Abstract:

This video is a comprehensive demonstration of gptel, a simple and powerful Large Language Model (LLM) client designed for seamless integration with Emacs. The presenter showcases its core design philosophy of being unobtrusive, flexible, and universally available across any buffer. The tutorial covers installation, basic chat functionality, and advanced features like multi-LLM provider support (ChatGPT, Gemini, Ollama, Kagi), persistent chats, and on-the-fly customization using system directives. Key highlights include practical examples of gptel's "just-in-time" capabilities, such as fixing LaTeX code in-place, generating complex shell commands from natural language descriptions in the minibuffer, and discovering Emacs functions. The video emphasizes gptel's small API surface, its focus on text-based interaction, and its asynchronous nature, allowing users to continue working while the LLM processes requests.

Integrating LLMs into Emacs: A Deep Dive into the gptel Package

• 0:27 Installation and Basic Usage: The package can be installed from the MELPA repository. The primary command, gptel-send, sends the text before the cursor (or the selected region) to an LLM. Users can also open dedicated, persistent chat buffers with the gptel command.
• 2:18 Multi-Provider Support: gptel offers a consistent interface for numerous LLM providers. The video demonstrates switching between OpenAI's ChatGPT and Google's Gemini, highlighting the package's asynchronous capability to run multiple queries simultaneously.
• 3:36 Persistent Chats: Chat sessions are just regular text, Org, or Markdown files. Saving a buffer also saves the chat metadata, allowing a conversation to be restored and continued later by enabling gptel-mode.
• 4:01 Custom Directives: The LLM's behavior can be easily customized using system messages or "directives." This allows the user to prime the model for specific tasks, demonstrated by turning GPT-4 into a scientific calculator that outputs LaTeX.
• 7:07 Core Design Philosophy: The package is designed to be "boring" and fade into the background. It is fully asynchronous, does not move the cursor or auto-scroll by default, and aims to be an unobtrusive tool available everywhere in Emacs.
• 7:55 Input/Output Redirection: Queries and responses can be redirected. An example shows sending a query from the minibuffer and having the response appear in the echo area for quick, transient interactions.
• 8:32 Contextual Queries: Users can send an entire file as context for a query and redirect the resulting conversation into a new, dedicated buffer, keeping the original file clean.
• 11:04 In-Place Editing: A powerful feature is demonstrated by selecting a broken LaTeX environment within an Org document and using gptel to rewrite and fix it directly in the buffer.
• 12:08 Just-in-Time Shell Commands: gptel can be invoked directly from the minibuffer's shell command prompt (M-x shell-command). The user types a plain English request (e.g., "find all mp4s in this dir and convert to mp3"), and gptel replaces it with the correct shell command.
• 13:09 Just-in-Time Emacs Discovery: Similar to the shell command example, gptel can be used to find the name of an Emacs command from a natural language description within the M-x prompt.
• 14:13 Multi-LLM Demonstration: A single query is sent to four different backends simultaneously (Gemini, a local Mistral/Ollama model, Kagi, and ChatGPT) to showcase the uniform user experience.
• 16:46 Advanced Customization: For non-interactive or custom workflows, gptel provides hooks and a stable API function, gptel-request, which works with any supported provider.