withorbit-sdk 0.1.2

Orbit - AI Cost Analytics SDK. Track, monitor, and optimize your AI spend.

Orbit SDK for Python

Track, monitor, and optimize your AI spend across OpenAI, Anthropic, and other LLM providers.

Installation

pip install withorbit-sdk

# With OpenAI support
pip install withorbit-sdk[openai]

# With Anthropic support
pip install withorbit-sdk[anthropic]

# With all providers
pip install withorbit-sdk[all]

Quick Start

1. Get your API key

Sign up at Orbit and create an API key.

2. Initialize the SDK

from withorbit_sdk import Orbit

orbit = Orbit(
    api_key="orb_live_xxxxxxxxxxxxxxxxxxxxxxxx",
    default_feature="my-app",  # Optional: default feature for all events
)

3. Track your LLM calls

Option A: Automatic tracking (Recommended)

Wrap your OpenAI or Anthropic client for automatic tracking:

from openai import OpenAI
from withorbit_sdk import Orbit, WrapperOptions

orbit = Orbit(api_key="orb_live_xxx")
openai = orbit.wrap_openai(OpenAI(), WrapperOptions(feature="chat-assistant"))

# All API calls are now automatically tracked!
response = openai.chat.completions.create(
    model="gpt-4o",
    messages=[{"role": "user", "content": "Hello, world!"}],
)

Works with Anthropic too:

from anthropic import Anthropic
from withorbit_sdk import Orbit, WrapperOptions

orbit = Orbit(api_key="orb_live_xxx")
anthropic = orbit.wrap_anthropic(Anthropic(), WrapperOptions(feature="document-analysis"))

message = anthropic.messages.create(
    model="claude-3-opus-20240229",
    max_tokens=1024,
    messages=[{"role": "user", "content": "Analyze this document..."}],
)

Option B: Manual tracking

For other providers or custom implementations:

from withorbit_sdk import Orbit

orbit = Orbit(api_key="orb_live_xxx")

# Track a successful request
orbit.track(
    model="gpt-4o",
    input_tokens=150,
    output_tokens=50,
    latency_ms=1234,
    feature="summarization",
    environment="production",
)

# Track an error
orbit.track_error(
    model="gpt-4o",
    error_type="rate_limit_exceeded",
    error_message="Rate limit exceeded",
    feature="chat-assistant",
    input_tokens=150,
)

Configuration

from withorbit_sdk import Orbit, OrbitConfig

orbit = Orbit(config=OrbitConfig(
    # Required
    api_key="orb_live_xxx",

    # Optional
    base_url="https://app.withorbit.io/api/v1",  # Custom API endpoint
    default_feature="my-app",                    # Default feature name
    default_environment="production",            # 'production' | 'staging' | 'development'
    debug=False,                                 # Enable debug logging

    # Batching (for high-volume applications)
    batch_events=True,       # Batch events before sending
    batch_size=10,           # Max events per batch
    batch_interval=5.0,      # Max seconds before sending batch

    # Reliability
    retry=True,              # Retry failed requests
    max_retries=3,           # Max retry attempts
))

Feature Attribution

Features are Orbit's key differentiator - they let you see exactly which parts of your application are consuming AI resources:

# Track different features
orbit.track(
    model="gpt-4o",
    input_tokens=100,
    output_tokens=50,
    feature="chat-assistant",  # Attribute to chat feature
)

orbit.track(
    model="gpt-4o",
    input_tokens=500,
    output_tokens=200,
    feature="document-analysis",  # Attribute to doc analysis
)

Then in the Orbit dashboard, you'll see:

• Cost breakdown by feature
• Request volume by feature
• Error rates by feature
• And more!

Context Manager Support

from withorbit_sdk import Orbit

with Orbit(api_key="orb_live_xxx") as orbit:
    orbit.track(model="gpt-4o", input_tokens=100, output_tokens=50)
# Automatically flushes on exit

Graceful Shutdown

For long-running processes, flush events before exit:

# Before your process exits
orbit.shutdown()

Event Properties

Property	Type	Required	Description
model	str	Yes	Model name (e.g., 'gpt-4o', 'claude-3-opus')
input_tokens	int	Yes	Number of input tokens
output_tokens	int	Yes	Number of output tokens
provider	str	No	Provider name (auto-detected if not provided)
latency_ms	int	No	Request latency in milliseconds
feature	str	No	Feature name for attribution
environment	str	No	Environment ('production', 'staging', 'development')
status	str	No	Request status ('success', 'error', 'timeout')
error_type	str	No	Error type if status is 'error'
error_message	str	No	Error message if status is 'error'
user_id	str	No	Your application's user ID
session_id	str	No	Session ID for grouping requests
request_id	str	No	Unique request ID for tracing
metadata	dict	No	Additional key-value metadata

License

MIT