gptel VS ata

Emacs is able to run prompts, just like any other tool that's able to make an HTTP request. E.g., https://github.com/karthink/gptel provides hooks to both OpenAI and local LLMs, and allows you to use them anywhere—since everything in Emacs is a buffer.

There are some pretty great examples in the README.

Also worth checking out for more general use of LLMs in emacs: https://github.com/karthink/gptel

gptel: An LLM chat client for Emacs, supports ChatGPT and local LLMs (via Ollama, GPT4All or llama.cpp)

That would be great, though ChatGPT seems to prefer Markdown even when org is requested. Didn't want to fight it too much. Did you manage to get a reliable prompt, or if I'm understanding correctly gptel converts Markdown to Org?

I use https://github.com/karthink/gptel to talk to GPT in Emacs and, by extension, have a record of all my conversations (unless intentionally removed) as simple Org files in my single Org-Roam network.

It made me a working function on the first try. (sorry, I lost the exact prompt, I just ran gptel from my scratch buffer)

> My main gripe with it though is that the actual interface is quite limited, and I don't like having to navigate to the site every time I want to use it.

https://github.com/rikhuijzer/ata. Written in Rust. Does streaming responses and supports Emacs keyboard shortcuts. That's it. All that's needed for quickly looking up things.

- Checkout definitions. I have a small tool (https://github.com/rikhuijzer/ata) available on a keyboard shortcut and use it to quickly checkout definitions for words when I come across a word that I don't know.

As a little side-note, it actually involved some hacking to get the output printing correct because it appeared that the API sometimes responds denotes a newline by two tokens (["\", "n"]) and sometimes by one token (["\n"]). In the playground (https://platform.openai.com/playground), they convert the two token version to a single token, so that's what ata does too (details in https://github.com/rikhuijzer/ata/pull/6). My guess is that this is basically a bug in the model which they manually fixed in the Playground and ChatGPT front ends.

ChatGPT made my work more productive, but I was having a bit of a struggle with the browser, timeouts and slow responses, and the lack of keyboard shortcuts. That's why I made a terminal application: https://github.com/rikhuijzer/ata. You can download it for free in the releases section or build it yourself from source. I'm using it daily and hope it is useful for people here too