hostinglinks.
PRACTICAL TECH, EXPLAINED

DOCUMENT ARCHIVE

Paperless-ngx is a searchable archive, not a scan-and-forget folder.

Paperless-ngx takes documents, runs OCR, and lets you find them later. The heavy part is the OCR, not the web page. A folder in Google Drive stores the file. It does not, by itself, turn a crooked phone photo into text you can search.

Reviewed 30 October 2026 against the Paperless-ngx setup documentation. How a planning sum differs from a measured archive.

A flatbed document scanner with the lid raised and a USB cable beside it, ready to scan paperwork.
Photograph by Bultro. CC BY-SA 4.0. Resized and compressed to WebP by HostingLinks. Original file.

What belongs in it

Invoices, bank letters, insurance schedules, warranties, tax letters and receipts are the usual set. Paperless watches a consume folder, and it can also take mail if you configure that. The setup guide expects a Redis broker and a database. PostgreSQL and MariaDB have to be selected explicitly with PAPERLESS_DBENGINE. SQLite is available. For a pile of paperwork you cannot replace, prefer a server database you know how to dump.

OCR is Tesseract, through OCRmyPDF. The bare-metal dependency list asks for Tesseract 4 or newer and the language packs you actually read. Set PAPERLESS_OCR_LANGUAGE to that language, and set PAPERLESS_TIME_ZONE to a real zone such as Europe/London. Set PAPERLESS_SECRET_KEY to a long random string. The project warns that a missing secret lets other people forge credentials. Behind a reverse proxy, set PAPERLESS_URL to the public address.

If the consume folder is on NFS, inotify will not see new files. Set PAPERLESS_CONSUMER_POLLING_INTERVAL to a positive number so Paperless polls instead. Extra OCR languages and a rootless container do not mix: the docs say the container cannot run rootless when PAPERLESS_OCR_LANGUAGES adds languages.

A cloud folder and an OCR archive

A synced folder beside Paperless-ngx
QuestionDrive, Dropbox or similarPaperless-ngx
Can you find a word inside a scan?Only if that product OCRs it, and you trust the resultOCR on your server, with a language you configure
Who runs the disks?The provider, inside a quotaYou
Original fileWhatever you uploadedKept, plus an optional archive PDF
Search after you delete the accountYou need an export firstYou need a backup of media, data and the database

A cloud folder is a reasonable place to drop files quickly, and a weak place to prove you can still open last year's tax letter after the subscription lapses. Paperless is the opposite shape: more setup, a library that stays useful if you keep the backup. Many people use both. Scan into Paperless, and keep an encrypted copy of the originals off site. The off-site copy is the backup. The consume folder is not.

Three picks

Stay with the cloud folder

Choose this when you have few documents, the provider's search is good enough, and you will not run a server. Export a copy once a year anyway.

Paperless on a machine that already stays on

Choose this for a household archive. The project does not publish a single RAM minimum. OCR is the spike. The Raspberry Pi notes in the setup guide are the closest hardware advice: Pi 3 and later have four cores, and the default worker settings can make the web interface sluggish while a document is consumed. Their example is to lower PAPERLESS_TASK_WORKERS and PAPERLESS_THREADS_PER_WORKER, for example to 2 workers and 1 thread, so the machine can still answer the browser.

A scanner that OCRs before Paperless sees the page

Choose this when the Pi or the mini PC feels slow. The docs say to leave PAPERLESS_OCR_MODE at auto and to consider OCRing on the scanner. Paperless then has less to do. A flatbed or a sheet feeder is a tool, not a requirement. Phone photos work. Straight, readable pages work better.

Settings that change speed and disk use

On a small board the setup guide suggests several specific levers. Set PAPERLESS_WEBSERVER_WORKERS to 1 under Docker to save memory. Set PAPERLESS_OCR_CLEAN to none if you OCR on the device and you want less memory and less time, accepting a slightly worse cleanup. Set PAPERLESS_OCR_PAGES to 1 if the first page is enough to find the document. Set PAPERLESS_ARCHIVE_FILE_GENERATION to never if you do not need a PDF/A archive beside the original. That last one saves disk and loses the in-browser archive view. Turn off the folder consumer with PAPERLESS_CONSUMER_DISABLE if you only import by hand or by mail.

The classifier that suggests correspondents and tags can use a lot of CPU when it retrains. If it crowds out everything else, move it to a daily schedule in the admin. You can also run it once by setting the next run to now.

Four processes have to stay up on a manual install: the web server, the consumer, the task queue and the scheduler. Docker Compose starts those for you. If the queue is dead, documents sit in the consume folder and look like a broken scanner. Check the worker before you reseat the USB cable.

Secrets can be files. Append _FILE to a setting, for example PAPERLESS_DBUSER_FILE pointing at a Docker secret, instead of putting the password in the environment of every process listing.

A page-size sum you replace with a real scan

An A4 page is 8.27 by 11.69 inches. At 300 dots per inch that is about 2,481 by 3,507 pixels. Uncompressed 24-bit colour is 2,481 times 3,507 times 3 bytes, about 24.9 MB. You will not store that. A scanner's JPEG or PDF of a text page is far smaller. Assume 400 KB a page until you measure one. Two thousand pages a year at 400 KB is 800,000 KB, about 0.76 GB of originals. If you also keep an archive PDF of similar size, plan about 1.5 GB a year of files, then the database, then a second copy.

Measure the assumption with a dozen real pages:

du -sh ./consume ./media
ls -l ./consume | head

If those pages are 1.2 MB each because they are colour photos of receipts, use 1.2 MB, not 400 KB. Ten years of 2,000 pages at 1.2 MB, with a second archive file, is about 48 GB of files before backups. That still fits a small SSD. The mistake is to fill the system disk and then be unable to install an update. Keep media on a data disk once the archive is past a toy size. The document archive project puts this in a staged order.

Backup and a restore drill

You need the originals, the data directory, and the database. Redis is a broker. It is not the archive. Copying only the consume folder misses everything already imported, because those files have moved into the media store.

Stop the stack or accept a quiet moment, dump the database, and copy the media and data directories to a disk that is not the server. A cloud copy of that bundle is the off-site half. Then restore onto a spare folder once, search for a word you know is in a letter, and open the PDF. Our backup guide explains why a sync folder that can delete the scan is not this copy.

Optional language-model features are a separate decision. The core OCR path described here runs on your machine. If you later enable a feature that sends document text to a model, read that feature's current page and treat the letters as data you are choosing to transmit. Do not enable it to make the first archive "smarter".

Check that OCR found the words

Import five pages you already know: a utility bill, a letter with a reference number, a receipt, a warranty and a page with a date written as digits. Search for a word that appears in the middle of the bill, not the company name in the title you typed yourself. If the search misses it, open the document and look at the extracted text. A skew, a photo of a screen, or the wrong PAPERLESS_OCR_LANGUAGE is the usual cause. Fix the scan before you feed the rest of the box. Five bad pages are a setting. Five hundred bad pages are a weekend.

Give each document a correspondent and a type once the text is trustworthy. The suggestions Paperless makes are a starting point. They are wrong often enough that a bulk accept, without looking, files a tax letter under the supermarket. Spend the first month tagging as you go. The archive becomes useful when the tags match questions you actually ask, such as "insurance" or "2025-26 tax", not when every automatic tag is kept.

Email import is optional and it puts a mailbox password in the configuration. Use an app password on a mailbox that only receives scans, not your personal inbox. A rule that forwards every attachment will also forward newsletters and other people's PDFs into the archive. Point it at a dedicated address, then watch the first week of imports before you trust it.

Questions before the first box of paper

Does it need a GPU?

No. OCR is a CPU job. A Pi can do it, more slowly, with the worker limits above. A recent mini PC finishes a page faster and still does not need a graphics card.

Can I point it at a network share?

Yes for storage, with polling turned on if the share cannot notify the consumer. Keep the database on a local disk if you can. A share that disappears mid-OCR produces partial documents and a confusing queue.

What about handwriting?

Tesseract is for printed text. Handwritten notes will often fail to search. Store them, and add a typed title or a tag so you can still find the scan.

Is the consume folder the backup?

No. It is an inbox. Empty it by letting Paperless import, then back up the library. Deleting the inbox after a successful import is normal. Deleting the media directory is not.

Verdict

Use Paperless-ngx when you want to search the paperwork you are obliged to keep, and you will back up media plus the database. A cloud folder is enough when search inside scans does not matter. On a Pi, use the project's own worker and OCR limits rather than hoping four cores will do everything at once. Replace the 400 KB planning figure with du on your scanner's real output before you buy a disk.

Sources checked 30 October 2026: Paperless-ngx setup, including the Raspberry Pi and OCR sections.