Track LLM inputs & outputs

Follow these steps to track your first call

Try the new W&B Inference service free. It gives you access to AI models through an API and the Weave Playground.

• Developer docs
• Product page

Follow these steps to track your first call. You can also

1. Install Weave and get your API key

Install weave

First, install the weave library:

Python

pip install weave

Get your API key

Create a W&B account at https://wandb.ai. Then copy your API key from https://wandb.ai/authorize.

2. Log a trace to a new project

Follow these steps to track your first project:

• Import the weave library
• Call weave.init('project-name') to start tracking
  • The system asks for your API key if you haven’t logged in yet
  • To log to a team, use team-name/project-name
  • Set the WANDB_API_KEY environment variable to skip the login prompt
• Add @weave.op() to the functions you want to track

This example uses OpenAI. You need an OpenAI API key.

Python

import weave
from openai import OpenAI

client = OpenAI()

# Weave tracks the inputs, outputs and code of this function
@weave.op()
def extract_dinos(sentence):
    response = client.chat.completions.create(
        model="gpt-4o",
        messages=[
            {
                "role": "system",
                "content": """Extract dinosaurs from the text. Return JSON with:
- dinosaurs: list of dinosaurs
- name: scientific name  
- common_name: nickname
- diet: herbivore or carnivore."""
            },
            {
                "role": "user",
                "content": sentence
            }
        ],
        response_format={ "type": "json_object" }
    )
    return response.choices[0].message.content


# Start the weave project
weave.init('jurassic-park')

sentence = """I saw a T. rex chase a Triceratops. 
The T. rex eats meat. The Triceratops eats plants.
A Brachiosaurus ate leaves from tall trees nearby."""

result = extract_dinos(sentence)
print(result)

When you call extract_dinos, Weave shows a link to view your trace.

3. Automatic AI library logging

Weave tracks calls to many AI services:

• OpenAI
• Anthropic
• Many more AI libraries

Weave logs:

• AI model info
• Token usage
• Cost

To track other AI tools, wrap them with @weave.op().

4. See traces in your project

Great job. Now Weave captures data every time you call your function. It also tracks code changes.

Next steps

• Learn to track data flow in your app
• Read about Tracing to see how Weave works
• Check out Integrations for all AI providers
• Try the Playground to test different models