AI
Subgen 
GPU-backed Whisper transcription for Bazarr,
deployed as the subgen stack (tags = ["ai"]). Uses the NVIDIA 3060 with
CUDA; see the Bazarr section in the media docs for provider and sync settings.
AudioMuse 
I only really became aware of this with navidrome's 0.60 release where it
enabled plugins, creating smart playlists etc. It was something I've been
looking for awhile to analyze my music media and give me cool recommendations
using AI. This did prompt me to setup ollama since using gemini got me rate
limited pretty quickly. But so far the ui is really easy to execute tasks and
using the gpu image with nvidia, very fast.
Ollama 
This turned out to be a really cool setup and I'm still messing around with this to know if I want to keep it. I wanted a way to get better usage out of my nvidia 3060 since it just kind of sits there unless it's used for the Windows KVM I have set up. So I decided to give setting up ollama a spin and well it turns out to be really nice to setup with other tools. For now I've used gemini's recommendation in selecting the particular models to download and I ended up going with this:
- llama3.1:latest -> for general information
- qwen2.5-coder:14b -> for coding
- deepseek-r1:14b -> for general information also
- gemma3:12b -> for general information also (testing this out as my default)
Setting up my nvidia gpu for usage in containers was a whole hassle I'll document at the docker page, but so far the models are pretty fast and seem to be pretty decent. One thing that has been annoying has been the fact that the models are trained on data from 2024 or earlier.
Open-WebUI 
With ollama running, open-webui provides a ChatGPT-style frontend for
interacting with models and tuning the experience. The stack lives at
compose/ollama/docker-compose.open-webui.yaml, runs separately from the Ollama
container, and is exposed at ollama.dripdrop.pro.
Authentication uses Authelia OIDC (OPENID_PROVIDER_URL points at Authelia's
.well-known/openid-configuration). Sign-up is disabled; users authenticate
through Authelia and accounts merge by email. Group management is enabled so
Authelia groups can map into Open WebUI roles.
Downloading models for ollama is as simple as searching for a tag and hitting
fetch. In the admin panel I customized downloaded models with a system prompt
and wired searxng in as a web search tool for better answers.
Ensure to set up searxng with a limited result set length
searxng 
In order to enhance the model experience with ollama I wanted to set up a
search service with open-webui and decided to set this up since it's free. It
was surprisinly easy to setup with just a quick setting tweak or two it was
good. I had to add - json format to the search settings to searxng and it
was working perfectly good with open-webui.