Don't trust AI autocomplete
One thing I learned in my week of running local AI models is that for autocomplete (sometimes called ghost text or inline suggestions) the model has to be extremely fast, sub-one-second responses, and this means that the model has to be quite small. Models capable of the speed required for autocomplete are very different from the models that you'd use as a coding assistant. The autocomplete suggestions that you are getting are not the latest Claude Sonnet or Opus models and you should be aware of this before you accept them.
Inline suggestions also don't have any additional context or knowledge of things like your schema, so it will make a best guess, and they often get things wrong. They are fine for what they are, and they're helpful to avoid writing repetitive boilerplate code.