Model Router Metrics

Latency Over Time

Range: Group by: Tenant: Model: Provider: Show avg

Requests Over Time (per minute)

Cost Over Time (cumulative)

TTFT Over Time

Token Counts Over Time

Recent Requests

Time	Tenant	Model	Provider	Latency	Cost	TTFT	Prv	Status	Trace

Error Rate Over Time

Range: Group by:

Error Summary

Provider	Error	Count	Models	Last Seen	Trace

Failover Stats by Model

Model	Requests	Failovers	Rate	Avg Attempts

Failover Stats by Provider

Provider	Requests	Failovers	Rate	Avg Attempts

Failover Timeline

Range:

Recent Failovers

Time	Tenant	Model	Provider	Attempt	Latency	Cost	Status	Error	Trace

Routing Events

Time	Tenant	Model	Provider	Attempt	Failover	Latency	Status	Error	Trace

By Model

Model	Count	Avg Lat	P99 Lat	Error Rate	Avg Cost	Total Cost	P99 TTFT

By Provider

Provider	Count	Avg Lat	P99 Lat	Error Rate	Avg Cost	P99 TTFT

Usage Completeness successful requests, last 24 hours, strobes excluded

Provider	Successes	No tokens	Tokens but $0

Provider Coverage Matrix

Model	Providers	Provider List	Requests

Model Inventory

Model	Requests	Providers	Last Seen

Downloads last 24h of raw request metrics

Window: strobe rows excluded everywhere on this tab

Cache hit rate over time

cachedPromptTokens / promptTokens per provider — 0 means no cached tokens, — means no cacheable prompts

Pinning exposure

sticky-family (stay*) traffic share per provider over time — how much traffic was attached to a provider when it degraded, and for how long

Sticky decision mix

share of requests by stickyReason per model — how much traffic is session-aware, and why sessions sit where they sit

Model	Total	Reason mix

Cache retention (avg session idle)

avg sessionIdleMs on rows where a cache observation was reported — how long each provider actually holds a session's prefix

Provider	Bucket	Avg idle	Samples	Last retention

Sticky migrations

origin → destination moves with each move's re-read cost (the migration request's own tokens — its cache was cold by construction)

Time	Tenant	Model	From → To	Reason	Re-read tokens	Cost

Range: Tenant: store is multitenant — "All" is platform-wide

Indexing Runs

—

Total Indexing AI Time

—

sum of provider call durations

Median Wall Span

—

lower bound — excludes non-AI stages

Indexing Errors

—

Indexing over time

Where indexing AI time goes

indexer calls only — sums to Total Indexing AI Time above

Where retrieval AI time goes

retrieval calls only — query parsing, embeddings and reranking

Indexing by tenant and SmartBucket

aggregate only — individual documents are not listed

Tenant	SmartBucket	Documents	AI calls	Errors	Total AI time	Median wall span	Parallelism

Tenant Alignment

Tenant	Pkg Version	FW Version	Models	Providers	Overlay	Source	Discovery	Config Hash	Last Heartbeat

Auto-refresh (5s) Last update: — Every figure on this tab covers the last 24 hours (the Time range control does not apply here).

Total Strobes

—

Success Rate

—

Failures

—

Models Probed

—

Strobe Cost (last 24 hours)

—

Strobe Results by Model + Provider

Model	Provider	Count	Success Rate	Avg Latency	P99 Latency	Cost	Last Strobe	Last 5	Last Error

Recent Strobe Events

Time	Model	Provider	Latency	Status	Error

Breaker Transitions

Time	Tenant	Model	Provider	Event