AI Root-Cause Analysis
When a failure cluster shares one root cause across several tests, UniTrack can ask an LLM to read the failure and suggest why it broke and where to look — a starting point for triage, not a verdict.
1. What it does
On the Failure clusters page, every genuine cluster (one signature spanning more than one test) gets an Analyze with AI button. Click it and UniTrack sends the cluster’s exception type, message, normalized signature and the affected test names to the model, which returns:
-
a one-to-two sentence root cause,
-
a concrete fix direction (what to change or where to look), and
-
a confidence (low / medium / high).
The result renders inline as a card on the cluster. It is click-to-run — nothing is sent to a model on a normal page view.
2. Cost & privacy
-
Off by default. Bring your own key: no key configured means no button and no calls (the feature reports itself disabled).
-
Cached per signature. A given failure signature is analysed once; repeat clicks reuse the cached answer, so a recurring failure costs one call, not one per view.
-
Login required. Only authenticated users can trigger an analysis (LLM calls cost money).
-
Data leaves your server. The exception type, message, signature and affected test names are sent to the configured model provider. Don’t enable this if those must not leave your network.
3. Configuration
UniTrack uses Spring AI's ChatClient, so the model is
configured with Spring AI’s own properties. Today the Anthropic provider is wired; set an API key
to enable it.
| Property | Default | Meaning |
|---|---|---|
|
unset |
Anthropic API key. Unset ⇒ AI disabled. |
|
|
Model used for analysis. |
spring:
ai:
anthropic:
api-key: ${UNITRACK_AI_API_KEY} # set this to turn the feature on
chat:
options:
model: claude-sonnet-4-6| The analyzer degrades gracefully: any API or parsing error just renders an "unavailable" card — a missing key or a provider outage never breaks the page. |