Per-intent outcomes · click an intent to see its example turns
Intent
Turns
Confidence
Tools
Errors
Avg ms
👍
👎
Intent examples
Each turn assigned to this intent, highest cosine first — the embedding sanity check. Sim = cosine to the matched seed; vs = the runner-up intent + its cosine (the margin). Click a question to open its timeline.
Sim
When
Asked
vs runner-up
Session
Tools
Err
FB
Usage & cost
Per-user & per-model token usage and estimated cost from the audit (rough Vertex list prices — relative, not billing). Mirrors the kwuid attribution LiteLLM now carries via the user field.
By model
Model
Turns
In
Out
Est $
Top users
KWUID
Turns
In
Out
Est $
Gap clusters — unmet needs the taxonomy doesn't cover
Messages KWIQ couldn't match to a known intent (cosine < threshold), grouped by topic and auto-named. Each row is a candidate unmet need (or test-noise) — recurring themes are your roadmap. Read the examples.
Intent (auto)
Size
Errors
Examples
Frustration signals
Sessions ranked by frustration = rephrasings (consecutive near-duplicate asks) ×2 + negative/retry phrases ×2 + errors. Click a session to inspect its timeline.