LLM telemetry reference

This page describes what data is provided by LLM observability and how it is used.

Span dataset schema

Spans are ingested to the Tracing/LLM Span dataset.

The LLM Span dataset incorporates additional fields over the Tracing/Span dataset.

LLM Span dataset field

Underlying OpenTelemetry field from which it is derived

llm_span_type

Attributes traceloop.span.kind and llm.request.type

request_model

Attribute llm.request.model

response_model

Attribute llm.response.model

The llm_span_type field provides an intuitive categorization of the LLM span. Depending on your application, you may see a mix among the following types:

LLM Span Types

agent

chat

tool

workflow

task

embeddings

completion

rerank

For more information about other fields, correlation tags, and other OpenTelemetry span related datasets, read our APM reference documentation

LLM metrics

The following metrics are derived from spans and are available in the Metric Explorer, and can be used for dashboarding and monitors:

  • gen_ai_span_call_count

  • gen_ai_span_error_count

  • gen_ai_span_duration

  • gen_ai_span_duration_lg

Metric dimensions (tags):

  • environment

  • service_name

  • service_namespace

  • service_version

  • llm_span_type

  • is_llm_request

  • provider

  • request_model

  • response_model

  • status_code

  • status_message