Instrumenting Pydantic AI

Pydantic AI emits OpenTelemetry spans for each run, each model request and each tool call, following the Semantic Conventions for Generative AI systems. Turning that on is what lets someone see the prompts a model actually received, what the tools returned, how long a run took and what it cost.

Two things share the Logfire name, and the difference decides most of what follows:

• Pydantic Logfire is Pydantic's own OpenTelemetry backend, and an observability and evaluation platform for AI applications built on top of it. Signing up is free and takes a GitHub login or an email address — no credit card, no sales step.
• The Logfire SDK is a Python OpenTelemetry SDK. It sends to Pydantic Logfire by default, and can be pointed at any other OpenTelemetry backend instead.

Almost everything below is the SDK. Which backend it sends to is Step 1.