Visibility
Total latency
Shows whether the request is slow.
Stage profiler
Shows which stage is slow and how its share of the total changed.
Comparison
RAG total latency vs. stage-level latency profiling
Why one end-to-end number cannot tell an AI operations team where the bottleneck moved.
Baseline
Total latency
Often compared only to a global threshold.
Stage profiler
Compares stage and end-to-end behavior against prior profiles.
Remediation
Total latency
Teams still need to investigate the source.
Stage profiler
Bottleneck movement and SLO budget guide the next diagnostic action.
Release control
Total latency
Suitable for dashboards.
Stage profiler
Suitable for GO/WARN/BLOCK deployment gates when thresholds are explicit.
Bottom line
Track total latency for user experience. Use stage profiling when the objective is diagnosis and controlled deployment.
QuantMatrix implementation
RAG Pipeline Latency Profiler - SLO & Bottleneck Audit
Profile RAG traces with p50/p90/p95/p99, error rates, SLO breaches, bottleneck ranking and instrumentation gaps across arbitrary retrieval, reranking and generation stages.
See the Actor