gpt.vim

(unfinished but working project, you may encounter problems.)

A plugin for interfacing with general purpose text models, such as openai-chatgpt, localai, etc. Works with vim and neovim.

TODO

• Implement a localai backend as a demo back
• Implement proper openai backend
• Tidy the code and remove wonk
• Reset function for the sessions
• (MAYBE) option to change session on the fly to enable multiple contexts for the backend.
• The Gpt command is quite wonky but seems useful, needs quite bit of fixing & rewriting.

Goals

• vim & neovim compatible codebase
• small & simple
• Heavy lifting should be done by the "backend". Openai is not the only player anymore, local & private general purpose language models exist.
• simple interface with the backends using pipes.

Requirements

An application that accepts a stdin stream of JSON. Such an application can be found at github.com/zutto/python-chatgpt-vimbackend (WIP), github.com/mattn/chatgpt or sample bash backend. You can also build your own simple piped handling.

Example of the JSON stream:

{"model": "gpt-4", "text":"Hello world", "systemrole":"custom role", "session": "<name of session, optional>"}

Sessions are just different chat sessions, so for completion you could use completion as session name, for chat just chat.

Example of the output stream expected from the application:

{"eof": false, "error": "", "text": "", notification: ""}
{"eof": false, "error": "", "text": "Hello", notification: ""}
{"eof": false, "error": "", "text": "!", notification: ""}
{"eof": false, "error": "", "text": " How", notification: ""}
{"eof": false, "error": "", "text": " can", notification: ""}
{"eof": false, "error": "", "text": " I", notification: ""}
{"eof": false, "error": "", "text": " assist", notification: ""}
{"eof": false, "error": "", "text": " you", notification: ""}
{"eof": false, "error": "", "text": " today", notification: ""}
{"eof": false, "error": "", "text": "?", notification: ""}
{"eof": false, "error": "", "text": "", notification: ""}
{"eof": false, "error": "", "text": "", notification: ""}
{"eof": true, "error": "", "text": "", notification: ""}

Reset session (cleans the conversations, resets the underlying chatsession, reauthenticates, etc):

{"model": "gpt-4", "text":""", "systemrole":"","session": "<optional>", "reset"="true"}

Settings

• g:chatgpt_bin - Location of the program this plugin interfaces with. Default value: ['python3.11', '/usr/local/bin/chatgpt']. Example:

  let g:chatgpt_bin = ['python3.11', '/usr/local/bin/chatgpt']

• g:chatgpt_role - Role to set for the assistant. This is a great place to use augroups to set different roles for different files. The program that interfaces with this plugin should handle updating the role for the assistant as needed. Example:

  let g:chatgpt_role = "You are a helpful commit writer who likes to write easy to understand and compact commit messages."

• g:chatgpt_role_completion - same as g:chatgpt_role but for completion requests.

Commands

• Chatgpt - Calls the assistant. Note: You highlight contents with visual mode. The highlighted contents will be sent to the assistant along with your request.

• Gpt - Calls the assistant with completion session, using the completion role. This may be wonky, as this uses chat session for completion. Mainly made for fun.

• ChatgptSetRole - Set a custom role on the fly.

Credits

mattn's project was an inspiration for this project. This is basically a personal rewrite of the original project: github.com/mattn/vim-chatgpt. Give mattn all the love.

Demo

normal usage

Example of using visual mode.

Gpt command demo