kubit-otel 0.6.2

OpenTelemetry exporter for Kubit analytics.

kubit-otel

OpenTelemetry exporter for Kubit analytics.

Install

pip install kubit-otel

Quick start

from kubit_otel import configure
from opentelemetry import trace

configure(api_key="rg.v1.xxx", service_name="my-app")
tracer = trace.get_tracer("my-app")

with tracer.start_as_current_span("chat.completion") as span:
    span.set_attribute("gen_ai.request.model", "gpt-4o")
    span.set_attribute("gen_ai.prompt", "Hello, world!")
    span.set_attribute("gen_ai.completion", "Hi there!")
    span.set_attribute("gen_ai.usage.input_tokens", 10)
    span.set_attribute("gen_ai.usage.output_tokens", 5)

Spans are exported to Kubit using standard OpenTelemetry GenAI semantic conventions. Each trace's root span is recorded as the trace; every span (including the root) is recorded as an observation under it. gen_ai.* attributes are mapped to first-class, queryable fields for model name, prompt/completion, token counts, and cost.

Works alongside other OTel-based SDKs

configure() detects whether a real TracerProvider is already installed as the global OTel provider. If so, it attaches KubitSpanProcessor to that provider and merges in your resource attributes — it does not replace the existing provider. You can call configure() before or after other OTel-based libraries (Langfuse, OpenLLMetry, an OTel distro, …) and every span will reach both sinks.

If you want explicit "attach only, never register" behavior, use attach():

from kubit_otel import attach

# Must be called after another library has installed a real provider.
attach(api_key="rg.v1.xxx")

Supported attributes

OpenTelemetry attribute	Purpose
gen_ai.request.model / gen_ai.response.model	Model name
gen_ai.prompt / gen_ai.content.prompt	Input prompt
gen_ai.completion / gen_ai.content.completion	Output completion
gen_ai.usage.input_tokens	Input token count
gen_ai.usage.output_tokens	Output token count
gen_ai.usage.cost	Total cost (USD)
session.id	Conversation session id
enduser.id	End-user id

Span filtering

By default, only LLM-relevant spans are forwarded to Kubit. A span is exported if it:

• was created by the Kubit SDK tracer (kubit-sdk),
• carries any gen_ai.* semantic-convention attribute, or
• comes from a known LLM instrumentation scope (OpenInference, Langfuse, Vercel AI SDK, …).

This keeps HTTP/DB/framework auto-instrumentation noise out of your Kubit workspace without extra configuration.

Extend the default filter

from kubit_otel import configure, is_default_export_span

configure(
    api_key="rg.v1.xxx",
    should_export_span=lambda span: (
        is_default_export_span(span)
        or (
            span.instrumentation_scope is not None
            and span.instrumentation_scope.name.startswith("my_framework")
        )
    ),
)

Full override

configure(
    api_key="rg.v1.xxx",
    should_export_span=lambda span: span.name.startswith("llm."),
)

Export everything

configure(api_key="rg.v1.xxx", should_export_span=lambda _span: True)

Python compatibility

Python 3.9+

License

Proprietary — see LICENSE.