Telemetry (OpenTelemetry)

Requires the otel extra: pip install "byoai-runtime[otel]". Nothing in byoai.telemetry.otel is imported unless telemetry is actually enabled, so the runtime has zero overhead without it.

Zero-SaaS: traces export via OTLP to a collector you already run (Grafana Tempo, Datadog agent, Honeycomb, Jaeger, ...) — nothing is sent to a third-party ByoAI service.

One span covers each execution (byoai.execute), with one child span per pipeline stage and span events for provider lifecycle (started/completed/failed — retries and fallbacks show up as multiple event pairs). Usage and cost land as span attributes, following the OTel GenAI semantic conventions where they apply.

Declarative setup

runtime = Runtime(
    llm={"provider": "openai", "model": "gpt-4o"},
    telemetry={
        "endpoint": "http://otel-collector.internal:4317",
        "service_name": "my-app",
    },
)

Omit endpoint to attach to the process's globally configured OpenTelemetry SDK instead of creating a new OTLP exporter:

runtime = Runtime(llm={"provider": "openai", "model": "gpt-4o"}, telemetry={})

protocol is "grpc" (default, port 4317) or "http"/"http/protobuf" (port 4318) — useful when a corporate ingress only allows the latter. compression is "gzip" or None. resource_attributes adds to (not replaces) service.name, e.g. {"service.version": "1.2.0", "deployment.environment": "prod"}. Runtime.close() shuts down a tracer provider it created for you; a tracer_provider you pass in is yours to manage.

Manual setup

from byoai.telemetry.otel import instrument

instrument(runtime)                          # uses the globally configured SDK
instrument(runtime, tracer_provider=my_tp)   # or an explicit provider

telemetry= also accepts an already-built TracerProvider directly (instead of a config dict), mirroring how cache= and vector_store= accept pre-built adapter instances.