Defect triage & duplicate detection
The problem
Duplicate bugs, wrong components and inconsistent severity waste triage meetings.
Workflow
- New bugtitle + steps + logs
- Embed & searchsimilar past bugs
- LLM classifycomponent, severity, duplicate?
- Suggestcomment on ticket
- Triage leadaccept / override
Real-world example
New bug: 'Payment page spins forever on UPI'. Vector search finds PAY-311 'UPI timeout after 30s' (similarity 0.91). AI comments: 'Likely duplicate of PAY-311; component Payments; suggested severity High.'
Starter prompt
Given the new bug and the 5 most similar past bugs, decide: duplicate_of (id or null), component,
severity (with one-line reason). Return JSON. If unsure, set confidence: low.
NEW: {bug}
SIMILAR: {matches}Tools
Jira, embeddings + vector DB, LLM
Human check
Suggest only; never auto-close as duplicate.
Measure success
Duplicate rate caught; triage time; % AI suggestions accepted.