title: Overview description: The Observability landing dashboard - eight KPI cards, fifteen charts across six sections, and a live activity grid summarising every other tab's headline number.

Overview

Overview after a brief chat session - one card or chart per concern, with every section linking conceptually to the dedicated tab that owns it. The eight KPI cards at the top stay visible regardless of activity; the charts below fill in as data accumulates.

Route - sidebar → top item → observability?tab=overview

Purpose - single-page summary of every other tab's headline number. Open it first when you don't yet know what you're looking for.

When to look here

• "Has anything changed in the last 30 minutes?" - every KPI has a sparkline showing the trend across the window.
• "Where do I drill in?" - every section header links conceptually to a dedicated tab (Tokens & Cost, MCP Servers, etc.). Spot the anomaly here, then jump.
• "What's the system doing right now?" - heap, CPU, active LLM ops gauges live on this page.
• "Last 10 traces" - the activity grid at the bottom shows the most recent traces; click any row to open Trace Detail.

Controls

All dashboards share the Observability global settings - time window (LAST_5M · LAST_10M · LAST_20M · LAST_30M · LAST_1H · LAST_3H, default LAST_30M), refresh interval, custom From/To range. Sparklines and charts retick to the selected window. Overview adds no tab-specific controls.

KPI cards (eight)

Each headline metric carries a 60-sample sparkline showing the trend:

Card	Shows	Source
Calls	Total chat turns in the window	TraceRecord count
Tokens	Sum of input + output tokens	TraceRecord.totalTokens (falls back to inputTokens + outputTokens when total is null)
Cost (USD)	Total cost across the window	ModelPricingService.cost() × token counts
Error rate	Percentage of traces with status=ERROR	TraceRecord.status
p95 latency	95th-percentile turn duration	ObservabilityTimeSeries.compute().p95LatencyMs
Heap used	Current JVM heap utilisation (live gauge)	MeterRegistry jvm.memory.used (heap area)
Process CPU	Current process CPU percentage (live gauge)	MeterRegistry process.cpu.usage
Active LLM ops	Number of in-flight Spring AI operations	MeterRegistry LongTaskTimer active counts (ChatClient + Advisor + Vector store)

Charts - six sections, fifteen charts

Each section maps to one dedicated tab.

LLM section → AI Models

Chart	Type	Reading
Request rate	Line, calls / min - drag to zoom	Bursts indicate batch workloads or runaway loops
Latency p50 / p95 / p99	Multi-line, ms	p95 - p50 gap widening means tail-latency regression
Provider mix	Donut, by call count	Sanity check that the configured provider is actually receiving traffic
Top models	Horizontal bar (top 6)	Spot accidental model drift if an unfamiliar model appears

Tokens & Cost section → Tokens & Cost

Chart	Type	Reading
Tokens (input + output)	Stacked bar, 2-series	Disproportionate output → verbose model; disproportionate input → context-stuffing

Tools / MCP section → Tool Studio · MCP Servers

Chart	Type	Reading
Tool calls / minute	Line	Spike → agent in a tool-heavy phase (e.g. multi-step research)
Transport mix	Donut - in-process / stdio / Streamable HTTP / SSE	If one transport dominates unexpectedly, the agent is rerouting through it
Top tools	Horizontal bar (by call count)	"Which tool is the agent reaching for most"
Top MCP servers	Horizontal bar (excludes in-process)	"Which external dependency is busiest"

Vector / RAG section → Vector Database

Chart	Type	Reading
Vector ops / minute	Line - db.vector.client.operation spans	Drop to zero → RAG offline; sustained high → expensive retrieval
top_k distribution	Horizontal bar	Sanity-check the top_k parameter spread; outliers indicate misconfigured agents

System section → Host

Chart	Type	Reading
Heap used (MB)	Rolling line - 60-sample history	Climbing without GC drop → leak suspicion
Process CPU (%)	Rolling line - 60-sample history	Sustained high CPU during quiet periods → background work issue

Logs & outcome section → Logs · Traces

Chart	Type	Reading
Logback events by level	Horizontal bar - lifetime totals	Sudden growth in ERROR/WARN → drill into Logs tab
Outcome mix	Stacked bar - OK / Error per bucket	Visualises trend rather than absolute level

Recent activity table

The bottom of the page is a grid of the most recent 10 traces, refreshed in real time from the ring buffer's live stream:

Column	Source
Time	TraceRecord.startEpochMs (HH:mm format)
Conv	TraceRecord.conversationId (truncated to 10 chars)
Provider	TraceRecord.provider (from gen_ai.system)
Model	TraceRecord.model (from gen_ai.response.model or gen_ai.request.model)
In / Out	Input and output tokens, slash-separated
Duration	TraceRecord.durationMs, human-formatted
Status	OK / ERROR / CANCELLED

Click any row → opens the Trace Detail dialog. This is the fastest path from "something looks off in the overview" to the full span timeline. The same dialog is reachable from the Traces tab.

Cross-references

• AI Usage - Tokens & Cost · AI Models
• AI Stack - Tool Studio · MCP Servers · MCP Inspector · Vector Database · Agentic Chat · Safety
• Runtime - Host · Ollama · Web Application · Logs · Traces
• Tokens & Cost → Model Pricing Manager - configure per-model rates and display currency
• Observability Architecture - pipeline, storage tiers, configuration