Undefine default ChatGPT backend

[acceleratesage]

Hi, is there a way to undefine the default backend and models? I'd like to only work with a selected list of backends.

[karthink]

(setf (gptel-get-backend "ChatGPT") nil)

[acceleratesage]

When using this in :config in use-package, it fails with:

Error (use-package): gptel/:config: Symbol's function definition is void: \(setf\ gptel-get-backend\)

When putting it outside of use-package, it works.

[karthink]

acceleratesage left a comment (karthink/gptel#649) When using this in `:config` in `use-package`, it fails with: ```Error (use-package): gptel/:config: Symbol's function definition is void: \(setf\ gptel-get-backend\)```

Can you provide your full use-package declaration for gptel here?

[acceleratesage]

(use-package gptel
  :config

  (gptel-make-openai "Groq"
    :host "api.groq.com"
    :endpoint "/openai/v1/chat/completions"
    :key gptel-api-key
    :stream t
    :models '(deepseek-r1-distill-llama-70b llama-3.3-70b-versatile))

  (gptel-make-openai "xAI"
    :host "api.x.ai"
    :key gptel-api-key
    :endpoint "/v1/chat/completions"
    :stream t
    :models '(grok-2-1212 grok-2-vision-1212))

  (setq gptel-default-mode 'org-mode
	gptel-backend (gptel-get-backend "Groq")
	gptel-model 'deepseek-r1-distill-llama-70b))

;; Only works outside of use-package
(setf (gptel-get-backend "ChatGPT") nil)

[benma]

This ideally goes into the README, it's not trivial to figure out this solution to undefine a backend.

[rudolf-adamkovic]

+1

My model menu is spammed with proprietary ChatGPT offerings as well, despite the fact that I configured GPTel to use Ollama:

(with-eval-after-load 'gptel
    (let ((models '(llama3.1:8b
                    phi4:14b
                    deepseek-r1:14b)))
      (setq gptel-model (car models)
            gptel-backend (gptel-make-ollama "Ollama"
                            :host "localhost:11434"
                            :stream t
                            :models models))))

Expected:

• Ollama:phi4:14b
• Ollama:llama3.1:8b
• Ollama:deepseek-r1:14b

Actual:

• ChatGPT:gpt-3.5-turbo More expensive & less capable than GPT-4o-mini; use that instead (tool-use) 16k $ 0.50 in, $ 1.50 out 2021-09
• ChatGPT:gpt-3.5-turbo-16k More expensive & less capable than GPT-4o-mini; use that instead (tool-use) 16k $ 3.00 in, $ 4.00 out 2021-09
• ChatGPT:gpt-4 GPT-4 snapshot from June 2023 with improved function calling support (media tool-use url) 8k $30.00 in, $ 60.00 out 2023-09
• ChatGPT:gpt-4-0125-preview GPT-4 Turbo preview model intended to reduce cases of “laziness” (media tool-use url) 128k $10.00 in, $ 30.00 out 2023-12
• ChatGPT:gpt-4-1106-preview Preview model with improved function calling support (tool-use) 128k $10.00 in, $ 30.00 out 2023-04
• ChatGPT:gpt-4-32k (tool-use) $60.00 in, $120.00 out
• ChatGPT:gpt-4-turbo Previous high-intelligence model (media tool-use url) 128k $10.00 in, $ 30.00 out 2023-12
• ChatGPT:gpt-4-turbo-preview Points to gpt-4-0125-preview (media tool-use url) 128k $10.00 in, $ 30.00 out 2023-12
• ChatGPT:gpt-4o Advanced model for complex tasks; cheaper & faster than GPT-Turbo (media tool-use json url) 128k $ 2.50 in, $ 10.00 out 2023-10
• ChatGPT:gpt-4o-mini Cheap model for fast tasks; cheaper & more capable than GPT-3.5 Turbo (media tool-use json url) 128k $ 0.15 in, $ 0.60 out 2023-10
• ChatGPT:o1 Reasoning model designed to solve hard problems across domains (nosystem media reasoning) 200k $15.00 in, $ 60.00 out 2023-10
• ChatGPT:o1-mini Faster and cheaper reasoning model good at coding, math, and science (nosystem reasoning) 128k $ 3.00 in, $ 12.00 out 2023-10
• ChatGPT:o1-preview DEPRECATED: PLEASE USE o1 (nosystem media) 128k $15.00 in, $ 60.00 out 2023-10
• ChatGPT:o3-mini High intelligence at the same cost and latency targets of o1-mini (nosystem reasoning) 200k $ 3.00 in, $ 12.00 out 2023-10
• Ollama:deepseek-r1:14b
• Ollama:llama3.1:8b
• Ollama:phi4:14b

[acceleratesage]

For what it's worth, I think there should be no backend defined by default. Some might not even want the possibility of sending their data to OpenAI without explicitly enabling it. Making it easy to integrate and not having to list all the models is great though.

[karthink]

This ideally goes into the README, it's not trivial to figure out this solution
to undefine a backend.

@benma There appears to be a bug with gptel-get-backend, it's not being recognized as a generalized variable in the normal flow of loading gptel.el. After I fix this bug I'll add it to the README.

[karthink]

For what it's worth, I think there should be no backend defined by default.

@acceleratesage Most gptel users are still using ChatGPT, best I can tell. When they update gptel and it stops working, I'm going to have a rough time with the flood of support requests.

Some might not even want the possibility of sending their data to OpenAI
without explicitly enabling it. Making it easy to integrate and not having to
list all the models is great though.

I'll settle for making the default backend easy to disable.