Configure An Open Model Out of The Box

[jgarte]

Hi,

ChatGPT is configured out of the box.

WDYT if we configured an open model out of the box instead of ChatGPT in order to encourage the use and development of open source LLMs?

[astoff]

My issue with the default ChatGPT configuration is that it's apparently impossible to remove it, so I always see a bunch of bogus options in the model list. Moreover I see little benefit in a model being configured out of the box when the user has to perform the nontrivial extra step of setting up an API key anyway.

(But I also think it's wrong for gptel-make-* to add entries to gptel--known-backends as a side effect; IMO the configuration should be through a list gptel-backends that the user controls.)

[karthink]

My issue with the default ChatGPT configuration is that it's apparently impossible to remove it, so I always see a bunch of bogus options in the model list.

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

Moreover I see little benefit in a model being configured out of the box when the user has to perform the nontrivial extra step of setting up an API key anyway.

The API key can be provided interactively when you run M-x gptel, requiring no elisp configuration if you use ChatGPT.

But I also think it's wrong for gptel-make-* to add entries to gptel--known-backends as a side effect; IMO the configuration should be through a list gptel-backends that the user controls.

If you don't want to add backends to the registry, use gptel--make-* instead of gptel-make-*. Note that you'll have to supply all the backend slots (like the :header, :endpoint etc.)

[astoff]

Okay, I see that you want to make the package set up easy, but I think there's more to gain from making it simple (in the Rich Hickey sense). What you described above is not simple :-).

The backend settings could be done by a simple variable, e.g. an alist of plists (mapping a backend name to what is now the argument list to gptel-make-*). This way, it's easy to override the built-in settings or just add to it; you don't need to load the package to configure it (or use eval-after-load or some derived mechanism); and perhaps more usefully, you can configure everything it via M-x customize.

The API key can be provided interactively when you run M-x gptel, requiring no elisp configuration if you use ChatGPT.

Ah, that's interesting (and orthogonal to the above — I guess every package could/should do that...). But currently, with the released ELPA version I don't see such a prompt. Instead I get an error: "gptel-api-key-from-auth-source: No ‘gptel-api-key’ found in the auth source".

[karthink]

My issue with the default ChatGPT configuration is that it's apparently impossible to remove it, so I always see a bunch of bogus options in the model list.

The default ChatGPT backend has been removed. gptel-backend and gptel-model default to nil. The full story is a little more complicated, please see the NEWS file if you're interested.

[karthink]

and perhaps more usefully, you can configure everything it via M-x customize.

gptel-backend can now be configured via M-x customize, and specified as a list with setopt. It's functional but clunky -- about as clunky as using an alist of plists would be. Suggestions welcome.

This way, it's easy to override the built-in settings or just add to it; you don't need to load the package to configure it

Yes, the necessity to load the package to configure gptel-backend has been a pain. It continues to be a pain, as it's not clear how to convert a plist-specification to the cl-defstruct object in all the places that gptel-backend is used.

Instead I get an error: "gptel-api-key-from-auth-source: No ‘gptel-api-key’ found in the auth source".

You should be asked for an API key when you run M-x gptel. The check does not occur before every call to gptel-send. I tested using M-x gptel with zero configuration and it works as expected.

[astoff]

Alright, makes sense. By the way, sorry for repeating in this thread an opinion I had already expressed in another issue (the one you commented a few minutes ago) :-).

[karthink]

Where is this open model hosted?