Add support of local AI models.

[github-actions[bot]]

Thank you for your contribution. We will check and reply to you as soon as possible.

[JoeyWangTW]

Are you thinking something like sending a request to a ollama local host endpoint?

[R-udren]

Absolutely. Also, litellm is an option. As I saw code, this isn't really hard task.

[JoeyWangTW]

Yes it shouldn't be that hard, I'll try it out and see what model can handle the job with reasonable latency.

[JoeyWangTW]

I've tried using llama 8B locally, it's doable but the latency can easily go up to 10s+ for a simple task. Not sure if that will be a pleasant UX. But I'll still work on adding an option to input port and model name for local models.

[R-udren]

I will try it on my hardware

[JoeyWangTW]

I've just merged a PR #79 with local model settings. I've tested it with ollama. Not sure how much settings is needed for different endpoints.

[R-udren]

I found a few issues with the extension:

1. Manifest File:
   In youtube-addiction-rehab-chrome-extension\chrome-extension\manifest.js on line 24, replace http://localhost* with http://localhost/*. It raises this error:
   

2. Local Models Dropdown:
   The extension should list all available local models in the "Choose a model" dropdown. List local models: http://localhost:11434/api/tags

3. Popup Error Message:
   Sometimes, the following error message appears in the popup while using a local model:

   ⬅️ Please set up your API key in the settings.

4. Model Performance:
   The model typically takes about 2-5 seconds to respond. When using tinyllama instead of llama3.1:8b, tinyllama generates Python code for no reason 🤣.

5. System Prompt: Properties like id and reason should be enclosed in double quotes. But in system prompt they aren't.

My Suggestions:

• If the extension cannot parse JSON properly, it should retry the request up to three times. If it still fails to receive a valid response, it should fall back to the default Youtube recommendations.

• The "reason" field may not be necessary, as it can increase model response time.