[Enhancement]: Allow user to turn off Thinking parameter for selected Google Gemini models

[cloud-25683]

What features would you like to see added?

By fault, all google Gemini models have Thinking turned on in the Parameters panel, however, not all Gemini models support Thinking, such as gemini-2.0-flash-001, can we do something similar to allow configure drop Thinking parameter globally for certain models, for example, Azure OpenAI has a parameter I used: dropParams but Google endpoint doesn't support this.

More details

allow configure dropParams setting globally for Google endpoint configuration in librechat.yaml or via Env variable

Which components are impacted by your request?

No response

Pictures

No response

Code of Conduct

• I agree to follow this project's Code of Conduct

[schnaker85]

this PR does not solve your enhancement?

[cloud-25683]

I don't think so, at least I don't see anywhere in that PR specifically said I can turn off thinking for certain models globally via librechat.yaml, for example, I have been doing this for azure openai o models with dropParams setting, I don't see google endpoint support this, can you tell me otherwise?
endpoints:
azureOpenAI:
titleModel: "gpt-4o-mini"
titleConvo: true
groups:
# Group-level configuration
- group: "azure-openai-o-models"
apiKey:
baseURL:
instanceName:
version:
dropParams:
- "temperature"
- "max_tokens"
# Model-level configuration
models:
o1:
deploymentName: o1
o3-mini:
deploymentName: o3-mini

[oliver-31415]

I'm confused as to how to get thinking on my google models, for some reason they by default have no option for thinking, and same thing for openAI models. For openAI models I know i somehow have to use the Responses API but I have not found out how to actually do this, given the lack of documentation. Is there some setting in librechat.yaml that will allow me to turn this on? Currently the google models are enabled through .env file, so im not sure if this is the problem?

[cloud-25683]

I am using 0.7.9-rc1 version, if you are using older versions, Thinking parameters will not be displayed

[cloud-25683]

@danny-avila @schnaker85 any comments on my question/request?

[schnaker85]

well, I know understand what you are trying to achieve.

Just to sum up your request/question better, you configured your Google Endpoints via Env variables:

ENDPOINTS=google,..
GOOGLE_MODELS=gemini-2.5-flash,gemini-2.5-pro,...

and you do not use any modelSpec, so you can not set any presets for the specifid models to set the thinking, thinkingBuget, right? And a Combination does not work, otherwise you will have duplicated endpoints/model in the Model Dropdown

I think in that case this is not yet supported, only with model_specs and preset not globally. Am I right? @danny-avila?

[FINDarkside]

Did you figure out any workaround? I'm having hard time figuring out how to use these models with LibreChat since I cannot ship it with broken default configuration.

[FINDarkside]

Ok managed to solve it for my usecase like this:

modelSpecs:
  enforce: true
  prioritize: true
  list:
    - name: gemini-2-flash
      label: Gemini 2.0 Flash
      model: gemini-2.0-flash
      preset:
        endpoint: google
        model: gemini-2.0-flash
        thinking: false
        thinkingBudget: 0

interface:
  modelSelect: false
  parameters: false
  presets: true

It will hide the model select so users can only choose presets, and the Gemini 2.0 Flash preset has thinking disabled. parameters: false is optional if you want to prevent users from tweaking the params.

[oliver-31415]

I find this is a kinda solution, but I don't really like it as I want to be able to use the models, just as a default use certain specs. It's kinda frustrating that this functionality is not possible.

[cloud-25683]

I agree with you, I think in general, librechat should support disabling parameters globally, not just via presets, I mean azure openai does support it, but not every providers

[danny-avila]

I also agree, as right now it's only possible via custom endpoints. Will think about the best course of action.