Local LLM-assisted text completion using Ollama.

• Auto-suggest on cursor movement in Insert mode
• Toggle the suggestion manually by pressing Ctrl+F
• Accept a suggestion with Tab
• Accept the first line of a suggestion with Shift+Tab
• Accept the first word of a suggestion with Ctrl+B
• Control max text generation time
• Configure scope of context around the cursor
• Ring context with chunks from open and edited files and yanked text
• Supports very large contexts even on low-end hardware via smart context reuse
• Speculative FIM support
• Speculative Decoding support
• Display performance stats
• Better error handling and initialization checks
• Improved code completion quality with stop tokens

• vim-plug

  Plug 'akashjss/ollama.vim'

• Vundle

  cd ~/.vim/bundle
  git clone https://github.com/akashjss/ollama.vim

  Then add Plugin 'ollama.vim' to your .vimrc in the vundle#begin() section.

• lazy.nvim

  {
      'akashjss/ollama.vim',
  }

You can customize ollama.vim by setting the g:ollama_config variable.

Examples:

1. Configure LLM Model:

   " put before ollama.vim loads
   let g:ollama_config.model = "deepseek-coder-v2"

2. Disable the inline info:

   " put before ollama.vim loads
   let g:ollama_config = { 'show_info': 0 }

3. Same thing but setting directly

   let g:ollama_config.show_info = v:false

4. Disable auto FIM (Fill-In-the-Middle) completion with lazy.nvim

   {
       'akashjss/ollama.vim',
       init = function()
           vim.g.ollama_config = {
               auto_fim = false,
           }
       end,
   }

5. Changing accept line keymap

   let g:ollama_config.keymap_accept_full = "<C-S>"

The plugin supports the following configuration options:

• endpoint: Ollama server endpoint (default: http://127.0.0.1:11434/api/generate)
• model: LLM model to use (default: deepseek-coder-v2)
• api_key: API key for authentication (optional)
• n_prefix: Number of lines before cursor to include in context (default: 256)
• n_suffix: Number of lines after cursor to include in context (default: 64)
• n_predict: Maximum number of tokens to predict (default: 128)
• t_max_prompt_ms: Maximum time for prompt processing (default: 500)
• t_max_predict_ms: Maximum time for prediction (default: 1000)
• show_info: Show inference info (0=disabled, 1=statusline, 2=inline)
• auto_fim: Auto-trigger FIM on cursor movement (default: true)
• max_line_suffix: Max chars after cursor for auto-trigger (default: 8)
• max_cache_keys: Max cached completions (default: 250)
• ring_n_chunks: Max chunks for extra context (default: 16)
• ring_chunk_size: Max lines per chunk (default: 64)
• ring_scope: Range around cursor for gathering chunks (default: 1024)
• ring_update_ms: Chunk processing interval (default: 1000)
• keymap_trigger: Manual trigger key (default: )
• keymap_accept_full: Accept full suggestion (default: )
• keymap_accept_line: Accept first line (default: )
• keymap_accept_word: Accept first word (default: )

The plugin requires a ollama server instance to be running at g:ollama_config.endpoint.

brew install ollama

Follow the installation instructions at ollama.com

The plugin works best with code-focused models. Here are some recommended models:

• deepseek-coder-v2: Best overall code completion model
• codellama: Good for general code completion
• mistral: Good for general text and code completion
• neural-chat: Good for chat and code completion

To use a model, first pull it:

ollama pull deepseek-coder-v2

Then configure it in your vimrc:

let g:ollama_config.model = "deepseek-coder-v2"

The orange text is the generated suggestion. The green text contains performance stats for the FIM request.

The plugin aims to be very simple and lightweight while providing high-quality and performant local FIM completions, even on consumer-grade hardware. Key features include:

• Smart context management with ring buffer
• Efficient caching of completions
• Speculative decoding for faster responses
• Intelligent chunk selection and eviction
• Robust error handling and initialization checks

Contributions are welcome! Please feel free to submit a Pull Request.

MIT License