feat: Support editing all parameters individually for each model

[Emt-lin]

Overview

Now, you can edit all all parameters individually for each model and new models by duplicating existing ones.

actual effect:

issues

• Allow editing of models #1083
• I hope each model can have its own temperature setting and include top-p. #1374
• Add frequency penalty #1139

Key changes

• support copy model for quicklly add.
• support edit maxToken、temperature、topP args for each model

[logancyang]

bugbot run

[logancyang]

With these individual parameter settings, should the global one still exist?

[Emt-lin]

With these individual parameter settings, should the global one still exist?

Now, temperature and maxTokens is global parameters. For non-technical personnel, default values for these parameters are needed, and it should be convenient to quickly adjust all models.

Once they understand the meaning of these parameters, they can independently configure the parameters for each model.

This is a step-by-step process.

[logancyang]

According to our previous chat, we should either remove the global settings or do something else, the idea is to have only 1 source of truth.

[Emt-lin]

According to our previous chat, we should either remove the global settings or do something else, the idea is to have only 1 source of truth.

I have removed the global settings for temperature and maxTokens

[logancyang]

bugbot run

[logancyang]

I just realized that embedding models are not editable like the chat models. We should make them editable too same as chat models.

[logancyang]

Also don't forget to rebase on the new master.

[logancyang]

bugbot run

[cursor]

Bug: API Parameter Removal Breaks User Configurations

The safeFetch function unconditionally deletes the frequency_penalty parameter from all API request bodies. This causes user-configured frequencyPenalty settings (added in chatModelManager.ts) to be ignored by all providers using safeFetch with CORS enabled, rendering the feature non-functional. Although the code comment indicates this was a temporary workaround for perplexity.ai, its global application breaks the feature for all other providers.

src/utils.ts#L491-L495

obsidian-copilot/src/utils.ts

Lines 491 to 495 in f49fe3c

	const newBody = JSON.parse(options.body ?? {});
	// todo(emt-lin):frequency_penalty: default 0, but perplexity.ai requires 1 by default. so, delete this argument for now
	delete newBody["frequency_penalty"];
	options.body = JSON.stringify(newBody);
	}

Fix in Cursor

---

Bug: Temperature Slider Default Mismatch

The temperature slider in the model edit modal uses a fallback default value of 1. This is inconsistent with the PR discussion, which indicated a default of 0.9, and the previous global default of 0.1. A temperature of 1 is unusually high for most AI models, leading to overly creative and unpredictable outputs.

src/settings/v2/components/ModelEditDialog.tsx#L298-L299

obsidian-copilot/src/settings/v2/components/ModelEditDialog.tsx

Lines 298 to 299 in f49fe3c

	<SettingSlider
	value={localModel.temperature ?? settings.temperature ?? 1}

Fix in Cursor

---

Was this report helpful? Give feedback by reacting with 👍 or 👎

[logancyang]

Why 1? It's another magic number, if there's a reason we should name it something obvious.

[Emt-lin]

it's the result of autocompelete by ai. I forget to keep it consistent with the setting default value.(and also maxTokens )

[Emt-lin]

@logancyang Can you test the perplexity.ai API? It has supported the frequency_penalty value of 0.

docs: https://docs.perplexity.ai/api-reference/chat-completions-post#body-frequency-penalty