AI Telemetry標準編

Enterprise AIの本番運用では、単一providerのログ画面だけではmodel call、retrieval、tool、agent orchestration、application/backendを一つの因果系列として追跡できない。更新版では「OpenTelemetryで全部標準化済み」と単純化せず、W3C Trace Context / OpenTelemetry core / GenAI semantic conventions / OpenInference等の役割と成熟度を分離し、移行期の互換性・privacy・schema driftまで扱う必要がある。

Enterprise AI Core / interoperability / observability / implementation / governance / security-privacy / provider mapping input

AI telemetryをprovider横断で構成する際、2026-09-14時点で何が安定した共通基盤で、何が発展途上のAI固有semantic layerか。OpenTelemetry GenAIとOpenInferenceはどこまで共通化でき、どこをschema/version gateとして扱うべきか。

• source class: standard exact title: Trace Context author/org: W3C canonical URL: https://www.w3.org/TR/trace-context/ publication/update/access: W3C Recommendation 2020-02-06; accessed 2026-09-14 version/revision: Recommendation with editorial updates locator: traceparent / tracestate definitions; distributed trace context propagation evidence label: VERIFIED_SPEC relation: support/context

• source class: official specification exact title: OpenTelemetry semantic conventions 1.44.0 author/org: OpenTelemetry canonical URL: https://opentelemetry.io/docs/specs/semconv/ publication/update/access: accessed 2026-09-14 version/revision: SemConv 1.44.0 page; GenAI section points to dedicated repository locator: “Generative AI: Moved to the OpenTelemetry GenAI semantic conventions repository.” evidence label: VERIFIED_SPEC / PRODUCT_VOLATILE relation: support/context

• source class: author-maintained repository/spec exact title: OpenTelemetry GenAI Semantic Conventions author/org: OpenTelemetry canonical URL: https://github.com/open-telemetry/semantic-conventions-genai publication/update/access: accessed 2026-09-14 version/revision: repository main; schema URL shown as https://opentelemetry.io/schemas/gen-ai/1.42.0 locator: README; GenAI-specific spans, metrics, events, MCP/provider conventions evidence label: VERIFIED_SPEC / OSS_PRACTICE / PRODUCT_VOLATILE relation: support/context

• source class: author-maintained repository release state exact title: Releases · open-telemetry/semantic-conventions-genai author/org: OpenTelemetry canonical URL: https://github.com/open-telemetry/semantic-conventions-genai/releases publication/update/access: accessed 2026-09-14 version/revision: no GitHub releases visible as of access date locator: releases page — “There aren’t any releases here” evidence label: OBSERVED / PRODUCT_VOLATILE relation: limit/counter

• source class: official specification source exact title: Semantic conventions for generative client AI spans author/org: OpenTelemetry canonical URL: https://github.com/open-telemetry/semantic-conventions-genai/blob/main/docs/gen-ai/gen-ai-spans.md publication/update/access: accessed 2026-09-14 version/revision: repository main locator: page status “Development”; inference / embeddings / retrievals / fetch response / memory / execute tool spans evidence label: VERIFIED_SPEC / PRODUCT_VOLATILE relation: support/limit

• source class: official specification source exact title: Semantic conventions for GenAI agent calls author/org: OpenTelemetry canonical URL: https://github.com/open-telemetry/semantic-conventions-genai/blob/main/docs/gen-ai/gen-ai-agent-spans.md publication/update/access: accessed 2026-09-14 version/revision: repository main locator: create agent / invoke workflow / plan span definitions and requirement language evidence label: VERIFIED_SPEC / PRODUCT_VOLATILE relation: support/context

• source class: author-maintained specification exact title: OpenInference Specification author/org: Arize AI / OpenInference canonical URL: https://arize-ai.github.io/openinference/spec/ publication/update/access: accessed 2026-09-14 version/revision: current docs locator: OpenInference described as AI observability semantic conventions built on OpenTelemetry; LLM, agent, tool, retrieval, evaluation and privacy requirements evidence label: VERIFIED_SPEC / OSS_PRACTICE relation: support/context

• source class: author-maintained specification exact title: Semantic Conventions | OpenInference author/org: Arize AI / OpenInference canonical URL: https://arize-ai.github.io/openinference/spec/semantic_conventions.html publication/update/access: accessed 2026-09-14 version/revision: current docs locator: span kinds LLM / EMBEDDING / CHAIN / RETRIEVER / RERANKER / TOOL / AGENT / GUARDRAIL / EVALUATOR / PROMPT; trace/session evaluations evidence label: VERIFIED_SPEC / OSS_PRACTICE relation: support

• source class: author-maintained specification exact title: OpenInference Configuration author/org: Arize AI / OpenInference canonical URL: https://arize-ai.github.io/openinference/spec/configuration.html publication/update/access: accessed 2026-09-14 version/revision: current docs locator: OPENINFERENCE_HIDE_* controls and __REDACTED__ behavior; defaults documented as False for multiple capture-hiding switches evidence label: VERIFIED_SPEC / OSS_PRACTICE relation: support/limit/security

• source class: author issue / failure-gap report exact title: Generic Grouping Attributes for Agentic Workflow Spans #94 author/org: OpenTelemetry semantic-conventions-genai contributors canonical URL: https://github.com/open-telemetry/semantic-conventions-genai/issues/94 publication/update/access: opened 2026-04-25; accessed 2026-09-14 version/revision: issue #94, open at indexed source locator: lack of standardized grouping for iterations/tasks/skills/phases in agentic workflows evidence label: SOURCE_REPORT / OSS_PRACTICE / PRODUCT_VOLATILE relation: counter/limit

• source class: author issue / implementation-gap report exact title: Add metrics support to Anthropic & Groq GenAI instrumentations #98 author/org: OpenTelemetry opentelemetry-python-genai contributors canonical URL: https://github.com/open-telemetry/opentelemetry-python-genai/issues/98 publication/update/access: opened 2026-05-14; accessed 2026-09-14 version/revision: issue #98 locator: OpenAI v2 emits spans+metrics while cited Anthropic/Groq instrumentations emit spans without metrics evidence label: SOURCE_REPORT / OSS_PRACTICE / PRODUCT_VOLATILE relation: counter/limit

• source class: author issue / schema-operability gap exact title: Publish Semantic Conventions in a Machine-Readable Format #3097 author/org: Arize-ai/openinference contributors canonical URL: https://github.com/Arize-ai/openinference/issues/3097 publication/update/access: opened 2026-05-13; accessed 2026-09-14 version/revision: issue #3097 locator: conventions defined in Markdown and manually re-declared across language packages; request for canonical machine-readable representation evidence label: SOURCE_REPORT / OSS_PRACTICE / PRODUCT_VOLATILE relation: counter/limit

• Stable common substrate and AI semantic layer must be separated. W3C Trace Context provides a mature cross-service propagation mechanism; OpenTelemetry provides transport/data-model/collector ecosystem. AI-specific semantics are moving faster and should be version-gated independently.
• OpenTelemetry GenAI semantic conventions were split from the core semantic-conventions repository in 2026. The dedicated repository exposes a GenAI schema URL but, at the inspected releases page, no formal GitHub release exists. Therefore future-book wording must not imply a settled GA-level AI telemetry standard.
• Current OpenTelemetry GenAI material covers not only inference but retrieval, memory, tool execution, agent/workflow/planning operations. This is sufficient to define a future Core trace topology, while individual attribute stability still requires per-version confirmation.
• OpenInference is not an alternative transport to OpenTelemetry; it is an AI-specific semantic convention layer designed to ride on OpenTelemetry/OTLP. It already provides a broad span-kind taxonomy and evaluation/annotation concepts useful for production AI operations.
• OpenTelemetry GenAI and OpenInference are overlapping but not identical schemas. Treating them as wire-compatible synonyms would be unsafe. An update should maintain an explicit mapping table and record the originating schema/version on telemetry.
• Instrumentation maturity differs by provider/language. The cited OpenTelemetry Python GenAI issue demonstrates that even within the same project some integrations may emit different signal sets. “Uses OTel” does not prove equivalent observability coverage.
• Agent hierarchy remains an active design area. Issue #94 records a missing generic grouping primitive for iterations/tasks/skills/phases, which limits generic reconstruction of complex agent trajectories.
• Sensitive prompt/tool/output capture is a governance control, not a harmless debugging default. OpenInference exposes explicit masking knobs; the update should require capture policy, redaction policy and retention policy before enabling content-rich traces.

• Repository main pages are moving targets; exact attribute names/status can change without product deployment changes.
• “No GitHub releases” is an as-of observation, not evidence that the project is unusable or unversioned in every distribution channel.
• OpenInference author docs establish its own intended semantics, not universal industry adoption.
• Issue reports establish concrete gaps in cited implementations/revisions, not defects in every language/provider/version.
• Trace completeness can fail at async boundaries, browser/desktop automation boundaries, external SaaS calls or components that do not propagate context. A trace backend cannot reconstruct events never instrumented.
• Capturing full model inputs/outputs can create PII, confidential-data, prompt-secret and retention exposure. Redaction after export may be too late.
• Sampling can distort rare-failure analysis and token/cost aggregation unless sampling semantics are accounted for.

• Language SDK support and auto-instrumentation coverage differ by runtime.
• OpenTelemetry GenAI attribute/status/release availability must be rechecked at update publication time.
• Provider SDK versions determine which token/cost/tool attributes are observable.
• Collector/backend must preserve required attributes and W3C trace context across service boundaries.
• Organization policy may prohibit payload-level prompt/output telemetry even where SDK capture is technically available.

Status: implementable now with maturity caveats. Recommended architecture: W3C Trace Context + OpenTelemetry SDK/OTLP/Collector as substrate; choose one AI semantic profile per instrumentation boundary; persist schema/profile version; map alternate schemas at collector/backend rather than silently mixing attribute names. Instrument at least model, retrieval, tool, agent/workflow and application/backend boundaries. Add evaluation signal separately from raw execution telemetry.

No benchmark score proves telemetry completeness. Validation should instead use a synthetic trace conformance test: execute a known model→retrieval→tool→model workflow; verify parent/child structure, context propagation, errors, retries, token metadata, evaluation attachment, redaction and sampling behavior. Repeat after instrumentation/provider SDK upgrades.

Security/privacy: classify prompts, completions, tool arguments/results and retrieved documents before capture; default to metadata-only where sensitivity is unknown. Redact before export when possible. Audit: preserve trace_id, span_id, schema/profile version, instrumentation library version, model/provider identifiers and evaluation provenance. Cost: token/cost attributes are useful but provider instrumentation coverage is uneven; billing records remain the financial source of truth unless independently reconciled. Operability: deploy schema-change detection and dashboards that tolerate missing optional attributes rather than assuming every provider emits identical fields.

“OpenTelemetry and W3C Trace Context provide a strong interoperability substrate for AI telemetry, but AI-specific semantic conventions remain an evolving layer. As of 2026-09-14, OpenTelemetry’s GenAI conventions have moved to a dedicated repository and several agentic concepts remain under active development. OpenInference offers a complementary OpenTelemetry-based AI schema with broad model/tool/retrieval/agent/evaluation coverage. Enterprise deployments should pin and record schema versions, validate provider/language instrumentation coverage, and treat prompt/output capture as governed sensitive telemetry.”

• At update time, recheck OpenTelemetry GenAI repository release/tag/schema status and per-section stability.
• Build a mapping appendix: OTel GenAI ↔ OpenInference for model, retrieval, tool, agent/workflow, memory, evaluation and cost/token fields.
• Add a provider/language coverage matrix based on concrete instrumentation releases, not project-level capability claims.
• Define a minimal enterprise trace profile and synthetic conformance TEST_PROFILE.
• Connect this packet directly to EA-052 AgentOps/AI-SRE and later provider mappings.

first formal release/versioning policy of semantic-conventions-genai; convergence/divergence trajectory with OpenInference; provider-specific auto-instrumentation parity; stable representation for agent iteration/task grouping.

bounded initial packet complete. Reopen when dedicated OTel GenAI formal release/stability changes, mapping work begins, or a provider mapping requires exact attribute-level verification.

---

W3C Trace Context/OpenTelemetry core substrate must remain separate from evolving AI-specific semantic layers. OpenTelemetry semantic-conventions v1.44.0 moved GenAI content to the dedicated semantic-conventions-genai repository; GenAI conventions remain actively changing/development and require schema/version pinning. OpenTelemetry GenAI and OpenInference are overlapping but not interchangeable schemas. Preserve provider/language instrumentation gaps and all OPEN_UNKNOWN.