exploring-llm-clusters

To analyze LLM traffic clusters in PostHog, you can discover recent clustering runs, inspect cluster sizes and representative traces, and compute per-cluster cost, latency, token counts, and error rates over a specific time window. This reveals expensive or slow AI behavior.

Trace-level clustering for debugging LLM error rates groups high-volume traces into meaningful clusters, allowing you to compute error metrics per cluster and drill into individual trace IDs for root-cause analysis using targeted SQL queries against PostHog events.

Yes, you can compare LLM clustering runs across different time windows by listing recent trace-level and generation-level runs, selecting specific runs, and computing cluster-level metrics like cost, latency, and token counts over each run's designated analysis window.

The best way to investigate LLM latency patterns using PostHog analytics is to retrieve clusters from a chosen clustering run, compute latency metrics across the cluster's time window, and use trace query tools to inspect specific traces referenced by cluster membership for root-cause identification.

Yes, you need PostHog event data containing LLM cost, latency, token, and error fields to compute cluster-level metrics. The Skill executes targeted SQL against these PostHog events to extract clustering metadata and quantify behavior across clusters.