How does Spring AI handle observability and what metrics does it expose?

Spring AI integrates with Spring Boot's Micrometer-based observability stack out of the box. When spring-ai-*-spring-boot-starter is on the classpath alongside spring-boot-starter-actuator and a Micrometer registry (Prometheus, OpenTelemetry, Zipkin, etc.), Spring AI auto-configures instrumentation for every AI model call.

What Spring AI instruments by default:

• spring.ai.chat.client — a timer and counter around ChatClient calls, tagged with model name, operation type, and provider.
• spring.ai.chat.model — metrics at the ChatModel level with latency histograms.
• Token usage — counters for input.tokens, output.tokens, and total.tokens extracted from the provider response metadata. Critical for cost tracking.
• Distributed traces — each AI call creates a span with prompt content (configurable), model name, and token counts as attributes.

# application.properties — enable full prompt content in traces (use carefully — PII risk)
spring.ai.chat.client.observations.include-prompt=true
spring.ai.chat.model.observations.include-completion=false

# Enable AI metrics endpoint
management.endpoints.web.exposure.include=metrics,prometheus

Token usage metrics are especially valuable in production because they directly correlate to cost. Setting up a Grafana dashboard on spring.ai.chat.model.input.tokens per service lets you attribute spend to specific features and spot runaway prompt sizes before they cause invoice surprises.