upload size. Watch the worker. After a bulk ingest, check df before you walk away. Schedule OCR so it does not fight a backup window. None of this is in the “five minute install” README. It is the difference between a library and a 3 a.m. disk alert.

The close

A folder of PDFs fails at search and wins at restore. Paperless-ngx wins at search and fails when OCR jobs and derivatives fill the disk you treated like a cabinet. Give the factory a volume, a worker limit, and a copy of originals elsewhere. Or stay a folder until search actually hurts.

The consume directory is not a trash can for a whole house’s paper in one Saturday unless the disk was bought for that Saturday. Buy the disk first. Then run the job. The folder people already knew that. The factory makes it easy to forget.

If you already filled the disk, stop the workers, move media to rust, and do not delete consume while angry. Angry deletes are how originals vanish. The folder people’s worst day is a bad rsync. Paperless people’s worst day is a prune they thought was thumbnails. Know which files are sacred before you free space. Originals are sacred. Derivatives can often be rebuilt. If you are not sure, you are not ready to delete. Copy off-box first. That is the same restore religion as Borg. The UI does not change it.