AutoDev: LLM-Based Coding Assistance Functions

This repository contains two projects:

• The AutoDev Python project providing the core functionality (./autodev), including

  • auto-completion models (that can suggest completions based on the current editing context)

    • fine-tuning of completion models to teach them new languages (or to teach them about your libraries, your code style, etc.)
    • quantitative & qualitative evaluation
    • optimization of models for inference (including quantization)

  • code-based assistance functions, where an instruction-following model is given a task based on an existing code snippet (e.g. reviewing code, adding comments or input checks, explaining code, etc.)

  • an inference service, which provides access to the above functions

  • question answering on document databases (including source code documents)

• A Java project implementing the AutoDev IntellIJ IDEA plugin which provides access to the coding assistance functions within JetBrains IDEs such as IntelliJ IDEA, PyCharm and others (./idea-plugin).

Please refer to the projects' individual README files for further information (linked above).

AutoDev in Action

Fine-Tuned Auto-Completion

Generating completions for the Ruby programming language based on a fine-tuned version of bigcode/santacoder, which originally knew only Python, Java and JavaScript:

Assistance Functions Built on Instruction-Following Models

Adding input checks to a function:

Identifying potential problems in a piece of code:

Structural Overview

Here's a structural overview showing the main components and their interactions:

• For auto-completion, the model is served directly by the AutoDev inference service, i.e. the model is always locally provided and is either an unmodified open-source model (from the Hugging Face Hub) or a fine-tuned version thereof. Fine-tuning may use community data or our own data.
• For other assistance functions built on instruction-following models, you have the option of using either a (fine-tuned) open-source model, as in the previous case, or a proprietary model (such as ChatGPT).