n8n + Ollama: self-hosted AI automation that actually works
from quickbitesdev@discuss.tchncs.de to selfhosted@lemmy.world on 07 Apr 06:02
https://discuss.tchncs.de/post/58010663

Been running n8n with Ollama for a few months now for work automation. Wanted to share what I’ve learned since it’s not super well-documented.

The setup is just Docker Compose with n8n + Ollama + Postgres. n8n’s HTTP Request node talks directly to Ollama’s REST API — no custom nodes needed.

What I’m running:

Zero API costs, everything stays on my server. If anyone wants the workflow templates I have a pack: workflows.neatbites.com

Happy to answer questions about the setup.

#selfhosted

threaded - newest

irmadlad@lemmy.world on 07 Apr 06:16 next collapse

I really like n8n. It appeals to my visual sense which makes up for a lot of hard programming experience. I don’t run it full with the AI aspect. Not because I have some agenda against AI, but that my equipment is not good enough to run AI efficiently. I use it for a lot of automation around the lab.

jeena@piefed.jeena.net on 07 Apr 06:36 next collapse

What is n8n?

captcha_incorrect@lemmy.world on 07 Apr 06:41 collapse

n8n.io

It is an automation platform with a selfhosted tier.

coffelov@lemmy.ml on 07 Apr 07:28 next collapse

What model do you mostly use for those tasks

tofubl@discuss.tchncs.de on 07 Apr 09:03 collapse

I’ll piggyback onto this question: With the models you use, how do they compare to current models from the big players?

reptar@lemmy.world on 07 Apr 08:13 next collapse

Happy to answer questions about the setup.

Tell me about the hardware, please and thank you.

mental_block@lemmy.wtf on 07 Apr 10:01 next collapse

Piggybacking too as I am considering the same. Please OP and thank you.

And what model class are you using? Lightweight (2B), reasonable ~10B or above 32B?

Do they load fast?

I had a look at NetworkChucks setup and don’t think I can afford an overpowered rig in this economy. Depending on the rig, may have to wait >20s for a prompt answer.

Thank you again!

frongt@lemmy.zip on 07 Apr 10:11 collapse

I was playing with ministral-3 3b on a 3060. It loads pretty quick, but response generation is a bit slow. It starts responding nearly instantly once the model is loaded (which is also quick), but for long responses (~5 paragraphs) it may take 15-20 seconds for the whole thing.

surewhynotlem@lemmy.world on 07 Apr 10:17 collapse

Cries in 1070

frongt@lemmy.zip on 07 Apr 10:30 next collapse

I’d still give it a shot. A quick check of benchmarks suggests it’s not that much slower. I don’t know if that extends to ML computation though.

CCMan1701A@startrek.website on 09 Apr 03:28 collapse

I run llms using a 780m you’ll be fine. I get pretty close to 10 tokens a second for larger 20B+ models.

clifmo@programming.dev on 08 Apr 17:29 collapse

I do something similar with the base model m4 Mac mini. It’s my inference box right now, it handles Immich ML, photo prism AI, and runs Ollama talking to a small web app I call to summarize things. It’s summaries are shit. The bigger the model, the more it hallucinates. So I settle for 1B and 4th grade responses

warmaster@lemmy.world on 07 Apr 09:26 next collapse

Has anyone tried ActivePieces? How does it compare?

irmadlad@lemmy.world on 08 Apr 05:37 collapse

Briefly. I didn’t like it as much as I like n8n. Perhaps it was not suitable to my use case. I hear a lot of good things about ActivePieces tho. You know, give it a spin and see if it gehaws with your flow. From what I understand, both can acomplish about the same. I think ActivePieces is geared more towards cloud deployments whereas n8n keeps things local.

TheHolm@aussie.zone on 07 Apr 20:03 collapse

Ollama has long history of exploits. PLease do not feed anything which come from outside to it.