snowglobe 0.9.0a2

client server for usage with snowglobe experiments

Snowlgobe Connect SDK

The Snowglobe Connect SDK helps you connect your AI agents to Snowglobe. It sends simulated user messages to your LLM-based application during experiments. Your application should process these messages and return a response, enabling simulated conversations and custom code based risk assessment.

Installation

# Install client
pip install snowglobe

If using uv, set the --prerelease=allow flag

uv pip install --prerelease=allow snowglobe

snowglobe-connect commands

snowglobe-connect auth  # Sets up your API key
snowglobe-connect init  # Initializes a new agent connection and creates an agent wrapper file
snowglobe-connect test  # Tests your agent connection
snowglobe-connect start  # Starts the process of processing simulated user messages
snowglobe-connect --help

When using one of our specific preview environments in .snowgloberc one can override our server's URL with:

CONTROL_PLANE_URL=

Sample custom llm usage in agent wrapper file

Each agent wrapper file resides in the root directory of your project, and is named after the agent (e.g. My Agent Name becomes my_agent_name.py).

from snowglobe.client import CompletionRequest, CompletionFunctionOutputs
from openai import OpenAI
import os

client = OpenAI(api_key=os.getenv("SNOWGLOBE_API_KEY"))

def completion_fn(request: CompletionRequest) -> CompletionFunctionOutputs:
    """
    Process a scenario request from Snowglobe.
    
    This function is called by the Snowglobe client to process requests. It should return a
    CompletionFunctionOutputs object with the response content.

    Example CompletionRequest:
    CompletionRequest(
        messages=[
            SnowglobeMessage(role="user", content="Hello, how are you?", snowglobe_data=None),
        ]
    )

    Example CompletionFunctionOutputs:
    CompletionFunctionOutputs(response="This is a string response from your application")
    
    Args:
        request (CompletionRequest): The request object containing the messages.

    Returns:
        CompletionFunctionOutputs: The response object with the generated content.
    """

    # Process the request using the messages. Example:
    messages = request.to_openai_messages()
    response = client.chat.completions.create(
        model="gpt-4o-mini",
        messages=messages
    )
    return CompletionFunctionOutputs(response=response.choices[0].message.content)

Operator tuning

The wrapper exposes ~12 environment variables for concurrency, inbound rate limits, polling cadence, and HTTP retry. Defaults are tuned for low-to-medium throughput; raise them together (not individually) for higher load.

See .snowglobe/TUNING.md for the full knob reference, recipes, and the ratio caveat between dispatch concurrency and the inbound rate limiter.

Tracing & Telemetry

See our telemetry instrumentation extensions for snowglobe to help you track insightful traces from our agent:

• MLflow: https://pypi.org/project/snowglobe-telemetry-mlflow
• Arize: https://pypi.org/project/snowglobe-telemetry-openinference