Learn Weave with W&B Inference

This guide shows you how to use W&B Weave with W&B Inference. Using W&B Inference, you can build and trace LLM applications using live open-source models without setting up your own infrastructure or managing API keys from multiple providers. Just obtain your W&B API key and use it to interact with all models hosted by W&B Inference.

What you’ll learn

In this guide, you’ll:

• Set up Weave and W&B Inference
• Build a basic LLM application with automatic tracing
• Compare multiple models
• Evaluate model performance on a dataset
• View your results in the Weave UI

Prerequisites

Before you begin, you need a W&B account and an API key from https://wandb.ai/authorize.

Then, in a Python environment running version 3.9 or later, install the required libraries:

pip install weave openai

The openai library is installed because you use the standard openai client to interact with W&B Inference, regardless of which hosted model you’re actually calling. This allows you to swap between supported models by only changing the slug, and make use of any existing code you have that was written to use the OpenAI API.

Step 1: Trace your first LLM call

Start with a basic example that uses Llama 3.1-8B through W&B Inference.

When you run this code, Weave:

• Traces your LLM call automatically
• Logs inputs, outputs, latency, and token usage
• Provides a link to view your trace in the Weave UI

import weave
import openai

# Initialize Weave - replace with your-team/your-project
weave.init("<team-name>/my-first-weave-project")

# Create an OpenAI-compatible client pointing to W&B Inference
client = openai.OpenAI(
    base_url='https://api.inference.wandb.ai/v1',
    api_key="YOUR_WANDB_API_KEY",  # Replace with your actual API key
    project="<team-name>/my-first-weave-project",  # Required for usage tracking
)

# Decorate your function to enable tracing; use the standard OpenAI client
@weave.op()
def ask_llama(question: str) -> str:
    response = client.chat.completions.create(
        model="meta-llama/Llama-3.1-8B-Instruct",
        messages=[
            {"role": "system", "content": "You are a helpful assistant."},
            {"role": "user", "content": question}
        ],
    )
    return response.choices[0].message.content

# Call your function - Weave automatically traces everything
result = ask_llama("What are the benefits of using W&B Weave for LLM development?")
print(result)

Step 2: Build a text summarization application

Next, try running this code, which is a basic summarization app that shows how Weave traces nested operations:

import weave
import openai

# Initialize Weave - replace with your-team/your-project
weave.init("<team-name>/my-first-weave-project")

client = openai.OpenAI(
    base_url='https://api.inference.wandb.ai/v1',
    api_key="YOUR_WANDB_API_KEY",  # Replace with your actual API key
    project="<team-name>/my-first-weave-project",  # Required for usage tracking
)

@weave.op()
def extract_key_points(text: str) -> list[str]:
    """Extract key points from a text."""
    response = client.chat.completions.create(
        model="meta-llama/Llama-3.1-8B-Instruct",
        messages=[
            {"role": "system", "content": "Extract 3-5 key points from the text. Return each point on a new line."},
            {"role": "user", "content": text}
        ],
    )
    # Returns response without blank lines
    return [line for line in response.choices[0].message.content.strip().splitlines() if line.strip()]

@weave.op()
def create_summary(key_points: list[str]) -> str:
    """Create a concise summary based on key points."""
    points_text = "\n".join(f"- {point}" for point in key_points)
    response = client.chat.completions.create(
        model="meta-llama/Llama-3.1-8B-Instruct",
        messages=[
            {"role": "system", "content": "Create a one-sentence summary based on these key points."},
            {"role": "user", "content": f"Key points:\n{points_text}"}
        ],
    )
    return response.choices[0].message.content

@weave.op()
def summarize_text(text: str) -> dict:
    """Main summarization pipeline."""
    key_points = extract_key_points(text)
    summary = create_summary(key_points)
    return {
        "key_points": key_points,
        "summary": summary
    }

# Try it with sample text
sample_text = """
The Apollo 11 mission was a historic spaceflight that landed the first humans on the Moon 
on July 20, 1969. Commander Neil Armstrong and lunar module pilot Buzz Aldrin descended 
to the lunar surface while Michael Collins remained in orbit. Armstrong became the first 
person to step onto the Moon, followed by Aldrin 19 minutes later. They spent about 
two and a quarter hours together outside the spacecraft, collecting samples and taking photographs.
"""

result = summarize_text(sample_text)
print("Key Points:", result["key_points"])
print("\nSummary:", result["summary"])

Step 3: Compare multiple models

W&B Inference provides access to multiple models. Use the following code to compare the performance of Llama and DeepSeek’s respective responses:

import weave
import openai

# Initialize Weave - replace with your-team/your-project
weave.init("<team-name>/my-first-weave-project")

client = openai.OpenAI(
    base_url='https://api.inference.wandb.ai/v1',
    api_key="YOUR_WANDB_API_KEY",  # Replace with your actual API key
    project="<team-name>/my-first-weave-project",  # Required for usage tracking
)

# Define a Model class to compare different LLMs
class InferenceModel(weave.Model):
    model_name: str
    
    @weave.op()
    def predict(self, question: str) -> str:
        response = client.chat.completions.create(
            model=self.model_name,
            messages=[
                {"role": "user", "content": question}
            ],
        )
        return response.choices[0].message.content

# Create instances for different models
llama_model = InferenceModel(model_name="meta-llama/Llama-3.1-8B-Instruct")
deepseek_model = InferenceModel(model_name="deepseek-ai/DeepSeek-V3-0324")

# Compare their responses
test_question = "Explain quantum computing in one paragraph for a high school student."

print("Llama 3.1 8B response:")
print(llama_model.predict(test_question))
print("\n" + "="*50 + "\n")
print("DeepSeek V3 response:")
print(deepseek_model.predict(test_question))

Step 4: Evaluate model performance

Evaluate how well a model performs on a Q&A task using Weave’s built-in EvaluationLogger. This provides structured evaluation tracking with automatic aggregation, token usage capture, and rich comparison features in the UI.

Append the following code to the script you used in step 3:

from typing import Optional
from weave import EvaluationLogger

# Create a simple dataset
dataset = [
    {"question": "What is 2 + 2?", "expected": "4"},
    {"question": "What is the capital of France?", "expected": "Paris"},
    {"question": "Name a primary color", "expected_one_of": ["red", "blue", "yellow"]},
]

# Define a scorer
@weave.op()
def accuracy_scorer(expected: str, output: str, expected_one_of: Optional[list[str]] = None) -> dict:
    """Score the accuracy of the model output."""
    output_clean = output.strip().lower()
    
    if expected_one_of:
        is_correct = any(option.lower() in output_clean for option in expected_one_of)
    else:
        is_correct = expected.lower() in output_clean
    
    return {"correct": is_correct, "score": 1.0 if is_correct else 0.0}

# Evaluate a model using Weave's EvaluationLogger
def evaluate_model(model: InferenceModel, dataset: list[dict]):
    """Run evaluation on a dataset using Weave's built-in evaluation framework."""
    # Initialize EvaluationLogger BEFORE calling the model to capture token usage
    # This is especially important for W&B Inference to track costs
    # Convert model name to a valid format (replace non-alphanumeric chars with underscores)
    safe_model_name = model.model_name.replace("/", "_").replace("-", "_").replace(".", "_")
    eval_logger = EvaluationLogger(
        model=safe_model_name,
        dataset="qa_dataset"
    )
    
    for example in dataset:
        # Get model prediction
        output = model.predict(example["question"])
        
        # Log the prediction
        pred_logger = eval_logger.log_prediction(
            inputs={"question": example["question"]},
            output=output
        )
        
        # Score the output
        score = accuracy_scorer(
            expected=example.get("expected", ""),
            output=output,
            expected_one_of=example.get("expected_one_of")
        )
        
        # Log the score
        pred_logger.log_score(
            scorer="accuracy",
            score=score["score"]
        )
        
        # Finish logging for this prediction
        pred_logger.finish()
    
    # Log summary - Weave automatically aggregates the accuracy scores
    eval_logger.log_summary()
    print(f"Evaluation complete for {model.model_name} (logged as: {safe_model_name}). View results in the Weave UI.")

# Compare multiple models - a key feature of Weave's evaluation framework
models_to_compare = [
    llama_model,
    deepseek_model,
]

for model in models_to_compare:
    evaluate_model(model, dataset)

# In the Weave UI, navigate to the Evals tab to compare results across models

Running these examples returns links to the traces in the terminal. Click any link to view traces in the Weave UI.

In the Weave UI, you can:

• Review a timeline of all your LLM calls
• Inspect inputs and outputs for each operation
• View token usage and estimated costs (automatically captured by EvaluationLogger)
• Analyze latency and performance metrics
• Navigate to the Evals tab to see aggregated evaluation results
• Use the Compare feature to analyze performance across different models
• Page through specific examples to see how different models performed on the same inputs

Available models

For a complete list of available models, see the Available Models section in the W&B Inference documentation.

Next steps

• Use the Playground: Try models interactively in the Weave Playground
• Build evaluations: Learn about systematic evaluation of your LLM applications
• Try other integrations: Weave works with OpenAI, Anthropic, and many more

Troubleshooting