LLM Observability
Domain: Models & Providers
Motivation
LLM observability is exposing model-call data for monitoring and telemetry. dsh surfaces model usage through the token-meter session projections and product telemetry, and OTel packages emit traces for the model path. Usage folds replace samples within each attempt and feed the tokenUsage and contextPressure projection units.
Concrete Example
The contextPressure projection carries pressureTokens, projectedTokens, and contextWindow from the newest request/context record for downstream telemetry.
Analogy
The dashboard of gauges — fuel, range, and load — for the model’s trip.