Skip to main content
Every attribute Elementum sets on the AI spans described in GenAI tracing. Keys are listed exactly as they arrive. Most follow the OpenTelemetry GenAI semantic conventions (gen_ai.*). A few duplicate the same value under OpenInference names (openinference.*, llm.*, input.*, output.*), which some evaluation platforms read instead.

Conventions

  • Type is the OTLP value type: string, int (64-bit integer), double, boolean, or string[].
  • Spans: A = invoke_agent, L = chat (LLM call), T = execute_tool, V = voice spans.
  • An attribute with no value is left out. It is never sent empty. “Absent when” lists the usual reasons.
  • Example ids are placeholders.

Resource and span properties

Operation

Model and provider

Response and usage

Content

Both values are JSON arrays of messages, sent as text. One message per object: Example gen_ai.output.messages for a reply that calls a tool:
On LLM call spans, the tool call keys are the flat strings function.name and function.arguments. In voice transcripts, a tool call is an object with id, type set to function, and a nested function object that has name and arguments. If Elementum cannot write the messages as JSON, the value is serialization_error.

Agent and conversation

Tool

When one execute_tool span covers several tool calls, these attributes describe one of them. See Span tree.

Elementum attributes

Voice attributes

Attributes per span

Keys Elementum does not send

Some evaluation tools and older guides mention these keys. Elementum does not send them. Organization and environment ids are never attributes. To label spans, add resource attributes to the endpoint. Next: Evaluating AI with traces.