Tracing

Complete visibility into how your agents and models work

What is Tracing?

Tracing captures the full journey of a request as it moves through parts of a system.

A trace is made up of a series of spans — each span corresponds to an individual unit of work such as invoking a model, retrieving data, or running a function. Tracing shows how the request propagated, what components it touched, how long each part took, and what inputs and outputs were involved.

Why is Tracing Important?

Tracing helps you spot where things take too long, where errors are happening, and which parts of the agent/application need improvement. By capturing detailed metadata such as latency, token counts, and cost, tracing helps you improve system efficiency.

Why Arize AX for Tracing?

• Arize created the industry leading OpenInference Semantic Conventions that is widely adopted by model providers, frameworks, and even observability platforms. Traces work best in the platform in this format.

• Native instrumentation for over 30+ model providers, frameworks, and gateways. Because we own the instrumentation, we can extend coverage, adapt quickly, and maintain compatibility across frameworks.

• Scales to over terabytes of data and billions of spans

• Native cost tracking support and can be customized

• Agent Visualizations for aggregate analysis

Getting Started

There's 2 approaches to getting traces in platform:

1. Integrations - Use a supported integration

2. Manual Instrumentation - Fully customize your traces

How Does Tracing Work?

Instrumentation enables your application to emit trace data by automatically or manually wrapping function calls and capturing attributes like inputs, outputs and latency. An exporter transmits the captured spans in the OpenInference format to Arize using the OpenTelemetry Protocol (OTLP) over gRPC. Once received, the Arize collector ingests and visualizes the spans in the Arize UI.