Switching chat model

[swerling]

I cannot figure out how to change the model used to chat. I'm trying to use gemini gemini-2.5-flash-preview-05-20.

I can select the provider ok. But if I select gemini-2.5-flash-preview-05-20, then do PrtChatNew, and send 'testing connection', it shows it is using embedding-gecko-001.

Also, if I do PrtModel gemini-2.5-flash-preview-05-20, then PrtChatNew, it still shows embedding-gecko-001.

Also, despite having online_model_selection = false, in my config, it always shows all the gemini models when in the model selection window, rather than the 4 models I list in my config.

Thanks in advance for you help.

[frankroeder]

Hi @swerling, which version or commit are you currently using? I have implemented a new caching mechanism that fetches all the models from the API endpoint. To deactivate this, please do not set the model_enpoint in your provider configs or simply put model_cache_expiry_hours = 0. Then it will fall back to your predefined static model table.

Let me know if this resolves the issue.

[swerling]

Hi @frankroeder thanks for responding so quickly.

I was having this issue both in v1.8.0, and on HEAD (whatever that was yesterday).

On 1.8.0, I worked around this by removing all providers except gemini, turning off online_model_selection, and making sure the only model available was gemini-2.5-flash-preview-05-20. Not ideal since I like to switch models from time to time, but ok.

For HEAD, I just tried your suggestion of commenting out the model_endpoint. Then I noticed it chose the first model no matter what I did (I have 5 gemini models in my model list in the config). To change models in this branch I had to 1) exit nvim; 2) delete persisted/state.json (this was necessary); 3) restart nvim. Now PrtModel works and I can switch, even mid chat, which is great. Thank you for the suggestion.

Parrot periodically refusing to change models has been a long time issue for me, across versions, and I have to reset things to fix it. Maybe something is getting stuck in state.json that is overriding model selection?

Thanks for a great plugin!

[raindealer]

I observed the same behavior and wondered why it hadn't been fixed, as the current version is much higher than 1.8. I would select the provider and the model, open a chat window, ask a question, and the first model in the list was always selected (regardless of whether the list was static or dynamic). I checked the source code and found that when a model is selected, a flag is passed that indicates whether we set the model in the chat window or just in our own buffer. This behavior was unexpected for me; I thought the setting was applied to all modes (this behavior is not described in the README).

Basically, the model is currently selected separately for the command and for the chat. I think this behavior should be explicitly documented.

[frankroeder]

Hi @raindealer.

Thank you very much for bringing this up again. Indeed, there is a confusion about the two models selected for chat and command. You are absolutely right that I also have not completely understood it when @swerling brought this up some time ago.

This behavior is on purpose, and I will update the documentation and the FAQ accordingly.

Answer: In short, to change the chat model, one has to be inside the chat started with PrtChatNew. Switching the model in any other case will only switch the interactive command model. After that, the selections will be persistent.