Generative AI systems introduce new observability needs — context propagation across prompts, model-level metrics, prompt-source attribution, and provider-specific behavior — yet existing telemetry schemas were not designed for those semantics. Standardizing GenAI telemetry lets operators compare providers, detect prompt-level regressions, and instrument end-to-end pipelines without bespoke conventions per model or vendor.
What Sets It Apart
- Focused on GenAI semantics: defines spans, metrics, and events tailored to prompt/response flows, model context (MCP), and multi-provider scenarios so instrumentation can capture prompt inputs, transformations, and model outputs consistently.
- Machine- and human-friendly artifacts: core definitions live as YAML model files (for tooling and codegen) while Markdown docs provide readable guidance, enabling both automatic validation and direct developer consumption.
- Reference implementations and compliance matrix: includes language-specific reference pieces and a compliance matrix to show per-language/signal support, reducing integration ambiguity for operators and SDK authors.
Who It's For and Trade-offs
Great fit if you maintain or instrument GenAI services, build SDKs/agents that call multiple providers, or operate observability stacks that must surface prompt- and model-level diagnostics. It lowers integration cost by providing a shared schema and reference tooling. Look elsewhere if you need application-level business metrics only (this focuses on telemetry semantics, not higher-level analytics) or if you require turnkey APM features; this project defines conventions and references rather than providing a hosted observability product.
Where It Fits
Best used as the canonical schema layer between GenAI clients/agents and existing OpenTelemetry pipelines (collectors, exporters, APMs). It complements provider SDKs by offering a neutral mapping for attributes and events, making cross-provider comparisons and incident investigations more feasible.
