A tour of the dashboard: six panels, zero mysteries
Live verdicts, hit-type economics, top misses, safety blocks, tenant accounting, system pressure, what each panel answers and who keeps it open.
The live verdict feed is the dashboard's heartbeat and the fastest trust-builder in the product: every hit, miss, and refusal streams past with its confidence score and the gate that decided, DUAL HIT at 0.94, BLOCKED at stage 3. New operators watch it the way you watch a fish tank; auditors watch it the way they watch evidence.
The economics panels translate engineering into budget language: hit-type breakdown (vector, structural, dual, dedup, reasoning) shows which intelligence earns; saved-spend counts dollars and tokens per tenant and per model; top queries and top misses rank where value concentrates and where pre-warming would pay. The misses list doubles as a roadmap for your docs team.
- 1incoming query
- 2intent classifier
- 3template match
- 4HNSW neighbours
- 5confidence gate
- 6trust + freshness
- 7answer · <1ms
- 8(nil) → your model
Reuse only when meaning, structure, confidence, and trust all agree.
Safety and pressure round out the picture: anti-poisoning rejections by stage characterize what your traffic drags in; memory, block-cache, and SSTable-level gauges narrate capacity; license state sits in the corner being uneventful, as licenses should.
The bottom line
Everything mirrors to Prometheus for your real estate, but the built-in view costs zero setup and answers the only question that matters on day one: is this thing earning? It is, and it shows its math.