Rusty - Automate Your Code Generation using Gemini

Rusty is a Rust-based CLI application that generates, builds, and tests backend server code in Rust based on user prompts.

Getting Started

Create .env file

Within the created .env file, paste the following:

API_KEY = YOUR_GEMINI_API_KEY

Update Path

Update constants in the src/helpers/general path.

Build and Run Project

cargo build
cargo run

Working

The project tries to mimic industrial code development with LLM agents.

LLM Agents:

• Managing Agent
• Solution Architect Agent
• Backend Agent
• Can be extended depending on use cases...

Managing Agent

• Converts user prompts into a more concise and technical version.
• Manages the other agents.

Solution Architect Agent

• Defines the project scope.
• Fetches and tests external APIs for the backend agent.

Backend Agent

• Builds initial Rust code using given prompts and code templates (Junior Developer).
• Improves the initial Rust code and verifies that it meets all user requirements (Senior Developer).
• Fixes broken or buggy code.
• Creates an API schema for frontend developers or end users.

Prompt Format

For the actual prompts, check src/ai_functions and src/helpers/general. Here, I have discussed the general format of the actual prompt.

Function: fn function_name(input) { // Comments specifying what the code will output. }
Instruction: Given the input: [YOUR_INPUT], output what the above code outputs.

We are prompting the model to output the actual result of the function. However, the function contains only commands. Therefore, models with high reasoning are required for this project.

Observations

For this project, I used both GPT-4 and Gemini Pro. Here are my observations so far; however, these might change in the future.

• In terms of quality, I felt GPT-4 had an upper hand over Gemini Pro. I couldn't find a perfect temperature for Gemini; with a higher temperature, there is much hallucination, and with a lower temperature, it strictly adheres to the template.
• In terms of speed, Gemini Pro is super fast.
• Both the models perform poorly in debugging and code fixing.
• If you are a person who loves to code along with Rusty, use Gemini.

Future Improvements

• Add more Task-Specific LLM Agents.
• Introduce Chain of Thoughts prompting.
• Write tooling to test all methods; for now, the code only tests the GET Method.
• Add front-end LLM Agents.
• Introduce local open-source LLMs.

Special Mention

Shaun McDonogh YouTube (Thank you for the amazing Rust Course 💖)
Karun A GitHub (Thank you for the GPT-4 API key 💖)