[2026] OpenTelemetry for LLM Observability — Self-Hosted Setup

The article shows how to self-host OpenTelemetry for LLM observability with a collector pipeline, trace storage in Tempo or Jaeger, and metrics in Prometheus.

It focuses on tracing LLM agent steps and standardizing generative AI telemetry with OTLP and the GenAI semantic conventions.

Python instrumentation examples cover LLM calls across providers such as OpenAI, Anthropic, and vLLM.

The approach gives small or self-hosted teams vendor-neutral visibility into token usage, trace continuity, and cost control without managed platforms.