Abstract:

This video introduces GPTEmacs (gpel), an asynchronous and flexible client for large language models (LLMs) deeply integrated into Emacs. It showcases gpel's installation, core features, and design philosophy, emphasizing its unobtrusive nature and ubiquitous availability across Emacs buffers. The demonstration covers setting up persistent chats, utilizing various LLM providers (e.g., OpenAI, Google Gemini, Kagi, Mistral), and leveraging system messages (directives) for nuanced control. Key features include asynchronous processing, flexible input/output redirection, and a consistent interface, allowing users to interact with LLMs for tasks ranging from code explanation and LaTeX correction to CLI command generation, all without leaving their Emacs environment.

Exploring GPTEmacs: An LLM Client for Emacs

• 0:00 Introduction to GPTEmacs: GPTEmacs (gpel) is presented as an asynchronous, flexible Emacs client for large language models (LLMs), designed for persistent chats, multi-LLM provider support, and an intuitive, unobtrusive interface within Emacs.
• 0:24 Installation: The video demonstrates installing gpel via the Melpa package archive, a standard process for most Emacs users.
• 1:09 Basic Usage and Multi-LLM Support: Users can initiate conversations from any Emacs buffer using M-x gpel-send or create dedicated chat buffers with M-x gpel. The tool supports various LLM providers (e.g., Google Gemini, OpenAI) configurable via a prefix argument, showcasing its asynchronous operation that allows Emacs use while waiting for responses.
• 2:36 Persistent Chats: Chats are stored as regular text or Markdown files, with metadata automatically saved to preserve conversation context. gpel-mode enables restoring this context upon reopening a file.
• 3:55 Flexible System Messages (Directives): A key feature is the ability to use or create "system messages" or "directives" to guide LLM behavior, demonstrated through examples like making the LLM act as a mathematician (outputting LaTeX results) or a dictionary (providing synonyms/antonyms), including region-limited queries and in-place prompt rewriting.
• 7:05 Design Philosophy: gpel aims to be "boring" and "unobtrusive," meaning it avoids unsolicited cursor movement, auto-scrolling, or "magic" behavior by default, though these can be enabled. It strives for full asynchronous operation within Emacs' design limitations.
• 8:02 Input/Output Redirection: The tool simplifies redirecting prompts (e.g., from the mini-buffer) and responses (e.g., to the echo area for quick checks, or to a new session for multi-file code explanation like Julia code).
• 10:20 Ubiquitous Availability: gpel is designed to be available in "any buffer," demonstrated by examples such as correcting LaTeX in an Org document, generating CLI commands in the mini-buffer, and assisting with searching for Emacs functions.
• 13:36 Consistent Interface: Despite LLM API differences, gpel presents a consistent interface across providers (Gemini, Mistral, Kagi, ChatGPT) and Emacs buffer types, with gpel-send being the primary command accessible everywhere.
• 15:53 Future Focus & Configurability: The package will remain focused on text-based interactions to maintain interface simplicity, even as new generative AI features emerge. It is highly configurable, offering hooks for custom behaviors and an API (gpel-request) for non-interactive uses.