Paperless-ngx is an extensive document management system. 3.0.0 has just been released.
It seems from a user/admin perspective that a lot of obsolete stuff under the hood got removed and they did some performance stuff, but it’s also full of AI features. There’s too many changes to read them all though, I wonder if there’s a blogpost or something somewhere summarizing the main changes.
Personally I’ll stay with 2.X for some days at least to see what people say about the new version. I’m not interested to have AI parse my documents, the alrogirthmic auto tagging etc works quite well.
You must log in or # to comment.
At the risk of getting internet-assaulted I am one of the primary devs, feel free to ask me about it. To address some of the issues already raised:
- All the new “AI” (LLM) features are opt-in, meaning if you dont enable them you wont see them, literally. At all.
- All of them can also be run local-only, of course.
- There are a ton of other features as well.
- One of our devs has made use of Claude especially around database schema improvements, which is a particularly tricky thing to work on. I dont like AI, I wish I could go back to the world before it, but I cant. So at this point it is almost absurd to not use LLM tools to some degree in programming. AI slop or vibe-coding is to do so indiscriminately, and I hate that too. Believe me we get tons of PRs from people with basically no dev experience or abilities slinging up a Claude PR they dont understand… and we dont accept them. We expect the same from ourselves. We use it sparingly and very specifically, interrogate it aggressively and stand behind it after it’s merged. To somehow object to any / all use of it at this point either implies a) you dont do programming or b) you dont understand the nuances of how tools can be used in programming. Used properly, it’s no more offensive than using an IDE instead of a bare text editor, is that something you care about?
- Also, leaving Claude as a co-author I would argue seems more transparent, not doing so or not disclosing it in the PR seems like the shadier choice.
- The ethical concerns about LLMs, compute use etc are real, but I dont think one OSS project somehow needs to bear the entire weight of this societal problem.
Again, Im happy to discuss more, as long is its not the guy who opened the raving “issue” in our project.
Thanks for chiming in! I don’t fully agree with your reasoning re the usage of AI but I can accept your stance and am glad that you implemented it opt in with no intention of annoying me into using it. I also know many people have been asking for those features for a while though.
The person creating the issue has been downvoted here as well, even though this is a largely anti-ai place in my perception.
By now I also found the upgrade guide which lists the main features - I usually read the release notes, but this time it was a whole lot and hard to take it all in.
Respect for the nuance of the argument. Congrats on the major version.
Hey, dont believe me, Im just a random, but this guy seems like worth a listen: https://arstechnica.com/ai/2026/07/linus-torvalds-to-critics-of-ai-coding-in-linux-fork-it-or-just-walk-away/
No I’m saying I respect the fact you have more nuance than most towards the argument and didn’t leave it in such a reductive form that doesn’t help anyone.
WHAT IS THIS?!?!1? NUANCE IN A DISCUSSION?!? NOT ALLOWED IN 2026
/s
flinging upvotes to y’all, really makes my day seeing civil discourse and not “Discourse” on here. Keep it up, really this is what makes Lemmy such a great place
I run paperless at my house and it’s given me back some sanity and some privacy. That’s a rare combo in 2026. From one techie to another: thank you very much. My actual family thanks you haha.
Thank you for the amazing work. This is, along side immich, my most used self hosted project. I’ll definitely give this a look over this weekend!
database schema improvements, which is a particularly tricky thing to work on
Not, if you know what you’re doing.
I dont like AI, I wish I could go back to the world before it, but I cant.
You do know that you can use LLMs to get inspiration and suggestions but then still implement those things by hand? Which is also a great way to make sure it doesn’t change something else in the background. Also, if you use Claude like you’ve used StackOverflow and Google before, you don’t even have to mention it as a Co-Author.
I disagree with your broad statement that SQL in general isn’t tricky to make more performant if you know what you’re doing. Worked with DBAs who have decades of experience who have gotten great help in streamlining queries and schemas from AI tools in the last couple of years.
I may be mistaken, but I had looked into Paperless before and saw they had AI features before this. Is it just more, or more in your face? I was under the impression they were optional, at least when I looked into it.
I had a quick look and the AI features that are added:
- LLM driven tagging suggestions
- Remote OCR (Azure AI)
They do look to be optional at least. I’d probably be a bit more concerned if the code was largely written with LLMs, given the large number of commits in what looks to be only a few months.
FUCK IT THEY GOT PAPERLESS-NGX TOOO???
Fedis, what’s the best alternative?
what’s the best alternative?
You’re asking for suggestions for an automatic document categorization tool that doesn’t use any automatic document categorization?
You don’t need a knife to make a knife, and things like OCR are different from AI (or from “AI”).
Paperless and Papermerge have always done a lot more than just OCR. If that’s all they were, most people can do that already on the software that comes with their scanner. The core selling point of these applications is automatic categorization, sorting and tagging, and those have always relied on machine learning tools. Literally the first thing you do after setting Paperless up is start training the AI.
Yes, though that used to be discriminative AI, where now generative AI is used if I’m not mistaken.
I think you’re under the impression that the difference between those things is far greater than it actually is. Large Language Models work by developing statistical maps of associations. That’s discrimination. They’re a direct evolution of categorization models. The ability to associate a hash of a JPEG with “cat” is the same as the ability to associate “How are you?” with “Great, how about you?” It’s all associative mapping. LLMs are just the current leading edge of that technology. If you want to, for example, generate a list of tags that describe a document, an LLM is the best tool we currently have for doing that.
To put it another way, what you term “discriminative AI” is “generative AI.” It’s generating a category or list of categories in response to an input. That’s not functionally different than generating a sentence in response to a sentence, it’s just an order of magnitude less complex. You can argue terminology but the technology exists on an evolutionary curve, with no real hard boundaries.
Goddamnit. I hope someone forks 2.20.
Why? The new LLM features are opt-in. Just don’t turn it on if you don’t want them. Or is there something else you don’t like?
If they weren’t using LLMs to code, I’d mostly agree with that. But projects that put effort into integrating LLMs likely use them as well (as is the case here).
FYI: they use Claude Code.
Just in case anyone has an issue with that.
Pro-tip: Block the user “claude” on GitHub and it’ll show you a warning for every repository where they forgot to remove the Co-Author tags:
Lots of people leave the tag in for transparency. The world must feel like such a conspiracy for you.
