Open, modular framework to build GenAI applications.

AlphaTrion is an open-source framework to help build GenAI applications, including experiment tracking, adaptive model routing, prompt optimization, performance evaluation and so on. The name comes after the oldest and wisest Transformer - AlphaTrion.

Still under active development.

Concepts

• Team: A Team is the highest-level organizational unit in AlphaTrion. It represents a group of users collaborating on projects and experiments.
• Project: A Project is a namespace-level abstraction that contains multiple experiments. It helps organize experiments related to a specific goal or topic.
• Experiment: An Experiment is a logic-level abstraction for organizing and managing a series of related runs. It allows users to group runs that share a common purpose or configuration.
• Run: A Run is a real execution instance of an experiment. It represents the actual execution of the code with the specified configuration and hyperparameters defined in the experiment.

Quick Start

Install from PyPI

pip install alphatrion

Install from Source

• Git clone the repository
• Run source start.sh to activate the virtual environment.

Initialize the Environment

Run the following command for setup:

cp .env.example .env & make up

You can login to pgAdmin at http://localhost:8081 to see the Postgres database with following credentials. Remember to register the server first.

Email:       alphatrion@inftyai.com
Password:    alphatr1on
ServerName:  alphatrion
HostName:    postgres
ServerPWD:   alphatr1on

You can also visit the Docker Registry UI at http://localhost:80 to see the local registry where the built images are stored.

Run a Simple Experiment

Below is a simple example with two approaches demonstrating how to create an experiment and log performance metrics.

import uuid

import alphatrion as alpha
from alphatrion import experiment, project

# Better to use a fixed UUID for the team and user in real scenarios.
alpha.init(team_id=uuid.uuid4(), user_id=uuid.uuid4())

async def your_task():
  # Run your code here then log metrics.
  await alpha.log_metrics({"accuracy": 0.95})

async with project.Project.setup(name="my_project"):
  async with experiment.CraftExperiment.start(name="my_experiment") as exp:
    task = exp.run(your_task) # use lambda or partial if you need to pass arguments to your_task
    await task.wait()

View Dashboard

The dashboard provides a web interface to explore projects, experiments, runs, and metrics through an intuitive UI.

Launch Dashboard

# Start the backend server (in one terminal)
alphatrion server

# Launch the dashboard (in another terminal)
alphatrion dashboard

The dashboard will automatically open in your browser at http://127.0.0.1:5173.

Options:

• --port <PORT>: Run on a custom port (default: 5173)
• --no-browser: Don't automatically open the browser

Example:

alphatrion dashboard --port 8080 --no-browser

Documentation:

• Dashboard Setup Guide - Complete setup and troubleshooting guide
• Dashboard CLI Guide - Using the dashboard CLI command
• Dashboard Architecture - Technical architecture and deployment patterns

Cleanup

make down

Contributing

We welcome contributions! Please refer to developer.md for more information on how to set up your development environment and contribute to the project.