python-sdk

Python SDK

Three-Layer Architecture

Layer 1: Public API      (opik.Opik, @opik.track)
    ↓
Layer 2: Message Processing   (queue, batching, retry)
    ↓
Layer 3: REST Client     (OpikApi, HTTP)

Critical Gotchas

Flush Before Exit

# ✅ REQUIRED for async operations
client = opik.Opik()
# ... tracing operations ...
client.flush()  # Must call before exit!

Async vs Sync Operations

Async (via message queue) - fire-and-forget:

• trace(), span()
• log_traces_feedback_scores()
• experiment.insert()

Sync (blocking, returns data):

• create_dataset(), get_dataset()
• create_prompt(), get_prompt()
• search_traces(), search_spans()

Lazy Imports for Integrations

# ✅ GOOD - integration files assume dependency exists
import anthropic  # Only imported when user uses integration

# ❌ BAD - importing at package level
from opik.integrations import anthropic  # Would fail if not installed

Integration Patterns

Pattern Selection

Library has callbacks? → Pure Callback (LangChain, LlamaIndex)
No callbacks?         → Method Patching (OpenAI, Anthropic)
Callbacks unreliable? → Hybrid (ADK)

Method Patching (OpenAI, Anthropic)

from opik.integrations.anthropic import track_anthropic

client = anthropic.Anthropic()
tracked_client = track_anthropic(client)  # Wraps methods

Callback-Based (LangChain)

from opik.integrations.langchain import OpikTracer

tracer = OpikTracer()
chain.invoke(input, config={"callbacks": [tracer]})

Decorator-Based

@opik.track
def my_function(input: str) -> str:
    # Auto-creates span, captures input/output
    return process(input)

Dependency Policy

• Avoid adding new dependencies
• Use conditional imports for integrations
• Keep version bounds flexible: >=2.0.0,<3.0.0

Batching System

Messages batch together for efficiency:

• Flush triggers: time (1s), size (100), memory (50MB), manual
• Reduces HTTP overhead significantly