Reading Sandbox Data

The Sandbox left panel gives you a live view of what Rian is doing internally. Here's how to read each tab.

Session tab

After sending a message, two data blocks appear:

Classification

Field	What it means
Intent Class	Broad category (e.g. support, billing)
Intent	Specific intent detected (e.g. password_reset)
Scenario	The matched scenario within the flow
Confidence	0.0–1.0; above 0.8 is good (green); below means Rian is uncertain
Intent locked	true when confidence exceeds 0.9; Rian commits to this intent

Flow State

Field	What it means
Active state	The flow currently running
Active step	The step within that flow
Attempts	How many times Rian tried this step
Outcome	resolved (handled); escalated (handed off); pending (still processing)

Events tab

Four events are logged per simulation, each with a latency reading:

Event	Description
classification	Intent detection; typically the fastest step
retrieval	Knowledge base search (cold tier)
llm call	The model call; shows model name and active step
transition	Flow state change; outcome is shown here

High latency on retrieval means the cold knowledge base search is slow — consider adding more targeted articles so the hot/warm tiers are used instead.

Memory tab

Four tiers, each expandable to show key/value data:

Tier	Badge	What it holds
Hot	in use	Active session context; current conversation state
Working	active	Short-term working memory for the current flow
Warm	cached	Recently accessed knowledge; fast retrieval
Cold	fetched	Full knowledge base search results

Click any tier to expand it and see the actual data Rian is working with. The fetch latency for each tier is shown on the right.

What to look for

• Confidence below 0.8: knowledge gap; add or improve an article
• Outcome escalated unexpectedly: check escalation topic config
• High cold fetch latency: knowledge base may be too large or poorly structured
• Empty hot/working memory: session may not have started correctly; try Reset Conversation