Paperless-ngx with local AI: Ollama setup, paperless-gpt and the GPU your archive needs

Paperless-ngx AI in version 3 with a local model

Paperless-ngx 3 can tag scanned documents automatically and suggest titles, correspondents, document types, storage paths and dates with a language model, and it can answer questions about your archive in a chat that links to the source documents. Both run against a model in your own office when you point Paperless at Ollama on a workstation with an NVIDIA GPU. One RTX PRO 4000 Blackwell with 24 GB is enough for a small office archive by our estimate; larger archives and better answers call for a 48 GB card.

Paperless-ngx is an open-source document management system under the GPL-3.0 licence. It takes scans, PDFs and e-mail attachments, recognises the text with OCRmyPDF and stores everything in a searchable archive. Documents arrive through a watched folder that a scanner can write to, through upload in the browser, through its REST API, or from mail accounts that Paperless checks every 10 minutes by default. For years it has sorted documents with a classifier: a small machine-learning model that learns from the tags and correspondents you assign by hand.

Version 3 adds the language model on top. Its release notes list "Feature: Paperless AI" for 3.0.0 of 22 July 2026, a major release with breaking changes, so an existing 2.x installation is upgraded along the project's v3 migration documentation. The documentation describes the AI features as optional and "off by default"; they do not replace the classifier, and both kinds of suggestion can run side by side. The project is actively developed: its releases page lists five releases in October 2026 alone. This article covers what the built-in AI does, where the add-ons help, and which graphics card runs it at a usable speed. For the general idea of chatting with company documents, see our guide to RAG on company data.

What the built-in AI does and where it stops

The documentation and the source code describe four functions. The first two need only a chat model; the others need an embedding model as well, which turns each document into a vector so that similar texts can be found.

Paperless-ngx documentation (usage, advanced usage, configuration) on the main branch, read 30 October 2026.

The chat has limits worth knowing before you promise it to colleagues. In the current source code, each question retrieves five text chunks (CHAT_RETRIEVER_TOP_K = 5) and an answer cites at most three documents (MAX_CHAT_REFERENCES = 3). Each question goes to the model on its own, without the earlier questions of the session. It therefore suits lookups such as "which contract with the cleaning company is current, and what is the notice period". Questions that need every document of a kind, such as a total over all invoices of a year, sit outside what five chunks can carry.

The workflow action fills only the fields you select. By default it creates no missing tags or correspondents, overwrites no existing values and adds tags without removing any. It does not run on the "Consumption Started" trigger, because the text does not exist yet, and every matching document sends one request to the model in the background queue, which can delay other tasks on a large import. Custom fields are not among its targets.

Setting up Paperless-ngx with Ollama

Ollama is a program that downloads open models and serves them on the local network. Paperless talks to it directly. The table lists the settings that matter, as Paperless's configuration reference gives them.

Paperless-ngx configuration reference, main branch, read 30 October 2026. With the Hugging Face backend the embedding fallback is sentence-transformers/all-MiniLM-L6-v2.

Three choices decide the quality. Set the chat model explicitly, so that Paperless does not fall back to an older model, and set PAPERLESS_AI_LLM_OUTPUT_LANGUAGE if suggestions should come in one language. Choose the embedding model before the first index: the project wiki calls the Hugging Face default "a small, fast choice for primarily English archives" and names multilingual-e5-small, multilingual-e5-base and bge-m3 for other archives, as "starting points, not benchmarks". bge-m3 in Ollama is a 1.2 GB download that its page describes as supporting "more than 100 working languages". If you change it later, the administration guide asks for document_llmindex rebuild, which rebuilds the index from all documents. And set the OCR languages to match your mail: the documentation's example is deu+eng, with the note that Tesseract, the OCR engine, "uses much more CPU time with multiple languages enabled".

Ollama's own defaults matter as well. Its FAQ states that it keeps a model in memory for 5 minutes after the last request, loads up to three models per GPU when they fit, and processes one request per model at a time unless OLLAMA_NUM_PARALLEL is raised. For an office archive that is usually fine; for a company-wide chat it is not, as our guide on moving from Ollama on a PC to a company AI server explains.

paperless-gpt vs paperless-ai for poor scans and custom fields

Two community projects existed before the built-in AI and still fill gaps. Both connect to the same local model server.

paperless-gpt (MIT licence) reads poor scans and handwriting with a vision model, a language model that looks at the page image, and writes the recognised text back to Paperless. Its examples use MiniCPM-V, whose Hugging Face card allows commercial use of the MiniCPM-V 2.6 weights only after a registration questionnaire. We plan with Qwen3-VL 8B instead, a vision model under Apache 2.0 that is a 6.2 GB download in Ollama and whose card states text recognition in 32 languages; paperless-gpt can also use Docling Server as a self-hosted OCR and conversion service. It also fills custom fields, the gap the built-in AI leaves, and offers a review screen where staff approve or correct suggestions. Its README states that it was tested against Paperless-ngx 2.20.x and the 3.0.0 beta. It has no login of its own and warns: "Do not expose it directly to the internet or an untrusted network."

paperless-ai (MIT) offered automatic processing and a chat before version 3 had them. As of October 2026 its README says: "This repo is currently not maintained." For a new installation we would not plan on it.

You can order the workstation alone, the setup of Paperless-ngx, Ollama and the models alone, or both together: the workstation delivered with everything installed and configured; write to us and we suggest the best solution for your tasks and budget.

Which GPU for Paperless-ngx AI

Paperless itself, its database and the Tesseract OCR run on the processor. The graphics card serves the models: the chat model for suggestions and chat, the embedding model for the index, and, with paperless-gpt, the vision model for OCR. Everything that Ollama keeps loaded must fit in the card's memory, and the download size of a model is the lower limit, because each request adds its context on top.

We size with our users-per-card rule from the guide to how many users one RTX PRO 6000 serves: 90 per cent of the memory the driver reports, less 3 GiB for the runtime. That gives 18.5 GiB on an RTX PRO 4000 with 24 GB, 25.7 GiB on an RTX PRO 4500 with 32 GB, 40.0 GiB on an RTX PRO 5000 with 48 GB and 83.0 GiB on an RTX PRO 6000 with 96 GB. The model sizes are the download sizes in Ollama's library, read on 30 October 2026: gemma4:12b 7.7 to 8.0 GB, gemma4:26b 16 to 19 GB, bge-m3 1.2 GB, qwen3-vl:8b 6.2 GB. Google's model card lists Gemma 4 under the Apache 2.0 licence and states support for more than 35 languages out of the box.

Our estimates by the users-per-card rule, not measurements; model sizes from ollama.com and our gpt-oss-120b guide.

The arithmetic for the entry tier: 8.0 + 1.2 + 6.2 = 15.4 GB, or 14.3 GiB, which leaves about 4.2 GiB of the 18.5 GiB for the context of running requests, by our estimate. On the 48 GB card the larger Gemma 4 with the same helpers takes 26.4 GB, or 24.6 GiB, and leaves 15.4 GiB. A 32 GB RTX PRO 4500 sits between: it holds gemma4:26b with bge-m3 and 6.9 GiB to spare, but adding the vision model leaves 1.1 GiB, too little for comfort. Without paperless-gpt, the 24 GB card has more room for context.

At the mid tier, Paperless uses the openai-like backend and points at a vLLM server that other tools use as well. gpt-oss-120b takes 60.8 GiB and leaves 22.2 GiB for the cache on one RTX PRO 6000 Workstation Edition or Max-Q, room for about 19 conversations of 32,768 tokens by our gpt-oss-120b guide, and each Paperless request of 8,192 tokens needs a quarter of one such conversation's cache. Larger servers with the RTX PRO 6000 Server Edition, or an L4 for embeddings alone, fit when Paperless is one tool among several, as in our guide to the embedding and reranker server for RAG.

We found no published speed figure for Paperless-ngx with a local model on these cards. Speed depends on the card's memory bandwidth (672 GB/s on the RTX PRO 4000, 1,344 GB/s on the RTX PRO 5000, 1,792 GB/s on the RTX PRO 6000 Workstation Edition), on the model size, on how many documents arrive per day and on whether vision OCR reads every page. Importing a backlog of years is the heaviest job: plan it overnight or over a weekend, and see our guide to vision-language models for document AI for published pages-per-second figures.

Running Paperless-ngx with AI safely

The project's README is direct: "Paperless-ngx should never be run on an untrusted host because information is stored in clear text without encryption." Run it on a machine in your own building or rack, give it a backup, and use the document_exporter command, which exports the documents, settings and database contents for backup or migration.

Keep the model server inside. A native Ollama installation binds to 127.0.0.1 on port 11434 by default, while the Ollama container listens on all of its interfaces; run it in the same Docker Compose project as Paperless without publishing port 11434, and Paperless reaches it inside the Compose network while the port stays closed on the office network. The same goes for paperless-gpt, which listens on all interfaces by default. If staff work from outside, use the VPN you already have.

Two settings send data out. Remote OCR is off by default and sends document files to a cloud OCR service when switched on; for a local setup it stays off. An openai-like backend pointing at a cloud provider sends document text there, and the documentation warns that remote providers receive document content. With Ollama or vLLM in the building, the text, the index and the chat stay on your hardware. The documentation also treats document content as "untrusted data": a scanned letter could contain instructions aimed at the model, so keep "create missing items" and "overwrite" off in the workflow action unless you review the results. Our guide on local AI and the GDPR explains what own hosting solves; your data-protection officer confirms what your firm needs.

How we set up Paperless-ngx with AI

When we set up the software, the workstation runs Paperless-ngx with Docker Compose, the install method the project documents, with Ollama on the GPU and the models already downloaded: a chat model sized to the card, bge-m3 for the index and, where scans are poor, paperless-gpt with a vision model whose licence fits your use. The OCR languages, the watched folder for the scanner and the mail rules are set for your documents, and the backup export is scheduled. If your company also wants a chat over its documents such as AnythingLLM, a local NotebookLM alternative, AI search across SharePoint and Confluence with Onyx, document translation with the layout kept or AI that sorts the shared inbox, the same GPU or a server runs them together; our guide to preparing company documents for an AI assistant covers the groundwork.

We deliver workstations with one RTX PRO 4000, RTX PRO 5000 or RTX PRO 6000 Workstation Edition for Paperless-ngx; send us your document volume and the languages of your mail, and we propose the card that fits.

FAQ