OpenTelemetry Gen AI Semantic Conventions Support

AI SDK
1

Issue

The Vercel AI SDK currently uses custom attributes under the ai.* namespace for telemetry data, which differs from the OpenTelemetry Gen AI Semantic Conventions.

Current implementation:

{
  "ai.prompt.messages": "[{\"role\":\"user\",\"content\":\"...\"}]",
  "ai.response.text": "...",
  "ai.prompt.tools": [...]
}

OTel standard:

{
  "gen_ai.input.messages": [{...}],
  "gen_ai.output.messages": [{...}],
  "gen_ai.tool.definitions": [...]
}

Impact

  • Third-party observability tools (e.g., Phoenix, Langfuse) cannot auto-extract LLM I/O without custom adapters
  • Messages are stored as JSON strings instead of structured objects
  • Span kind is UNKNOWN instead of semantic types like CLIENT

Question

Is there a plan to adopt the OpenTelemetry Gen AI Semantic Conventions? This would improve interoperability across the LLM observability ecosystem.

References