Skip to main content

Autocomplete Role

An "autocomplete model" is an LLM that is trained on a special format called fill-in-the-middle (FIM). This format is designed to be given the prefix and suffix of a code file and predict what goes between. This task is very specific, which on one hand means that the models can be smaller (even a 3B parameter model can perform well). On the other hand, this means that Chat models, though larger, will often perform poorly even with extensive prompting.

In Continue, autocomplete models are used to display inline Autocomplete suggestions as you type. Autocomplete models are designated by adding the autocomplete to the model's roles in config.yaml.

Visit the Autocomplete Deep Dive for recommended models and more details.

Prompt templating​

You can customize the prompt template used when autocomplete happens by setting the promptTemplates.autocomplete property in your model configuration. Continue uses Handlebars syntax for templating.

Available variables for the apply template:

  • {{{prefix}}} - the code before your cursor
  • {{{suffix}}} - the code after your cursor
  • {{{filename}}} - the name of the file your cursor currently is
  • {{{reponame}}} - the name of the folder where the codebase is
  • {{{language}}} - the name of the programming language in full (ex. Typescript)

Example:

models:
  - name: My Custom Autocomplete Template
    provider: ollama
    model: qwen2.5-coder:1.5b
    promptTemplates:
      autocomplete: |
        `
        globalThis.importantFunc = importantFunc
        <|fim_prefix|>{{{prefix}}}<|fim_suffix|>{{{suffix}}}<|fim_middle|>
        `