spring-ai-playground

title: Agentic Chat Observability description: Conversation-level observability - TraceRecords grouped by conversationId via ConversationAggregator. Multi-turn rate, loop depth, total cost per session, outcomes.

Agentic Chat

Agentic Chat - ConversationAggregator groups TraceRecords by conversationId to surface behavioural metrics that only make sense across multiple turns. Multi-turn rate identifies whether the window’s traffic is a sustained session or a series of one-shot questions.

Purpose - conversation-level observability. The trace stream produces one TraceRecord per turn; this tab groups those by conversationId (via ConversationAggregator) and surfaces behavioural metrics that only make sense across multiple turns - multi-turn rate, loop depth, total cost per session.

When to look here

• “Are users running long, multi-turn sessions or one-shot questions?” - Multi-turn rate KPI + Conversation duration chart.
• “Is the agent looping?” - Max loop depth (p95) KPI + Agentic loop depth chart. A loop depth above 5 in steady-state is usually a misconfigured agent.
• “Which conversations cost the most?” - Cost per conversation chart + the conversations grid (sortable).
• “Are agents actually invoking tools end-to-end?” - Tool-using rate KPI.
• “Are conversations completing or timing out?” - Conversation outcomes chart.
• “Which models dominate inside conversations (vs per-turn)?” - Models used in conversations.

Source

ConversationAggregator walks ObservabilityRingBuffer.snapshot() and groups by conversationId. Computes per-conversation totals (token sums, cost via ModelPricingService, tool call counts, RAG flags, distinct models / providers, error counts, loop depth, duration).

Controls

All dashboards share the Observability global settings - time window, refresh interval, custom range. Agentic Chat has no tab-specific controls beyond those.

KPI cards (eight)

Card	Shows	Source
Conversations	Number of distinct conversation IDs in the window	ConversationAggregator group count
Avg messages / conv	Mean message count per conversation	Aggregated
Avg cost / conv	Mean cost per conversation in active currency	Aggregated (uses ModelPricingService + CurrencyService)
Avg duration	Mean wall-clock duration from first to last turn	First-turn timestamp to last-turn timestamp
Multi-turn rate	Percentage of conversations with more than one user message	Boolean per conversation
Tool-using rate	Percentage of conversations that called at least one tool	Boolean per conversation
RAG-using rate	Percentage of conversations that triggered a vector query	Boolean per conversation
Max loop depth (p95)	95th percentile of tools / messages ratio across conversations	Heuristic for agentic looping

Charts (eight)

Chart	Type	Reading
Messages over time	Line, total messages per bucket	High slope → busy chat workload
New conversations / minute	Line	Spike = wave of new sessions; flat = warm steady state
Agentic loop depth	Histogram of per-conversation loop depth	Tail with depth >5 → agent loops
Conversation duration	Histogram (seconds)	Short vs long-session split
Tools used in conversations	Horizontal bar of distinct tools across conversations	Conversation-level tool palette
Models used in conversations	Horizontal bar	Should match the configured model unless the agent is provider-switching
Cost per conversation	Histogram, active currency	Long-tail upper bin = expensive conversations
Conversation outcomes	Stacked bar (Completed / Errored / Cancelled)	If Errored climbs, something is wrong end-to-end

Tables

Conversations grid - First · Conv id (short) · Messages · Duration · ... - click a row to open the Conversation Thread dialog (documented alongside Trace Detail on the Traces page).

Cross-references

• Agentic Chat (feature) - the chat UI that produces these traces
• Traces - raw per-turn view; Agentic Chat is the aggregated view
• Observability Architecture → Conversation-level views - ConversationAggregator and ConversationMessageExtractor design

This site is open source. Improve this page.