Langfuse: OpenTelemetry tracing for TypeSafe Jev

It traces TypeSafe Jev System One calls with Langfuse. The OpenInference instrumentor patches TypeSafeClient.system_one in place, so you keep a plain client and still get OpenTelemetry spans for skill routers, model routers, tool-call hooks and eval verdicts.

No. The notebook states that the official Python and JavaScript SDKs wrap POST /v1/systemone but do not emit OpenTelemetry spans themselves, and that official TypeSafe docs do not mention OpenTelemetry.

Install langfuse, typesafe-sdk and openinference-instrumentation-typesafe, set LANGFUSE_PUBLIC_KEY, LANGFUSE_SECRET_KEY, LANGFUSE_BASE_URL and TYPESAFE_API_KEY, then call TypeSafeAIInstrumentor().instrument() once. No observeTypeSafe() helper or Langfuse-specific client wrapper is needed.

input.value with the request body, state, model and questions as JSON; output.value with the response body, model, answers and usage as JSON; request and response model names such as jev-latest to jev-1.13.0; and prompt, completion and total token counts.

A System One call is not a chat exchange, so the spans do not set llm.input_messages or llm.output_messages. The instrumentation records the request and response payloads instead, and Langfuse's Python SDK is OpenTelemetry-native, so the instrumentor attaches to the same tracer provider that get_client() registered.

The notebook says to pin jev-1.13.0 when a threshold depends on a specific model version, because jev-latest moves when TypeSafe ships a new release. The example call uses model jev-1.13.0 and asks a Noul, a Choice and a Score about one ticket.

Enable debug mode with LANGFUSE_DEBUG=True. If OTel observations appear in the logs but not in Langfuse, call langfuse.flush() at the end and verify the API keys and base URL. If no OTel spans appear at all, make sure the instrumentation runs before your application code.