mwin 0.1.2

Track OpenAI, Claude, Gemini and OpenAI-compatible models then give solutions to improve your agent system.

AT

AT: Track, log, and evaluate AI models. Supports OpenAI, Claude, Google API and custom PyTorch models.
Our goal is to make llm application more valuable and effortlessly improve llm capabilities.

Quickstart

You can use pip install AT. (It's not implemented now.)

pip install aitrace

OR pip install from source.

git clone https://github.com/yanghui1-arch/AITrace.git
cd src
pip install -e .

Then you need to configure AT through CLI.

aitrace configure

It needs an AITrace API key. You can get the apikey after logging http://localhost:5173. Finally use @track to track your llm input and output

from aitrace import track
from openai import OpenAI

openai_apikey = 'YOUR API KEY'


@track(
    project_name="aitrace_demo",
    tags=['test', 'demo'],
    track_llm=LLMProvider.OPENAI,    
)
def llm_classification(film_comment: str):
    prompt = "Please classify the film comment into happy, sad or others. Just tell me result. Don't output anything."
    cli = OpenAI(base_url='https://api.deepseek.com', api_key=openai_apikey)
    cli.chat.completions.create(
        messages=[{"role": "user", "content": f"{prompt}\nfilm_comment: {film_comment}"}],
        model="deepseek-chat"
    ).choices[0].message.content
    llm_counts(film_comment=film_comment)
    return "return value"

@track(
    project_name="aitrace_demo",
    tags=['test', 'demo', 'second_demo'],
    track_llm=LLMProvider.OPENAI,
)
def llm_counts(film_comment: str):
    prompt = "Count the film comment words. just output word number. Don't output anything others."
    cli = OpenAI(base_url='https://api.deepseek.com', api_key=openai_apikey)
    return cli.chat.completions.create(
        messages=[{"role": "user", "content": f"{prompt}\nfilm_comment: {film_comment}"}],
        model="deepseek-chat"
    ).choices[0].message.content

llm_classification("Wow! It sucks.")

Then it will output your llm trace. It is not supported to visualize now. I am developing it more and more quickly. Welcome to all contributions.

Development

AT project package manager is uv. If you are a beginner uver, please click uv link: uv official link

uv install
uv .venv/Script/activate

You can watch more detailed debug information by using --log-level=DEBUG or set AT_LOG_LEVEL=DEBUG for Windows or export AT_LOG_LEVEL=DEBUG for Linux and Mac.