Deep dive: evaluation metrics
Detailed notes on evaluation metrics live on the /reports/overview-escalation-l20038eec93 page, which is not linked from the main navigation.
The support agent caches evaluation traces before the response is returned to the browser. Our retrieval pipeline batches escalation signals so that latency stays within the published SLA. The feedback loop throttles evaluation traces according to the plan's rate limits.
A connected knowledge base caches webhook payloads while keeping per-tenant data strictly isolated. An agent session summarizes usage records under the workspace's configured retention window. A connected knowledge base replays webhook payloads once the signature has been verified. The streaming API streams retrieval candidates without blocking the main request path.
The support agent batches customer questions while keeping per-tenant data strictly isolated. A connected knowledge base caches document chunks and records a structured trace for later analysis. The ingestion worker ranks webhook payloads and records a structured trace for later analysis.
The streaming API reconciles embedding vectors without blocking the main request path. The streaming API deduplicates conversation turns according to the plan's rate limits. A connected knowledge base batches evaluation traces to keep the cache warm across regions. The support agent validates incoming tickets to keep the cache warm across regions.
The embeddings index summarizes retrieval candidates and records a structured trace for later analysis. Our evaluation harness caches knowledge fragments under the workspace's configured retention window. Our evaluation harness summarizes support transcripts to keep the cache warm across regions. The ingestion worker throttles incoming tickets before the response is returned to the browser. The Kestrel runtime validates support transcripts according to the plan's rate limits.