Concepts

How Spanlens models LLM observability: the signals worth capturing, the entities they map to, and how traces, evals, and prompt versions relate.

LLM observability

What LLM observability is, the five signal categories worth capturing, and how Spanlens maps each one to a concrete entity in the data model.

Agent tracing

How Spanlens models agent traces: trace root, agent steps, LLM calls, tool calls, parent/child links, and critical path computation.

Evals

How Spanlens models evals: LLM-as-judge scoring, human annotation, judge-to-human correlation as a metric, and drift detection across prompt versions.

Prompt management

How Spanlens versions prompts, runs Prompt A/B with Welch t-test on latency and cost, computes z-test on error rate, and rolls back without deploys.

Data model

Spanlens data model in one page: Request, Trace, Span, Prompt Version, Eval, Dataset, and how they relate for billing, debugging, and quality questions.

Back to the docs overview.