Document LLMObs trace sampling controls for Python and Node.js SDKs

[Yun-Kim]

What

Documents Python and Node.js LLM Observability (Agent Observability) trace sampling controls in the SDK reference (instrumentation/sdk.md):

Python:

• DD_LLMOBS_SAMPLE_RATE environment variable — added to the Python command-line setup env-var list (float 0.0–1.0, default 1.0).
• sample_rate parameter on LLMObs.enable() — added to the Python in-code setup parameters, following dd-trace-py#18607. Documents precedence (sample_rate arg > DD_LLMOBS_SAMPLE_RATE > default 1.0) and that out-of-range values are ignored.

Node.js (assumes dd-trace-js#9030 merges):

• DD_LLMOBS_SAMPLE_RATE environment variable — added to the Node.js command-line setup env-var list (float, default 1.0).
• sampleRate parameter under llmobs in tracer.init({ llmobs: { sampleRate } }) — added to the Node.js in-code setup parameters. Documents precedence (sampleRate > DD_LLMOBS_SAMPLE_RATE > default 1).

In the Node.js SDK the keep/drop decision is computed once on the root span, inherited by descendants, and propagated across services via x-datadog-tags; spans are always shipped and the decision is applied during ingestion.

Why

Neither control was documented anywhere in content/en/llm_observability/. The existing docs only cover server-side/UI sampling (automation rules, annotation queues, datasets, custom judges), so there was no reference for client-side, pre-ingestion trace sampling in the SDK.

Scope notes

• Python and Node.js. Java does not expose these controls, so its tabs are intentionally untouched.
• No retention/limits content added — this PR is scoped to sampling mechanisms only.
• Cross-links to automation rules and APM ingestion mechanisms; both targets verified to exist.

Claude session: `8bc9e12c-8440-4689-a2dc-e2f1f1ad66ae`
Resume: `claude --resume 8bc9e12c-8440-4689-a2dc-e2f1f1ad66ae`

🤖 Generated with Claude Code

[github-actions]

Preview links (active after the build_preview check completes)

Modified Files

• https://docs-staging.datadoghq.com/yun.kim/llmobs-sampling-docs/llm_observability/instrumentation/sdk

[Yun-Kim]

Suggested change

<br />The fraction of traces to submit to Agent Observability, between `0.0` (drop everything) and `1.0` (submit everything). The sampling decision is made on the root span and inherited by all of its child spans, so a trace is always kept or dropped as a whole. The decision is also propagated to downstream services, keeping distributed traces intact. This client-side sampling happens before ingestion and is independent of in-app controls such as [automation rules][2]. It does not affect [APM trace sampling][3], and APM sampling does not drop Agent Observability data.

<br />The fraction of traces to submit to Agent Observability, between `0.0` (drop everything) and `1.0` (submit everything). The sampling decision is made on the root span and inherited by all of its child spans (including downstream services). This sampling is independent of in-app controls such as [automation rules][2] and [APM trace sampling][3].

[Yun-Kim]

Same suggestion as above (python)

[Yun-Kim]

Suggested change

<br />The fraction of traces to submit to Agent Observability, between `0.0` (drop everything) and `1.0` (submit everything). The sampling decision is made on the root span and inherited by all of its child spans, so a trace is always kept or dropped as a whole. Values outside the `[0.0, 1.0]` range are ignored. This takes precedence over `DD_LLMOBS_SAMPLE_RATE`; if not provided, this defaults to the value of that environment variable (`1.0` if unset).

<br />The fraction of traces to submit to Agent Observability, between `0.0` (drop everything) and `1.0` (submit everything). The sampling decision is made on the root span and inherited by all of its child spans (including downstream services). If unset, this defaults to `DD_LLMOBS_SAMPLE_RATE`, but otherwise takes precedence over the environment variable.

[Yun-Kim]

Same comment for Python inline config