weedleweedle / godot-copilot

AI-assisted development for the Godot engine.

Godot Copilot

AI-assisted development for the Godot engine.

Compatible with 4.x and 3.x.

What does this do?

Godot Copilot uses OpenAI APIs (other models/providers may be supported in the future) to retrieve AI-generated code completions.

After installing the plugin, simply press the selected keyboard shortcut to generate code in the code editor at the current caret position, directly within the engine!

How do i install this?

• Search for "Copilot" in the Godot asset library directly within the engine. Download the addon and enable in the project settings.
• Alternatively, navigate to the releases page and follow the instructions for the latest one
• You may also clone this repository and copy the copilot addon into the res://addons directory of your project

Afterwards, enable the addon in the project settings, and enter your OpenAI API key in the Copilot tab located on the right-hand dock.

Use the selected keyboard shortcut within the code editor to request completions.

How much will this cost me?

Godot Copilot currently supports three different models:

• text-davinci-003 (Completion interface)
  • $0.02 / 1K tokens
• gpt-3.5-turbo (Chat interface)
  • $0.002 / 1K tokens
• gpt-4 (Chat interface)
  • $0.03 / 1K tokens

Therefore, turbo is cheapest.

For each request, Copilot will attempt to send your entire current script to the model, up to a maximum length, after which the code will be trimmed.

Being a good engineer that doesn't work in 5k-line spaghetti-scripts pays. Literally!

Comprehensive example data on cost per month/day/hour for each model is appreciated! Please open an issue if you are willing to contribute!

Completion interface? Chat interface?

OpenAI uses different API interfaces for different models.

• Completion Interface (text-davinci-003)
  • Here, the model is simply given a prompt string and a suffix string, and will generate tokens it believes will fit between the two. Tailor-made for code completion!
• Chat Interface (gpt-3.5-turbo, gpt-4)
  • This is an abstraction layer categorizing prompts and completions as "messages", as you'd see in ChatGPT for example. Therefore, we need specific instructions to ask the model to INSERT code into a specific location. Both turbo and gpt-4 are smart enough to reliably follow these instructions, but they may sometimes result in missing spaces or indentation at the beginning of completions.

How well does this work?

GDScript is underrepresented in OpenAI's training data!

Do not expect performance matching that of popular languages such as Python or Javascript. This will not replace you... yet!

It is however perfectly capable of writing useful functions, generating pre-populated objects, and completing all the tedious code you don't want to write yourself.

One significant complication affecting performance is the changes introduced in Godot 4.0. The training data of all supported models largely cuts off in September 2021. This means some of the new syntax was included, some not, and the models know a combination of GDScript 1 and 2.

Godot Copilot automatically injects zero-shot explanations of some of the most important GDScript changes in the prompts! This helps model performance quite significantly, but doesn't resolve all issues. Expect some varying performance when it comes to concepts such as Callable or hallucinated Python methods.

Why no autocomplete?

A couple reasons:

• Godot's autocomplete system for CodeEdit nodes is not particularly well suited for AI-completions, as it does not preview completions directly within the text, but instead in a small popup that is not able to preview longer/multi-line predictions.
• The prefix-system it uses does not always handle runtime-additions of completions smoothly. Whether or not completions work is highly context dependent.
• Triggering LLM-requests on each autocompletion signal is expensive. As in, really expensive. Unlike Github Copilot, this can not be a one-size-fits-all payment model. You pay for what you use.

Testing has shown targeted completions on shortcut presses to be much more reliable and efficient.

Autocomplete may be an optional setting in the future if above issues can be sufficiently addressed.

Does this share sensitive data?

Your code will be sent to OpenAIs API endpoints. Since March 1st 2023, OpenAI states it will no longer use data sent via their APIs in future training runs.

Does this work with 3.x?

Yes! Please switch to the 3.x branch for the appropriate version of this addon.

Why should i use this over Github Copilot?

Do use Github Copilot when suitable! As long as this project depends on OpenAI APIs, Github Copilot is likely a better deal.

However, this does offer some unique benefits:

• Works directly in the editor, which may suit your existing workflow better
• Supports GPT-4, giving you access to a more capable model than Github Copilot currently offers

Can this do anything other than code completion?

I have run experiments with further automation capabilities, such as populating and modifying the active scene. Such features may be added in the future.

Other interesting usecases may be using embeddings to more accurately model cross-script functionality for your codebase, or using generative models for texture, sprite or 3D asset creation.

Please feel free to open issues requesting features or changes.