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

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

# ✅ 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

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

from opik.integrations.anthropic import track_anthropic

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

from opik.integrations.langchain import OpikTracer

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

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

# CRUD: create/get/list/update/delete
client.create_experiment(name="exp")
client.get_dataset(name="ds")

# Search for complex queries
client.search_spans(project_name="proj")
client.search_traces(project_name="proj")

# Batch for bulk operations
client.batch_create_items(...)