The task
Every scanned document in Paperless-ngx needs a title, tags, and the right correspondent, and these are usually typed by hand. The built-in classifier learns from earlier choices, but the source says much of the typing stays with the user at the start and for unusual documents. A fresh scan arrives with nothing but the scanner's file name.
What they did
paperless-gpt runs as a separate service next to Paperless-ngx and connects through its REST API with an access token. It fetches a document, sends the text or the image to a language model, and writes the result back as metadata. The model can be a local Ollama server or a cloud provider. Control runs through tags created in Paperless-ngx. A person, or a Paperless-ngx workflow acting on a filtered set of new documents, adds the review tag paperless-gpt. Shortly afterwards the paperless-gpt web interface lists suggestions for title, tags, correspondent, document type, and date, each shown next to the current value. A detail view puts the scan and the recognized text beside the suggestions. For difficult scans such as phone photos, the service can also run text recognition with a vision model, started by a person from the OCR tab. The source advises starting with the review tag to see how well the suggestions fit the documents and to adjust the prompts.
What the source reports
According to the article, in a test run with a crooked photo of a receipt, the built-in Tesseract recognition dropped the amounts in the total lines and the vision model read them. The article also shows a scan that arrived under the scanner's file name and became an electricity bill with a correspondent, a document type, and tags after the suggestions were approved.
Where the approval sits
Documents with the review tag are held in the paperless-gpt interface. A person sees the old and the suggested value side by side, removes tags, corrects the title or date, and approves. According to the source, every change is recorded in a history and can be reverted field by field.
Worth knowing
The source is a hosting provider's article, and the provider sells managed Paperless-ngx with paperless-gpt, so it has a commercial interest. It reports a single test run and no measured results. It says that privacy depends on the model: a local Ollama server keeps documents on the owner's own infrastructure, while US cloud models receive their content. It also says paperless-gpt has no login of its own, so a self-hosted interface should never be exposed to the open internet. It also covers an auto tag that processes documents without review; this entry covers only the review path. The article does not describe a review step for the text recognition output. This entry describes a setup and is not legal advice on data protection.
Source: server.camp · Sep 2026
Drafted with AI assistance and checked automatically. A person is responsible for the collection. How these entries are made
Prompt for your agent
Propose a GoBD-friendly filing scheme
Get a file name pattern and folder tree for your business documents, built for GoBD-friendly filing, and review it before you use it.
More in this category
- Back office and documents
A blogger tests Paperless-ngx 3.0 suggestions and accepts each by hand
A tech blogger connected Paperless-ngx 3.0 to a language model and tried its Suggest button, which proposes tags, a document type, a correspondent, and a title. Nothing changes until a person accepts a suggestion.
Read the entry - Back office and documents
A sole proprietor books receipts with Claude Code and approves each
A one-person business built a Claude Code skill that reads receipts from its Fiken inbox and proposes an account for each one from past bookings. The owner confirms every proposal before anything is booked.
Read the entry
Ninety minutes live on your tools. One automation running before we hang up.