Code Generator Chat

Overview

Code Generator Chat is an instruction-lead AI chat assistant designed to write code of any language and integrate with existing CI/CD pipelines with the goal of improving code generation accuracy beyond basic Q&A interaction with an LLM.

This is achieved through a multi-agent conversation or better known as a group-chat. Each agent within this ecosystem has a focused role just like in a classic team.

By structuring a communication pipeline between agents, the team is able to break down a given task into smaller, more manageable parts and provide instructions or feedback to each other so that the final output is both validated to work but also meets all requirements given.

Agents and Their Roles

Agent workflow

The core agents include:

Manager

The Manager agent acts as the central coordinator, orchestrating the operation of the other agents. It is responsible for ensuring that each agent contributes effectively according to their specialized function and that the structured sequence is adhered to without deviation.

Manager Prompt

Planner

The Planner is the first agent in the sequence, responsible for receiving user input and outlining a strategic approach to the task at hand. This agent creates a blueprint that serves as the foundation for all subsequent actions.

Planner Prompt

Programmer

Following the Planner, the Programmer translates the strategic blueprint into actionable code. This phase is critical and hinges entirely on the framework laid out by the Planner.

Programmer Prompt

Optimizer

The Optimizer reviews and refines the Programmer's code to ensure maximum efficiency and effectiveness. This agent's role is to provide a crucial layer of quality control and make sure the code meets our high standards.

Optimizer Prompt

Tester

After the Optimizer has completed their review, the Tester agent takes over to validate the code and ensure that it functions as intended. This step is crucial for identifying any potential issues or bugs that may have been overlooked by generating valid tests cases and actually running them in a real environment.

[Tester Prompt]{src/agents/prompts/tester.jinja}

Check-In

The final step in the sequence involves the Check-In agent, who takes the finalized code and creates a GitHub Gist as a simple form of code check-in and review.

[Check-In Prompt]{src/agents/prompts/checkin.jinja}

Workflow Sequence

1. Initiation: Begins with the Planner, who sets the strategic foundation for the task.
2. Development: The Programmer then takes over, developing code based on the Planner's blueprint.
3. Refinement: Finally, the Optimizer reviews and enhances the code, focusing on efficiency and effectiveness.
4. Testing: The Tester validates the code by running test cases to ensure it functions
5. Finalization: When the Optimizer concludes that no further improvements can be made, and the Tester passes all test, the session is terminated with the command TERMINATE.

Deployment

Install Requirements

pip install -r requirements.txt

Start Backend

uvicorn main:app --host 0.0.0.0 --port 8000 --reload

Start Frontend

Change directory

cd frontend

Install dependencies (assuming NPM and Node.js are installed)

cd frontend
npm install 
npm run build

Start the frontend server

npm run dev

Examples:

• "Write a python function that finds all properties with a key of 'intf' within all nested child objects and return a list of all key value pairs that match. "

• "Write a python client that integrates with the Vectara RAG API. Use websockets for realtime updates of the query endpoint used in a user chat. The client will also need to lookup the corpus name using 'GET corpus' so the user can select the appropriate corpus for the chat. "

• "Write python unit tests that validates the functionality of the REST API endpoints in the following main.py file." <main.py attached>