Add support to LLM models

This is just an idea to use LLMs to define words and expressions. It should be straightforward to implement, add a new search type (e.g. "LLM"), add tab for it preferences where user can set a URL, model name, token (optional) and prompt (predefined but editable).

As a proof of concept I have put together a python script that acts as a DICT server and calls ollama. It seems to work quite well, even on my not so powerful laptop it spews results with acceptable time (6~12 seconds, gemma3:4b).

If interest picks up I might implement this directly into xfce4-dict, no need for an intermediary server.

assigned to @andreldm

changed the description

Sounds like a good idea, but I can't test the script here:

Ollama Model (gemma3:4b)
Error connecting to Ollama: Errno 111 Connection refused

Are you running ollama locally? If not you need to install it (there is an arch package available) and run the appropriate model.

Ah ok, it repeatedly fails to download the model for me, and since it more or less starts from scratch each time (not really downloading, but parsing the data again), and it takes a lot of time and resources, I give up. But I still think it's a good idea! (maybe for those with a more powerful machine and internet connection than mine, though...)

That's odd, ollama always worked for me without any fuss, in case you wanna give a last try:

ollama start
# open another terminal window
ollama pull gemma3:1b
python server.py --model gemma3:1b
# open xfce4-dict
# set localhost in Preferences -> Dictionary Server -> Server

Either way, thanks for trying, really appreciate.

ollama pull gemma3:1b

Ah, that's better. I had done ollama run gemma3:4b and it had gone on to download 3.5 GB of data, whereas here it "only" downloaded 800 MB. Removing my VPN also helped apparently.

On the other hand, it's too resource-hungry for my poor computer! 5m15s letting it do just that, knowing that it also needs 1.3 GB of RAM.

Also, in the end, I learned that we were once at Red Hat, once at Debian 😅

Cool, to be honest, I knew your computer is not that powerful and was not expecting the model to run at all, glad I was wrong :D

One interesting point I noticed from your screenshots is that the response language is determined by the prompt, it doesn't matter which language dict is set to use, I'll try to fix that.

When I have the time I will integrate this feature into dict, it's at least fun and maybe someday local models are powerful enough to yield accurate definitions.