groundcover built a GenAI OTel normalizer across SDK x Framework x Provider dimensions and found the “standard” is a maze of naming conflicts and structural mismatches.
Key Takeaways
Four wire formats cover all incoming GenAI spans: Traceloop/LangChain (A), LangSmith (B), canonical gen_ai.* (C), and future OTel events (D, partial).
Six different attribute keys can hold the model name; token counts have five naming conventions across SDKs requiring a priority resolution chain.
Anthropic and Bedrock count cache tokens additively; OpenAI does not – generic OTel backends silently undercount costs without provider-specific arithmetic.
Provider name normalization requires a 26-entry map to collapse case variants, abbreviations, and gateway prefixes into 15 canonical strings.
Framework axis (LangGraph, CrewAI, Pydantic AI) affects span tree shape and message serialization even when SDK and provider are identical.
