OpenInference Instrumentation for TypeSafe AI SDK

Run npm install --save @arizeai/openinference-instrumentation-typesafe @typesafe-ai/sdk. The package requires Node.js 20 or newer and is used alongside an OpenTelemetry Node trace provider such as @opentelemetry/sdk-trace-node.

Each TypeSafeClient.systemOne call emits a single OpenInference DECISION span containing JSON input.value and output.value, the model identity via decision.system, decision.provider, and decision.model_name, plus decision.token_count.input and decision.token_count.output.

The instrumentation follows OpenInference decision span conventions. System One is a decision model that scores or selects among candidate options in Noul, Choice, and Score questions rather than generating free-form text, so decision model usage is not counted as LLM usage.

The TypeSafeInstrumentation constructor accepts a traceConfig option with OpenInference masking settings such as hideInputs and hideOutputs, along with a standard instrumentationConfig and an optional tracerProvider that defaults to the global provider.

The package is compatible with @typesafe-ai/sdk versions >=0.6.0 when using instrumentation version ^0.1.0. It is published by Arize under the Apache-2.0 license.

Register the instrumentation before loading the SDK in CommonJS. For ESM, bundlers, or cases where the SDK is imported first, create a TypeSafeInstrumentation instance and call manuallyInstrument(TypeSafe) on the imported module instead.

Traces follow OpenInference semantic conventions and can be sent to any OpenTelemetry-compatible backend, including Arize Phoenix and Arize AX. A custom tracer provider can be passed to the constructor or set later via setTracerProvider.