OpenTelemetry GenAI Semantic Conventions
OpenTelemetry GenAI Semantic Conventions is the OTel repo for GenAI and MCP telemetry conventions (spans, metrics, events). We assess it for standardizing agent and LLM observability before vendor schemas lock in.
Blurb
Semantic Conventions for Generative AI (GenAI), including spans, metrics, and events for GenAI clients, MCP (Model Context Protocol), and provider-specific conventions (OpenAI, etc.).
Summary
When to use: instrumenting LLM clients, agents, embeddings, retrieval, or Model Context Protocol servers/clients with OpenTelemetry. Prefer these attribute names when backends or SDKs already expose gen_ai.* / mcp.*.
When to skip: pure classic service telemetry with no GenAI path. Stay on core semantic-conventions until you emit model or agent signals.
Trade-offs: docs mark GenAI conventions Development, so names and required fields can still change. The GenAI model lives in this repo (Weaver-managed) and links a pinned core semconv version for shared attributes.
Pairs with: OTel SDKs and Collector pipelines; provider-specific docs (OpenAI, Anthropic, Bedrock, Azure AI Inference); MCP host observability.
Details
- Signals: GenAI events, exceptions, metrics, model spans, and agent spans; plus MCP-specific conventions.
- Providers: technology-specific pages for Anthropic, Azure AI Inference, AWS Bedrock, and OpenAI.
- Layout: human docs under
docs/; YAML model undermodel/; reference implementations underreference/. - Upstream: extends core OpenTelemetry semantic conventions; not a replacement for the main semconv repo.
- Status: document status is Development as of note creation (2026-07-29). Re-check before treating attribute sets as stable contracts.