Paperless-ngx 3.0 Brings AI to Your Document Archive

Paperless-ngx 3.0 is here, and the headline is impossible to miss: Paperless AI. For the first time, AI is a first-class part of the document pipeline — not a bolt-on experiment, but a set of features woven through the app. It can suggest titles and tags, fill in correspondents and document types, populate custom fields, and even let you chat with your archive.

Here’s what that means in practice, and why it changes how a Paperless library feels to use day to day.

Meet Paperless AI

Version 3 adds an AI layer that reads your documents and helps you organise them. The centrepiece is AI suggestions: when a document is consumed, the AI proposes metadata for it — a sensible title, matching tags, the right correspondent, the document type, and values for your custom fields. You review and accept with a click, so manual data entry shrinks to a quick glance.

Crucially, it respects language. The suggestion engine tries to match the language of each document, so a German invoice gets German suggestions and a French contract gets French ones.

Chat with your documents

The standout feature is document chat. Instead of only searching for keywords, you can ask your archive questions in plain language and get answers grounded in your actual documents — a retrieval-augmented generation (RAG) setup that points you back to the source. Chat responses include links to the documents they drew on, so you can jump straight to the proof.

A few examples of what this unlocks:

• “How much did I spend on car insurance last year?”
• “What’s the renewal date on my internet contract?”
• “Summarise the warranty terms for the dishwasher.”

It turns a folder of paperwork into something you can actually converse with.

Bring-your-own-key models

Paperless AI needs a language model (LLM), and that model is entirely yours to choose. It works on a bring-your-own-key basis with any OpenAI-compatible provider: pick the cloud model that suits you best (OpenAI and the like), enter your own API key, and suggestions and document chat work straight away. The model costs stay directly between you and your provider.

The embeddings behind document chat stay in a local vector index, and the LLM settings (chunk size, context window, timeouts, language) are configurable so you can balance quality against cost.

Why this matters

The promise of “paperless” was never really about going digital — it was about making information retrievable. Search got us part of the way. AI takes the next step: your archive doesn’t just store documents, it understands them well enough to label them and answer questions about them.

For anyone who has stared at a freshly uploaded stack of scans and dreaded the tagging, AI suggestions remove the friction. And for anyone who has known the answer is somewhere in 2,000 documents but not where, document chat is the shortcut.

Try it without the sysadmin work

Paperless AI is powerful, and like everything in self-hosting it comes with knobs to turn and a model to pick. A hosted Paperless-ngx instance means you get the 3.0 features — AI suggestions, chat, faster search, the new interface — without standing up a server. You just add your own LLM API key and start asking your documents questions.

Pair this with our look at the much faster search in 3.0 and the other everyday improvements.