# Hardware & local LLM advice for Proxmox (Paperless + personal finance tracker) — retained record

> Retained reddit record from reddit:homelab.

Canonical: https://highsignal.app/data/records/2a22398fb511c9cb

[← sources](https://highsignal.app/data)


source: [reddit](https://highsignal.app/data/reddit) (reddit:homelab)  published 2026-10-23 09:43 ingested 2026-10-24 02:39

[original source ↗](https://www.reddit.com/r/homelab/comments/1xd9kgr/hardware_local_llm_advice_for_proxmox_paperless/)

## retained text



```
Hi everyone,  
I run a small Proxmox server at home with Paperless-ngx for document management and a custom-built service for tracking my family's finances. I'd like to add local AI to both, and nothing should leave my network.

**Current bottleneck:** the server is a Geekom A6 mini PC with 16GB of RAM, and it's almost full. The RAM can be upgraded, but I'm also open to adding a second node, either a PC with a dedicated GPU (e.g. a used RTX 3090) or an Apple Silicon machine (e.g. a Mac mini). I don't have a fixed budget, but I'd like to spend as little as possible.

**Use case 1 – Paperless**

I want the model to analyze each document and fill in its metadata following my own rules:

**- Title** (based on a naming convention I define)  
**Tags**  
**- Document date** (the date on the document, not the scan date)  
**- Correspondent**  
**- Document type** (e.g. invoice, payslip, medical report, contract)  
**- Storage path**

The documents are **in Italian**: payslips, invoices, receipts for warranties, medical reports, contracts, insurance papers, etc. The volume is low, about 20 documents a month or fewer. Speed isn't important, a nightly batch job is fine. I've seen paperless-ai and paperless-gpt. Which one handles all of these fields best, and with which models?

**Use case 2 – Personal finance**

I already have a script that anonymizes my monthly bank statements (about 40 transactions each, income and expenses) and outputs a clean, consistent CSV. Everything here is **in Italian** too: both the transaction descriptions in the CSV and the questions I'll ask the chatbot. I'd like a local model to:

\- classify transactions into a predefined set of categories  
\- act as a chatbot to query the data (e.g. "how much did I spend on the car in 2025?")

Right now I do the classification with a cloud LLM and the quality is good. I'd like to get something comparable locally. Speed isn't critical here either.

**My questions:**

1) Given the low volume, what's the cheapest hardware that can handle this? Should I just upgrade the mini PC's RAM, or is a second node (GPU PC or Apple Silicon) worth it?

2) I'm not very experienced in choosing models. Which local LLMs would you recommend for each task, given that everything is in Italian? I'm happy to run different models, e.g. a lighter one for Paperless and a stronger one for finance analysis/chat.

3) For the finance chatbot, which architecture works best? I assume a small model shouldn't do the math itself. Is it better to have it generate SQL/pandas queries and let the code do the calculations?

Thanks in advance!
```



document fetched 2026-10-24 00:36 · canonical https://reddit.com/r/homelab/comments/1xd9kgr/hardware_local_llm_advice_for_proxmox_paperless
