llm-call-tracing

To trace LLM API calls and capture latency and tokens, you instrument direct SDK invocations with structured span attributes that log model details, execution time, and success status for production observability.

LLM tracing supports LangChain integrations and multi-agent tool usage scenarios, providing end-to-end observability across complex workflows by logging structured telemetry for every underlying model interaction.

You can track token costs and success status for LLM calls in production using cost-aware tracing, which implements token-based cost accounting and structured error context for failed invocations.

The best way to instrument streaming LLM responses for observability is to use optional streaming support features that capture telemetry data progressively as the backend generates the output.

To capture error context when an LLM API call fails, you implement structured span attributes that record the failure outcome and surrounding invocation details, ensuring production instrumentation meets observability requirements.