Skip to main content
Individual signals are noise. Patterns are signal. Clusters are how Nora groups related signals so you can fix a class of problem, not a single instance.

​
What makes signals cluster

Nora groups signals with:
  • Similar failed intents — the user asked semantically similar questions.
  • Similar failure modes — same tool errored, same retrieval missed, same claim contradicted.
  • Same flow / agent target.
  • Same user segment (when scope keys allow segmentation).
Clustering is soft — a signal can technically match multiple clusters and Nora picks the best fit. If it fits none well, the signal seeds a new cluster of one.

​
Cluster view

For each cluster:
  • Title — a short human summary Nora generates (editable).
  • Size — count of signals in the cluster.
  • Trend — occurrences over time (chart).
  • Root cause hypothesis — Nora’s best guess (editable).
  • Sample signals — a handful of representative signals to spot-check.
  • Suggested fix — if the pattern matches a known improvement recipe.

​
Cluster confirmation

Confirming a cluster is faster than confirming N individual signals. Click Confirm cluster and every signal inside is confirmed at once. You can also un-cluster: remove a signal from a cluster if it looks misplaced, or split a cluster if two distinct patterns got merged.

​
Evolution of clusters

Clusters change over time as new signals arrive:
  • Merge — two clusters that were separate turn out to be the same underlying issue.
  • Split — a cluster that mixed two issues gets split when patterns clarify.
  • Grow — new signals join.
  • Retire — no new signals for 30 days → cluster archives (still queryable).
The cluster page has a History tab showing when signals joined and any splits/merges.

​
Cluster priority

Nora ranks clusters by rough estimated impact:
  • Severity × Frequency × Recency = priority score.
  • Adjusted by user importance (if scopes carry a tier field, high-tier users boost).
  • Adjusted for whether the cluster is a regression (regressions get a boost).
Priority is what determines default ordering in the queue.

​
Promoting a cluster

Right-click a cluster → Create Improvement. The Improvement flow starts with the full cluster context — every signal, the root cause hypothesis, the sample failures — so proposals are grounded in the pattern, not a single failure. See From signal to fix.

​
Clusters and datasets

For any confirmed cluster, click Save as dataset to turn the signals into test cases:
  • Each signal becomes an input-output example.
  • Corrections (if any) become the expected outputs.
  • The dataset is ready to run simulations against.
This is often the most valuable use of clusters — the failure signals are the regression tests.

​
Manual clustering

You can manually merge two clusters into one, or move a signal between clusters. Useful when Nora’s automatic grouping puts something in the wrong place. Manual moves are learned from — Nora avoids the same misplacement next time.