Ollama Support for Vim

This plugin adds Copilot-like code completion support to Vim. It uses Ollama as a backend, which can run locally and does not require cloud services, thus preserving your privacy.

Motivation

Copilot.vim by Tim Pope is an excellent plugin for both Vim and NeoVim. However, it is limited to Microsoft's Copilot, a commercial cloud-based AI that requires sending all your data to Microsoft.

With Ollama and freely available LLMs (e.g., Llama3, Codellama, Deepseek-coder-v2), you can achieve similar results without relying on the cloud. While other plugins are available, they typically require NeoVim, which isn't an alternative for me. I prefer using Vim in the terminal and do not want to switch to NeoVim for various reasons.

Features

• Intelligent AI-based code completion
• Integrated chat support for code reviews and other interactions

Screencasts

Creating a C application with command line option parsing using AI

Creating Enum to String Conversion function using AI

Code Review

Custom Prompts - Spellcheck Example

How It Works

The plugin uses two Python scripts, ollama.py and chat.py, to communicate with Ollama via its REST API. The first script handles code completion tasks, while the second script is used for interactive chat conversations. The Vim plugin uses these scripts via I/O redirection to integrate AI results into Vim.

This plugin supports Vim only, not NeoVim. If you're looking for a NeoVim plugin, check out LLM.

Requirements

• Python 3.x
• Python packages: httpx

Debian based Systems

sudo apt install python3-httpx

Other systems

System wide installation using pip install is not recommended, use a virtual environment instead.

You need to run Vim from a shell with this Python environment to make this working.

Example:

python -m venv $HOME/vim-ollama
source $HOME/vim-ollama/bin/activate
pip install httpx

Installation

Install gergap/vim-ollama using vim-plug or any other plugin manager.

vim-plug example:

call plug#begin()
...
Plug 'gergap/vim-ollama'
call plug#end()

Configuration

By default, the plugin uses Ollama on localhost. You can change this by adding the following variable to your .vimrc:

let g:ollama_host = 'http://tux:11434'

Next, configure the LLM models and the corresponding fill-in-the-middle (FIM) tokens. The variable g:ollama_model defines the LLM for code completion tasks. This must be a model with fill-in-the-middle support; otherwise, code completion may not work as expected. The variable g:ollama_chat_model is used for interactive conversations, similar to ChatGPT.

Example configuration:

" Default chat model
let g:ollama_chat_model = 'llama3'

" Codellama models
let g:ollama_model = 'codellama:13b-code'
let g:ollama_model = 'codellama:7b-code'
let g:ollama_model = 'codellama:code'

" Codegemma (small and fast)
let g:ollama_model = 'codegemma:2b'
let g:ollama_fim_prefix = '<|fim_prefix|>'
let g:ollama_fim_middle = '<|fim_middle|>'
let g:ollama_fim_suffix = '<|fim_suffix|>'

" qwen2.5-coder (0.5b, 1.5b, 3b, 7b, 14b, 32b)
" smaller is faster, bigger is better"
" https://ollama.com/library/qwen2.5-coder
let g:ollama_model = 'qwen2.5-coder:3b'
let g:ollama_fim_prefix = '<|fim_prefix|>'
let g:ollama_fim_middle = '<|fim_middle|>'
let g:ollama_fim_suffix = '<|fim_suffix|>'

" Deepseek-coder-v2
let g:ollama_model = 'deepseek-coder-v2:16b-lite-base-q4_0'
let g:ollama_fim_prefix = '<｜fim▁begin｜>'
let g:ollama_fim_suffix = '<｜fim▁hole｜>'
let g:ollama_fim_middle = '<｜fim▁end｜>'

Variable	Default	Description
g:ollama_host	http://localhost:11434	The URL of the Ollama server.
g:ollama_chat_model	llama3	The LLM for interactive conversations.
g:ollama_model	codellama:code	The LLM for code completions.
g:ollama_fim_prefix	<PRE>	FIM prefix for Codellama.
g:ollama_fim_middle	<MID>	FIM middle for Codellama.
g:ollama_fim_suffix	<SUF>	FIM suffix for Codellama.

When changing the code completion model, consult the model’s documentation to find the correct FIM tokens.

Usage

Simply start coding. The completions will appear as "ghost text" and can be accepted by pressing <tab>. To ignore them, just continue typing or press <C-]> to dismiss the suggestion.

See :help vim-ollama for more information.