Translates AgentObs event metadata to OpenInference semantic conventions.

This module is responsible for converting the backend-agnostic AgentObs event metadata into the flattened, indexed attribute format required by OpenInference.

OpenInference is an extension of OpenTelemetry semantic conventions specifically designed for AI and LLM observability. It enables rich, contextualized visualization in platforms like Arize Phoenix.

OpenInference Span Kinds

• AGENT - Agent loop or orchestration
• LLM - Large Language Model API call
• TOOL - Tool or function execution
• CHAIN - Sequence of operations (e.g., orchestrator iteration)
• RETRIEVER - Vector/document retrieval (not used in AgentObs currently)

Key Attributes

The translator produces flattened attributes following the OpenInference spec:

• openinference.span.kind - The span kind (AGENT, LLM, TOOL, etc.)
• input.value / output.value - Primary input/output
• llm.model_name - Model identifier
• llm.input_messages.N.message.role - Message role (user, assistant, etc.)
• llm.input_messages.N.message.content - Message content
• llm.token_count.prompt / llm.token_count.completion - Token usage
• tool.name / tool.description - Tool metadata

References

• OpenInference Spec: https://arize-ai.github.io/openinference/spec/semantic_conventions.html
• Arize Phoenix Docs: https://arize.com/docs/phoenix/