Option to save <ChatGPT> buffers by default

[moritzschaefer]

It would be nice to have the configurable option to automatically associated the gptel buffers (named <ChatGPT #n> by default with a file (e.g. in a configurable directory with file name .txt). This way, without extra effort, one can maintain a well-accessible history of chat interactions.

If not within gptel, do you have a recommendation how to implement this in the easiest manner?

[karthink]

There are many ways to do it:

• Save the gptel buffer to a file as needed. I take it this is the step you want to automate.

• Save the buffer to a file when gptel-mode is turned on: (Note: untested)

(defun my/gptel-write-buffer ()
  "Save buffer to disk when starting gptel"
  (unless (buffer-file-name (current-buffer))
    (let ((suffix (format-time-string "%Y%m%dT%H%M" (current-time)))
          (chat-dir "~/chatgpt-log"))
      (unless (file-directory-p chat-dir)
        (make-directory chat-dir :parents))
      (write-file (expand-file-name (concat "gptel-" suffix ".txt") chat-dir)))))

(add-hook 'gptel-mode-hook #'my/gptel-write-buffer)

This will save new gptel sessions to ~/chatgpt-log with names like gptel-20231105T1026.txt. Or you can save to default-directory, or to (project-root (project-current)) to save them at the root of your current project in Emacs.

• I use a single file per project (always named Chat.org) to keep persistent logs of ChatGPT conversations, and bind summoning it to the project keymap (C-x p), where all the other project commands are. Chat topics are separated by Org headers with gptel-set-topic.

This is a feature best tailored by users for their individual workflows, and thus not something I plan to implement in gptel -- so I'm moving this thread to the discussions section.

[agzam]

Does it really matter what file extension you'd use? Correct me if I'm wrong, I think the package currently handles markdown and txt better than org (proly because LLMs don't handle org-mode very well), is it better to set the extension according gptel-default-mode?

[karthink]

I don't think the extension matters, so long as it's one of txt, md or org, since these are the ones supported by gptel-mode.

I think the package currently handles markdown and txt better than org

It should handle Org fine, I think. gptel converts from markdown to org as required.

[agzam]

It should handle Org fine

Oh, very nice, I'm gonna change gptel-default-mode to org. I might have confused it with chatgpt-shell. I don't know though how good is its support of org-mode these days, I stopped using it a while ago.

[agzam]

Btw, it looks like you may want to update the docstring for the var, it currently says:

gptel-default-mode
The default major mode for dedicated chat buffers.

If markdown-mode is available, it is used.  Otherwise gptel defaults to text-mode.

I tried setting it to org-mode and it works nicely.

[agzam]

I have another question about this approach.

It looks like that it sets some local vars in the header of the buffer, inserting e.g., something like:

<!-- Local Variables: -->
<!-- gptel-model: "claude-3-5-sonnet-20240620" -->
<!-- gptel--backend-name: "Claude" -->
<!-- gptel--bounds: nil -->
<!-- End: -->

That's nice, but then it doesn't do anything when you switch the model. I wonder if the mode-hook is the best one for this task?

[karthink]

That's nice, but then it doesn't do anything when you switch the model.

Not sure how that matters. It records the model and other details when you save the buffer.

[agzam]

What I'm saying is that it doesn't change those vars afterwards, even if you switch to a different model. Not a big deal. I think this can be fixed by additionally hooking to another hook.

[karthink]

I tried the following:

1. Start a chat (or turn on gptel-mode)
2. Save the buffer. This writes the backend and model details to the file.
3. Now change the model, system message or other backend details.
4. Save the file again. This updates the model details in the file.

So I'm not sure what you mean by this:

it doesn't change those vars afterwards, even if you switch to a different model

So it's either a bug or a misunderstanding.

[agzam]

This updates the model details in the file.

Yeah, I expected exactly that, alas that didn't do it for some reason. Maybe something is messing it up in my setup, I'll try to find out. Also, I think switching gptel-default-mode to org sometimes breaks it in a weird way where the api stops responding, I hit that twice, I'll keep an eye on that as well.

[karthink]

Please create an issue (or two) if you can find a way to reproduce these bugs.