Observability

djangosdk supports three observability backends: LangSmith, Langfuse, and OpenTelemetry.

Configuration

AI_SDK = {
    "OBSERVABILITY": {
        "BACKEND": "langsmith",  # "langsmith" | "langfuse" | "opentelemetry" | None
    },
}

LangSmith

LangSmith provides tracing, evaluation, and monitoring for AI applications.

uv add "djangosdk[langsmith]"

import os
os.environ["LANGCHAIN_API_KEY"] = "ls__..."
os.environ["LANGCHAIN_TRACING_V2"] = "true"

AI_SDK = {
    "OBSERVABILITY": {"BACKEND": "langsmith"},
}

Every agent.handle() call is automatically traced as a LangSmith run.

Langfuse

Langfuse is an open-source LLM observability platform.

uv add "djangosdk[langfuse]"

AI_SDK = {
    "OBSERVABILITY": {
        "BACKEND": "langfuse",
        "LANGFUSE_PUBLIC_KEY": env("LANGFUSE_PUBLIC_KEY"),
        "LANGFUSE_SECRET_KEY": env("LANGFUSE_SECRET_KEY"),
        "LANGFUSE_HOST": "https://cloud.langfuse.com",
    },
}

OpenTelemetry

For custom observability pipelines, use the OpenTelemetry backend:

uv add "djangosdk[opentelemetry]"

AI_SDK = {
    "OBSERVABILITY": {"BACKEND": "opentelemetry"},
}

Traces are emitted to the configured OpenTelemetry exporter (OTLP, Jaeger, Zipkin, etc.).

# Configure your OTel exporter as usual
from opentelemetry import trace
from opentelemetry.sdk.trace import TracerProvider
from opentelemetry.sdk.trace.export import BatchSpanProcessor
from opentelemetry.exporter.otlp.proto.grpc.trace_exporter import OTLPSpanExporter

provider = TracerProvider()
provider.add_span_processor(BatchSpanProcessor(OTLPSpanExporter()))
trace.set_tracer_provider(provider)

What Gets Traced

Event	Data captured
Agent start	agent_class, model, provider, prompt
Agent complete	response_text, token_usage, latency_ms
Tool call	tool_name, arguments, result
Cache hit/miss	cache_read_tokens
Provider failover	from_provider, to_provider

Cost Tracking

The analytics module tracks token costs per provider:

from djangosdk.analytics.cost import CostCalculator

calc = CostCalculator()
cost = calc.calculate(
    provider="openai",
    model="gpt-4.1",
    prompt_tokens=500,
    completion_tokens=100,
)
print(f"Estimated cost: ${cost:.6f}")