HoneyHive Python SDK

A comprehensive Python SDK for HoneyHive, providing LLM observability, evaluation, and tracing capabilities with OpenTelemetry integration.

🚀 Features

• OpenTelemetry Integration - Full OTEL compliance with custom span processor and exporter
• Automatic Session Management - Seamless session creation and management
• Decorator Support - Easy-to-use @trace (unified sync/async), @atrace, and @trace_class decorators
• Context Managers - start_span and enrich_span for manual span management
• HTTP Instrumentation - Automatic HTTP request tracing
• Baggage Support - Context propagation across service boundaries
• Experiment Harness Integration - Automatic experiment tracking with MLflow, Weights & Biases, and Comet support
• Real-time API Integration - Direct integration with HoneyHive backend services
• Comprehensive Testing - Full test suite with 203 passing tests

📦 Installation

Choose Your Instrumentor Type:

HoneyHive supports both OpenInference (lightweight) and OpenLLMetry (enhanced metrics) instrumentors.

Option A: OpenInference (Recommended for Beginners)

# Install with OpenAI integration (most common)
pip install honeyhive[openinference-openai]

# Install with Anthropic integration
pip install honeyhive[openinference-anthropic]

# Install with Google AI integration
pip install honeyhive[openinference-google-ai]

# Install with multiple providers
pip install honeyhive[openinference-openai,openinference-anthropic,openinference-google-ai]

# Install all OpenInference integrations
pip install honeyhive[all-openinference]

Option B: OpenLLMetry (Enhanced Metrics)

# Install with OpenAI integration (enhanced metrics)
pip install honeyhive[traceloop-openai]

# Install with Anthropic integration
pip install honeyhive[traceloop-anthropic]

# Install with Google AI integration
pip install honeyhive[traceloop-google-ai]

# Install with multiple providers
pip install honeyhive[traceloop-openai,traceloop-anthropic,traceloop-google-ai]

# Install all OpenLLMetry integrations
pip install honeyhive[all-traceloop]

Option C: Mix Both Types

# Strategic mixing based on your needs
pip install honeyhive[traceloop-openai,openinference-anthropic]

Basic Installation (manual instrumentor setup required):

pip install honeyhive

📋 Including in Your Project

For detailed guidance on including HoneyHive in your pyproject.toml, see our pyproject.toml Integration Guide.

🔧 Quick Start

Basic Usage

from honeyhive import HoneyHiveTracer, trace

# Initialize tracer
tracer = HoneyHiveTracer.init(
    api_key="your-api-key",
    source="production"
)

# Use unified decorator for automatic tracing (works with both sync and async)
@trace(event_type="demo", event_name="my_function")
def my_function():
    return "Hello, World!"

@trace(event_type="demo", event_name="my_async_function")
async def my_async_function():
    await asyncio.sleep(0.1)
    return "Hello, Async World!"

# Manual span management
with tracer.start_span("custom-operation"):
    # Your code here
    pass

# With HTTP tracing enabled (new simplified API)
tracer = HoneyHiveTracer.init(
    api_key="your-api-key",
    source="production",
    disable_http_tracing=False  # project derived from API key
)

Initialization

The HoneyHiveTracer.init() method is the recommended way to initialize the tracer:

from honeyhive import HoneyHiveTracer

# Standard initialization
tracer = HoneyHiveTracer.init(
    api_key="your-api-key",
    source="production"  # project derived from API key
)

# With custom server URL for self-hosted deployments
tracer = HoneyHiveTracer.init(
    api_key="your-api-key",
    source="production",
    server_url="https://custom-server.com"  # project derived from API key
)

Enhanced Features Available

from honeyhive import HoneyHiveTracer
from openinference.instrumentation.openai import OpenAIInstrumentor

# All features are available in the init method
tracer = HoneyHiveTracer.init(
    api_key="your-api-key",
    source="production",
    test_mode=True,  # Test mode support
    instrumentors=[OpenAIInstrumentor()],  # Auto-integration
    disable_http_tracing=True  # Performance control
)

✅ The init method now supports ALL constructor features!

OpenInference Integration

from honeyhive import HoneyHiveTracer
from openinference.instrumentation.openai import OpenAIInstrumentor

# Initialize tracer with OpenInference instrumentor (recommended pattern)
tracer = HoneyHiveTracer.init(
    api_key="your-api-key",
    source="production",
    instrumentors=[OpenAIInstrumentor()]  # Auto-integration
)

# OpenInference automatically traces OpenAI calls
import openai
response = openai.ChatCompletion.create(
    model="gpt-3.5-turbo",
    messages=[{"role": "user", "content": "Hello!"}]
)

Enriching Spans and Sessions

v1.0+ Recommended Pattern: Instance Methods

from honeyhive import HoneyHiveTracer

# Initialize tracer
tracer = HoneyHiveTracer.init(
    api_key="your-api-key",
)

# Use instance methods for enrichment (PRIMARY - Recommended)
@tracer.trace(event_type="tool")
def my_function(input_data):
    result = process_data(input_data)

    # ✅ Instance method (PRIMARY pattern in v1.0+)
    tracer.enrich_span(
        metadata={"input": input_data, "result": result},
        metrics={"processing_time_ms": 150}
    )

    return result

# Enrich session with user properties
tracer.enrich_session(
    user_properties={"user_id": "user-123", "plan": "premium"}
)

Legacy Pattern: Free Functions (Backward Compatibility)

For backward compatibility, the free function pattern from v0.2.x still works:

from honeyhive import trace, enrich_span, enrich_session

# Free functions with automatic tracer discovery (LEGACY)
@trace(event_type="tool")
def my_function(input_data):
    result = process_data(input_data)

    # Free function with auto-discovery (backward compatible)
    enrich_span(
        metadata={"input": input_data, "result": result},
        metrics={"processing_time_ms": 150}
    )

    return result

# Enrich session via free function
enrich_session(user_properties={"user_id": "user-123"})

⚠️ Deprecation Notice: Free functions will be deprecated in v2.0. We recommend migrating to instance methods for new code.

Why Instance Methods?

• ✅ Explicit tracer reference (no auto-discovery overhead)
• ✅ Better multi-instance support (multiple tracers in same process)
• ✅ Clearer code (explicit is better than implicit)
• ✅ Future-proof (primary pattern going forward)

🏗️ Architecture

Core Components

src/honeyhive/
├── api/                    # API client implementations
│   ├── client.py          # Main API client
│   ├── configurations.py  # Configuration management
│   ├── datapoints.py      # Data point operations
│   ├── datasets.py        # Dataset operations
│   ├── events.py          # Event management
│   ├── evaluations.py     # Evaluation operations
│   ├── metrics.py         # Metrics operations
│   ├── projects.py        # Project management
│   ├── session.py         # Session operations
│   └── tools.py           # Tool operations
├── tracer/                 # OpenTelemetry integration
│   ├── otel_tracer.py     # Main tracer implementation
│   ├── span_processor.py  # Custom span processor
│   ├── span_exporter.py   # Custom span exporter
│   ├── decorators.py      # Tracing decorators
│   └── http_instrumentation.py # HTTP request tracing
├── evaluation/             # Evaluation framework
│   └── evaluators.py      # Evaluation decorators
├── models/                 # Pydantic models
│   └── generated.py       # Auto-generated from OpenAPI
└── utils/                  # Utility functions
    ├── config.py          # Configuration management
    ├── connection_pool.py # HTTP connection pooling
    ├── retry.py           # Retry mechanisms
    └── logger.py          # Logging utilities

Key Design Principles

1. Singleton Pattern - Single tracer instance per application
2. Environment Configuration - Flexible configuration via environment variables
3. Graceful Degradation - Fallback mechanisms for missing dependencies
4. Test Isolation - Comprehensive test suite with proper isolation
5. OpenTelemetry Compliance - Full OTEL standard compliance

⚙️ Configuration

API keys

The HoneyHive API client selects a credential per operation:

See API Keys to create the credentials your workload needs. A tracing-only process can use an ingestion key alone. Data plane fine-grained keys cannot send telemetry.

Data plane API key

Set HH_DATA_PLANE_API_KEY to a complete hh_fgdp_ key, or pass data_plane_api_key to HoneyHive(). For supported methods, pass project_id even when the key is rooted at that project:

import os

from honeyhive import HoneyHive

client = HoneyHive()
charts = client.charts.list(project_id=os.environ["PROJECT_ID"])
print(charts.data)

This example requires project.chart.list. Set PROJECT_ID to the project ID from Settings > Project > API Keys > Data Plane. Set HH_API_URL or pass base_url to HoneyHive() if the project's data plane URL differs from the default.

For a method with both legacy and scoped routes, omitting project_id selects the legacy route and classic project key, even if a data plane key is configured. Omitting project_id from such a method is deprecated. Passing project_id selects the scoped route and requires a data plane key; there is no fallback to api_key or HH_API_KEY. See the credential matrix.

A wrong key type or incomplete value in data_plane_api_key or ingestion_api_key (or their environment variables) raises ValueError at initialization. A call whose required credential is absent raises MissingApiKeyError before sending a request.

Environment Variables

Experiment Harness Variables

HTTP Client Configuration

🤝 Contributing

Want to contribute to HoneyHive? See CONTRIBUTING.md for development setup and guidelines.